AIGP Practice Test

Artificial Intelligence Governance Professional (AIGP)

AIGP overview

  • 200 practice questions
  • 32 flashcards
  • 4 study guide topics
  • 40 key terms

What you'll study

  • Foundations of AI systems, their use cases and impacts25%
  • AI laws, regulation and frameworks for responsible governance25%
  • AI life cycle and AI risk management25%
  • Emerging and debated issues in AI governance25%

The Artificial Intelligence Governance Professional (AIGP) is a professional credential offered by the IAPP for people who need to understand and execute responsible AI governance. The IAPP frames it as a demonstration of competency in AI development, ethical AI deployment, and the use of best practice in AI management to ensure safety and trust - and notes that, now that AI has made its debut, professionals in all industries need this capability, not just technologists or lawyers.

The exam's content is organized around four areas the IAPP sets out directly. First, foundational knowledge of AI systems, their use cases, the impacts of AI, and responsible AI principles. Second, how current and emerging laws apply to AI systems, and how major frameworks can be responsibly governed. Third, the AI life cycle, the context in which AI risks are managed, and implementation of responsible governance. Fourth, awareness of unforeseen concerns and the debated issues surrounding AI governance.

Preparation materials are published by the IAPP: the AIGP Body of Knowledge and Exam Blueprint, a downloadable study guide covering exam format with sample questions, and broader options such as textbooks, flexible training, and self-assessment tests. The IAPP also strongly encourages candidates to read its certification candidate handbook for testing policies and procedures before sitting the exam.

This page is an independent study resource. It is not affiliated with, endorsed by, or operated by the IAPP.

Common mistakes

When a stem describes a use case (e.g., predicting loan default from structured tabular data, or ranking candidates from free-text CVs), the candidate selects a model family by buzzword association - picking a generative LLM or 'deep learning' because it appears in recent headlines - rather than by matching data type, label availability, need for calibrated uncertainty, latency, and explainability constraints.

Candidates memorise model names from news coverage and skip the requirements-to-technique mapping step; generative AI dominates attention, so it becomes the default answer even for problems where simpler supervised or even rule-based methods are the better governance choice.

What to do instead

Before answering, write down the four or five use-case constraints from the stem (data modality, labelled vs unlabelled, need for probability estimates, latency/compute budget, explainability demand) and eliminate options that fail a stated constraint; practise classifying use cases into supervised, unsupervised, reinforcement, generative, and symbolic/rule-based categories with one governance-relevant rationale per choice.

The candidate equates 'AI system' with machine learning or with large language models, so questions about autonomy, inference, adaptivity, or rule-based expert systems get answered as if only trained statistical models count; consequently they misjudge whether a described system falls in or out of scope of a definition or obligation.

Most public discourse and vendor marketing uses 'AI' as a synonym for generative ML, so candidates import that narrow usage into exam stems that deliberately describe a non-ML or hybrid system.

What to do instead

Memorise the widely used governance definition (a machine-based system that infers, from input it receives, how to generate outputs such as predictions, content, recommendations or decisions, with varying levels of autonomy and adaptivity) and test each stem scenario against each element - inference, autonomy, adaptivity, goal orientation - rather than against 'is it a neural network'.

The candidate places a described system in the wrong tier of the EU AI Act's risk pyramid, or treats general-purpose AI obligations as if they were an additional high-risk requirement, or assumes a prohibited practice becomes permissible with disclosure or consent.

The Act layers several distinct regimes (prohibited practices, high-risk under Annex III and product-safety routes, limited-risk transparency duties, minimal-risk, plus a separate GPAI chapter), and candidates collapse them into a single 'high-risk or not' question.

What to do instead

Build and rehearse a one-page decision tree: first ask 'is the practice prohibited?', then 'is it a product-safety or Annex III high-risk use?', then 'is there a transparency obligation (chatbot, synthetic content, emotion recognition, deepfake)?', and treat GPAI provider duties (technical documentation, copyright policy, systemic-risk evaluation) as a parallel track. Note that transparency and consent do not cure a prohibition.

The candidate assigns obligations to the wrong actor - describing a deployer's duties as a provider's, or assuming a user who substantially modifies or rebrands a system stays a deployer rather than becoming a provider.

Frameworks use role-based allocation (provider/developer, deployer/user, importer, distributor, authorised representative), and candidates learn 'the organisation must...' lists without attaching each duty to a specific role.

What to do instead

For every framework you study, make a three-column table: actor, trigger for that role, core obligations. Specifically rehearse the re-branding/substantial-modification rule that converts a deployer into a provider, and the deployer-specific duties (human oversight, input data relevance, monitoring, informing affected persons).

The candidate treats NIST AI RMF, ISO/IEC 42001, the OECD AI Principles, and UNESCO's Recommendation as interchangeable - for example, describing the NIST framework as certifiable, ISO/IEC 42001 as government-enforced law, or the OECD Principles as a technical control catalogue.

All are described in similar 'responsible AI' language, so candidates match on content keywords and ignore the structural differences in issuer, legal force, auditability, and intended audience.

What to do instead

Memorise a comparison grid with four columns: issuer and type (standards body, government agency, intergovernmental organisation), voluntary vs certifiable vs legally binding, what artefact it produces (management system, risk process, principles, control set), and who it is written for. Anchor each framework with one distinctive fact (e.g., ISO/IEC 42001 is a certifiable management-system standard; the NIST AI RMF is voluntary and process-oriented with govern/map/measure/manage functions).

The candidate answers lifecycle questions as if governance ends at the pre-deployment approval or procurement sign-off, so they omit post-deployment monitoring, performance drift detection, incident response, change management, re-validation after model updates, and decommissioning or retirement of the system.

Organisational narratives and exam-prep summaries foreground 'approve the model' gates, and candidates reproduce a stage-gate diagram while dropping the continuous-operations half of the lifecycle.

What to do instead

Draw the lifecycle as a loop, not a line, and for each phase name one governance artefact (use-case intake and triage, data provenance record, evaluation and red-team report, deployment decision record, monitoring dashboard with thresholds, incident log, retirement record). In any scenario question, ask 'what happens after go-live?' before selecting an answer.

The candidate uses risk vocabulary loosely - treating inherent risk and residual risk as the same number, answering an impact-assessment question with a risk-assessment answer (or vice versa), or claiming a risk is 'acceptable' without referencing the organisation's stated risk appetite or tolerance.

The terms appear in overlapping documents and candidates learn definitions in isolation without a model of how they relate arithmetically and procedurally.

What to do instead

Fix a mental formula: inherent risk = likelihood x severity before controls; residual risk = after controls, compared against a documented tolerance threshold set by risk appetite. Separately learn that an impact assessment examines effects on people, groups, and rights (who is affected, how severely, can they contest), while a risk assessment examines the organisation's objectives and likelihood of harm - then check which noun the stem actually uses.

The candidate treats a vendor's assurances, a third-party audit certificate, or a model card as transferring accountability, and writes answers that omit contractual controls, ongoing supplier monitoring, and the deploying organisation's own due diligence and validation before and during use.

Certificates and supplier documentation feel authoritative and reduce perceived workload, so candidates assume they close the risk rather than inform it - a classic outsourcing reflex carried over from other compliance domains.

What to do instead

Adopt the rule 'accountability does not transfer with the model'. For every supply-chain scenario, answer from the deploying organisation's duties: pre-contract due diligence and use-case fit, contractual terms (audit rights, incident notification, data-use limits, change notification), independent validation, and periodic reassessment including supplier model-version changes.

On emerging-issues questions the candidate picks the most absolute option available - 'AI output can never be copyrighted', 'an AI system can never bear legal responsibility', 'open-weight release is always irresponsible' - instead of the option that reflects the current state of the debate and the conditions under which positions shift.

Ambiguity feels unsafe in a multiple-choice format, and candidates convert a contested, fact-dependent question into a bright-line rule so they can answer decisively.

What to do instead

For each debated topic, write a two-sided card: what is settled, what is contested, and which facts move the answer (degree of human creative control; documented human oversight and attribution; capability level and available safeguards; jurisdiction). Practise choosing the conditional answer ('depends on whether X is documented') over the categorical one when the stem supplies unresolved facts.

The candidate collapses distinct transparency concepts - answering a question about disclosing that a person is interacting with an AI, or about the use of training data, with an answer about explaining an individual decision, or vice versa; similarly, they mix input-side IP questions (lawful use of protected works in training) with output-side questions (infringement or authorship in generated content).

All of these are tagged 'transparency' or 'IP' in preparatory material, and candidates pattern-match the keyword rather than the specific obligation and the direction of the data flow.

What to do instead

Build a disambiguation table with rows for disclosure of AI interaction, disclosure of synthetic or manipulated content, explanation of a specific automated decision, publication of training-data summaries or copyright policy, and model/system documentation - each with its own trigger and audience. For IP, always label the question 'input (training)' or 'output (generation)' before evaluating any option.

Exam requirements can change. Confirm current details with the exam provider.

AIGP practice

200 questions. Each one lists the choices, the correct answer, and an explanation. With JavaScript, this list opens one question at a time. The checker at the top of this page saves progress in this browser when it is available.

  1. Question 1Foundations of AI systems, their use cases and impacts

    Which statement best characterizes the AI systems widely deployed in organizations today?

    1. They can reason across any domain at human level.
    2. They learn with no need for data or feedback.
    3. They possess self-awareness and independent intent.
    4. They are typically designed for specific, narrow tasks. Correct answer.

    Explanation

    AI systems in widespread commercial use are narrow: they are trained or configured for a defined task such as classification, ranking, or generation. General-purpose, self-directed reasoning at human level remains a research goal rather than a deployed capability, and every current learning approach still depends on data or feedback.

  2. Question 2Foundations of AI systems, their use cases and impacts

    A team trains a model on 50,000 emails, each already tagged 'spam' or 'not spam'. Which learning paradigm is being used?

    1. Supervised learning Correct answer.
    2. Unsupervised learning
    3. Reinforcement learning
    4. Self-supervised clustering

    Explanation

    Supervised learning trains on examples where each input is paired with the correct output, so the model learns a mapping from input to label. Spam detection using pre-tagged emails is the canonical case of that setup.

  3. Question 3Foundations of AI systems, their use cases and impacts

    A retailer groups its customers by purchasing behavior without defining any categories in advance. Which approach does this describe?

    1. Reinforcement learning
    2. Unsupervised clustering Correct answer.
    3. Supervised classification
    4. Rule-based expert system

    Explanation

    Unsupervised learning finds structure in data that carries no pre-assigned labels, and clustering customers by behavior is a typical use. The absence of predefined categories is the decisive signal that the approach is unsupervised.

  4. Question 4Foundations of AI systems, their use cases and impacts

    After a model has been trained and deployed, it scores incoming loan applications. What is this phase called?

    1. Inference Correct answer.
    2. Feature engineering
    3. Backpropagation
    4. Data labelling

    Explanation

    Inference is the phase in which a trained model is applied to new inputs to produce outputs. Training, by contrast, is the earlier phase in which model parameters are estimated from data.

  5. Question 5Foundations of AI systems, their use cases and impacts

    What distinguishes deep learning from classical machine learning?

    1. It produces only rule-based decision trees.
    2. It only works with structured tabular data held in spreadsheets.
    3. It requires no labelled training data and no human oversight.
    4. Layered neural networks learn representations automatically. Correct answer.

    Explanation

    Deep learning refers to neural networks with many hidden layers that learn feature representations directly from raw inputs rather than relying on hand-engineered features. It still requires substantial labelled data and computing power.

  6. Question 6Foundations of AI systems, their use cases and impacts

    In supervised learning, what is a label?

    1. A weight adjusted during training
    2. A metric used to score model accuracy
    3. The raw input values that are fed into the model during training
    4. The known target answer paired with a training example Correct answer.

    Explanation

    A label is the ground-truth output attached to a training example, which the model is optimized to reproduce. Inputs, learned weights, and evaluation metrics are separate concepts in the pipeline.

  7. Question 7Foundations of AI systems, their use cases and impacts

    Which task is the clearest example of computer vision?

    1. Generating a written summary of a meeting transcript
    2. Flagging pedestrians in vehicle camera frames Correct answer.
    3. Ranking search results by click-through rate
    4. Predicting next quarter's sales from historical tables

    Explanation

    Computer vision extracts information from images and video, such as detecting objects or people in camera frames. The other options involve tabular forecasting, text summarization, and ranking, which belong to different task families.

  8. Question 8Foundations of AI systems, their use cases and impacts

    How does a large language model generate text?

    1. By predicting the next token from prior context Correct answer.
    2. By applying hand-written grammar rules in sequence
    3. By ranking documents retrieved from the internet
    4. By looking up stored answers in a fixed answer database

    Explanation

    Large language models are autoregressive next-token predictors: they estimate a probability distribution over the next token given preceding text and sample from it repeatedly. They do not retrieve canned answers or follow hand-written grammar rules.

  9. Question 9Foundations of AI systems, their use cases and impacts

    A model scores 99% on its training data but only 62% on new customer records. What is the most likely cause?

    1. Underfitting caused by an overly simple model
    2. Overfitting to the training data Correct answer.
    3. Inference latency being too high
    4. A held-out test set that is too large

    Explanation

    A large gap between very high training performance and much lower performance on unseen records is the signature of overfitting, where the model memorizes idiosyncrasies of the training set. Typical remedies include more data, regularization, or a simpler model.

  10. Question 10Foundations of AI systems, their use cases and impacts

    Why is a held-out test set kept separate during model development?

    1. To speed up the training process on very large datasets
    2. To tune hyperparameters before the final training run
    3. To estimate performance on unseen data Correct answer.
    4. To store backups of the training data

    Explanation

    A held-out test set is never used for fitting, so its results approximate how the model will behave on genuinely new data. Repeatedly tuning against the test set leaks information and inflates the reported performance.

  11. Question 11Foundations of AI systems, their use cases and impacts

    A robotics team rewards a simulated agent for each successful grasp and penalizes collisions. Which paradigm is being used?

    1. Rule-based expert system
    2. Supervised learning
    3. Reinforcement learning Correct answer.
    4. Unsupervised learning

    Explanation

    Reinforcement learning trains an agent by rewarding desired actions and penalizing undesired ones through interaction with an environment. No dataset of labelled correct actions is supplied to the agent.

  12. Question 12Foundations of AI systems, their use cases and impacts

    What is a defining characteristic of a foundation model?

    1. It is pretrained broadly and adapted to many tasks Correct answer.
    2. It requires no compute resources to run
    3. It is trained from scratch on data from one narrow task only
    4. It produces only numerical outputs, never text

    Explanation

    Foundation models are trained on broad data at scale and then adapted, through prompting or fine-tuning, to a wide range of downstream tasks. This contrasts with task-specific models built for a single purpose.

  13. Question 13Foundations of AI systems, their use cases and impacts

    A streaming service suggests films by identifying users whose viewing histories resemble a given user's history. What is this technique called?

    1. Optical character recognition
    2. Speech synthesis
    3. Collaborative filtering Correct answer.
    4. Content moderation

    Explanation

    Collaborative filtering recommends items by exploiting similarity between users' or items' interaction histories. It can operate without any understanding of the content of the items themselves.

  14. Question 14Foundations of AI systems, their use cases and impacts

    A hiring model trained on a decade of past decisions consistently ranks candidates from one demographic lower. What is the most direct explanation?

    1. Bias in historical labels was learned by the model Correct answer.
    2. The model has too few parameters to learn properly
    3. The inference hardware is misconfigured
    4. The model was trained for too many epochs

    Explanation

    Models learn whatever patterns are present in their training data, including historical inequities embedded in the labels. Disparate outcomes across groups therefore usually trace back to the training signal rather than to hardware or capacity.

  15. Question 15Foundations of AI systems, their use cases and impacts

    What is the primary purpose of human-in-the-loop review in a high-stakes AI workflow?

    1. To convert unstructured data into a structured format
    2. To reduce the volume of training data needed
    3. To let a person check or override high-risk outputs Correct answer.
    4. To eliminate the need for any model monitoring

    Explanation

    Human-in-the-loop review places a person in the decision path to verify, correct, or escalate model outputs where errors carry serious consequences. It complements rather than replaces model monitoring and governance.

  16. Question 16Foundations of AI systems, their use cases and impacts

    A fraud detection model degrades sharply after customers begin paying in a new way. Which phenomenon does this illustrate?

    1. A feature scaling error
    2. Underfitting of the training data
    3. Overfitting to the test set
    4. Model drift Correct answer.

    Explanation

    Model drift is the deterioration of predictive performance as the world diverges from the conditions captured during training. It can be driven by changes in the inputs, in the input-to-output relationship, or in both.

  17. Question 17Foundations of AI systems, their use cases and impacts

    Why is explainability especially important for AI used in credit decisions?

    1. It removes the need for training data
    2. Applicants may need reasons for a denial Correct answer.
    3. It makes the model train faster on GPUs
    4. It increases the raw accuracy of the model

    Explanation

    Explainability supports accountability: decision subjects and reviewers need to understand the basis of an adverse decision and to challenge it where appropriate. It does not itself improve accuracy, reduce data needs, or accelerate training.

  18. Question 18Foundations of AI systems, their use cases and impacts

    Which activity belongs to the post-deployment stage of the AI lifecycle?

    1. Monitoring live performance and retraining when needed Correct answer.
    2. Defining the business problem and success metrics
    3. Collecting and labelling the initial training data set for the project
    4. Selecting the model architecture and loss function

    Explanation

    Post-deployment work centers on operating the model in production, which includes monitoring performance and data quality and retraining when conditions change. Data collection, architecture selection, and problem definition all precede deployment.

  19. Question 19Foundations of AI systems, their use cases and impacts

    Classifying product reviews as positive or negative is a task within which area?

    1. Speech synthesis
    2. Natural language processing Correct answer.
    3. Computer vision
    4. Robotic process automation

    Explanation

    Sentiment classification operates on written text, which places it within natural language processing. Vision, robotic process automation, and speech synthesis address different data modalities or process functions.

  20. Question 20Foundations of AI systems, their use cases and impacts

    A model predicts house prices from size, number of bedrooms, and location. These inputs are called:

    1. Hyperparameters
    2. Features Correct answer.
    3. Labels
    4. Epochs

    Explanation

    Features are the input variables a model uses to make a prediction. Labels are the target outputs, epochs are passes through the training data, and hyperparameters configure the learning process.

  21. Question 21Foundations of AI systems, their use cases and impacts

    A hospital has only 2,000 labelled X-rays. Fine-tuning a model pretrained on millions of general images outperforms training from scratch. Why?

    1. Pretraining encodes visual features that transfer to the new task Correct answer.
    2. Pretraining removes all bias from the original image data
    3. Fine-tuning eliminates the need for any labelled X-rays
    4. Fine-tuning guarantees the model cannot overfit small datasets

    Explanation

    Pretraining on large image corpora produces generic visual features, such as edge and texture detectors, that fine-tuning can repurpose with modest amounts of labelled data. Fine-tuning still requires some labels and offers no guarantee against overfitting or inherited bias.

  22. Question 22Foundations of AI systems, their use cases and impacts

    How does retrieval-augmented generation reduce hallucination in a language model assistant?

    1. It retrains all of the model's weights from scratch on every user query
    2. It blocks the model from producing any text output
    3. It lowers the sampling temperature all the way to exactly zero
    4. It places retrieved documents in the prompt so answers are grounded Correct answer.

    Explanation

    RAG retrieves relevant passages at query time and inserts them into the prompt, so the model conditions its answer on source material. This grounds outputs and reduces fabrication without changing the model's weights.

  23. Question 23Foundations of AI systems, their use cases and impacts

    How does a classical expert system differ from a neural network?

    1. It requires GPUs to perform inference at scale
    2. It applies rules written explicitly by domain experts Correct answer.
    3. It learns patterns from large datasets of labelled examples
    4. It cannot represent domain knowledge in any form

    Explanation

    Expert systems encode human-authored if-then rules and a curated knowledge base, so their reasoning is explicit and inspectable. Neural networks instead learn statistical patterns from data and are not hand-programmed with rules.

  24. Question 24Foundations of AI systems, their use cases and impacts

    A factory runs defect detection on a device attached to the production line rather than sending video to the cloud. What is the main benefit?

    1. It eliminates all hardware costs
    2. Model accuracy always increases when inference runs at the edge
    3. Lower latency and less exposure of sensitive footage Correct answer.
    4. It removes the need to train the model at all

    Explanation

    Running inference locally on the device cuts the round trip to a remote server and keeps sensitive video on the premises. Edge deployment changes where computation happens, not whether training is required.

  25. Question 25Foundations of AI systems, their use cases and impacts

    A model's inputs keep the same distribution, but the relationship between those inputs and the target changes after a policy change. What is this called?

    1. Data drift in the input features
    2. Overfitting to the training set
    3. Class imbalance in the labels
    4. Concept drift Correct answer.

    Explanation

    Concept drift is a change in the relationship between inputs and the target, so identical inputs now imply different outcomes. Data drift, by contrast, is a change in the distribution of the inputs themselves.

  26. Question 26Foundations of AI systems, their use cases and impacts

    Which description best characterizes a 'narrow' AI system?

    1. A system that can rewrite its own objectives without human input
    2. A system that matches human performance across every cognitive task
    3. A system built for one bounded task, such as transcribing speech Correct answer.
    4. A system that needs no data to reach expert-level accuracy

    Explanation

    Nearly all deployed AI is narrow: it is optimized for a defined task and performs poorly outside that task's scope. Generality across all cognitive tasks, self-authored goals, and data-free competence are not properties of current production systems.

  27. Question 27Foundations of AI systems, their use cases and impacts

    After a model has been trained and deployed, the process of using it to score new inputs is called:

    1. Annotation
    2. Tokenization
    3. Inference Correct answer.
    4. Backpropagation

    Explanation

    Training adjusts the model's parameters; inference is the forward pass that applies the trained parameters to new data to produce an output. In production, inference dominates compute cost and latency.

  28. Question 28Foundations of AI systems, their use cases and impacts

    Which characteristic distinguishes unsupervised learning from supervised learning?

    1. It can only be applied to image data
    2. It works with data that carries no target labels Correct answer.
    3. It always requires a human to label every input
    4. It guarantees higher accuracy than supervised methods

    Explanation

    Unsupervised methods look for structure such as clusters, groupings, or lower-dimensional representations without a target variable. Supervised learning, by contrast, requires labelled examples pairing inputs with the desired output.

  29. Question 29Foundations of AI systems, their use cases and impacts

    Converting a recording of a customer's spoken complaint into written text relies on which capability?

    1. Machine translation
    2. Automatic speech recognition Correct answer.
    3. Image segmentation
    4. Optical character recognition

    Explanation

    Automatic speech recognition maps an audio waveform onto a sequence of words. The other options operate on images or on text in a different language.

  30. Question 30Foundations of AI systems, their use cases and impacts

    What distinguishes generative AI models from traditional predictive models?

    1. They are trained without any labelled data at all
    2. They never make errors on unseen inputs
    3. They can only produce numeric risk scores
    4. They generate new content such as text or images Correct answer.

    Explanation

    Generative models learn the distribution of their training data well enough to synthesize new samples, such as passages of text or pictures. Classical predictive models instead output a discrete class or a numeric estimate.

  31. Question 31Foundations of AI systems, their use cases and impacts

    A team wants to analyze customer records stored as rows and columns in a relational database. How should this data be described?

    1. Synthetic data generated by a model
    2. Structured, tabular data Correct answer.
    3. Unstructured data such as free-text notes
    4. Streaming sensor telemetry

    Explanation

    Rows and columns with a predefined schema are the defining form of structured data, which maps cleanly onto spreadsheets and database tables. Free text, images, and audio are unstructured because they lack that fixed schema.

  32. Question 32Foundations of AI systems, their use cases and impacts

    Which statement correctly describes the relationship between AI, machine learning, and deep learning?

    1. Deep learning is the broadest of the three categories overall
    2. AI is a subset of machine learning restricted to neural networks
    3. Machine learning and deep learning are unrelated disciplines
    4. Deep learning sits inside machine learning, which sits inside AI Correct answer.

    Explanation

    The terms form nested sets: AI is the broad goal of machine intelligence, machine learning is the subset that learns from data, and deep learning is the subset of machine learning built on multi-layer neural networks.

  33. Question 33Foundations of AI systems, their use cases and impacts

    Which of the following is a hyperparameter?

    1. A single pixel value in a training image
    2. The learning rate chosen before training begins Correct answer.
    3. The weight a model learns for one input feature
    4. The value of the target in a test record

    Explanation

    Hyperparameters are settings fixed by the practitioner before training, such as learning rate, batch size, or tree depth. Parameters, by contrast, are learned from the data during training.

  34. Question 34Foundations of AI systems, their use cases and impacts

    A bank wants a model that assigns each loan applicant to one of two risk grades. Which task type is this?

    1. Clustering
    2. Dimensionality reduction
    3. Regression
    4. Binary classification Correct answer.

    Explanation

    Assigning each case to one of two discrete categories is binary classification. Regression would apply if the output were a continuous quantity such as an exact interest rate.

  35. Question 35Foundations of AI systems, their use cases and impacts

    A model flagged 200 transactions, and 150 of those were genuinely fraudulent. Over the same period there were 300 actual frauds, of which the model caught 150. What is the model's precision?

    1. 0.33
    2. 0.75 Correct answer.
    3. 0.60
    4. 0.50

    Explanation

    Precision is the share of positive predictions that are correct, here 150 true positives out of 200 flagged, or 0.75. The 150 of 300 caught figure is recall, a different metric.

  36. Question 36Foundations of AI systems, their use cases and impacts

    A dataset contains 1,000 fraudulent transactions among 1,000,000 total. A model that always predicts 'not fraud' reports 99.9% accuracy. What does this demonstrate?

    1. The model has learned the fraud pattern with high precision
    2. The dataset is too small to train any reliable model
    3. Accuracy hides failure when one class dominates Correct answer.
    4. Accuracy is only a valid metric for regression tasks

    Explanation

    When the negative class is overwhelming, a trivial predictor achieves high accuracy while catching zero frauds. Metrics focused on the minority class, such as recall, precision, or F1, are needed alongside accuracy.

  37. Question 37Foundations of AI systems, their use cases and impacts

    In semi-supervised learning, the training data consists of:

    1. Labels that are withheld from the model at prediction time
    2. Some labelled data plus a larger unlabelled set Correct answer.
    3. Synthetic examples that have no relation to real data
    4. Only unlabelled examples, with no target values at all

    Explanation

    Semi-supervised learning combines a small labelled set with a much larger unlabelled set, using the unlabelled data to learn structure that improves the decision boundary. It is useful when labelling is costly but raw data is plentiful.

  38. Question 38Foundations of AI systems, their use cases and impacts

    Why does an ensemble of decision trees often generalize better than a single tree?

    1. It removes the need for a labelled training set altogether
    2. Averaging varied trees reduces the variance of predictions Correct answer.
    3. It guarantees the absence of bias in every prediction
    4. It makes each individual tree easier to interpret

    Explanation

    Individual deep trees are unstable: small changes in the data can alter them substantially. Combining many decorrelated trees cancels much of that variability, lowering variance and improving generalization.

  39. Question 39Foundations of AI systems, their use cases and impacts

    A team discovers that 12% of the dates in its training records are in the wrong format, producing impossible values. What is the most appropriate first step?

    1. Profile and clean the data before training any model Correct answer.
    2. Deploy the model and monitor for errors
    3. Add more layers to the neural network to compensate
    4. Increase the learning rate so the model adapts

    Explanation

    Corrupted fields violate the assumptions of both the data pipeline and the model, so they must be profiled and corrected before training. Model-side changes cannot recover signal from values that are simply wrong.

  40. Question 40Foundations of AI systems, their use cases and impacts

    Why can including personal data in a model's training set create a privacy risk even if the raw records are never published?

    1. Training data is always publicly downloadable by anyone
    2. Models can memorize and later reveal training records Correct answer.
    3. Personal data may not legally be stored in any database
    4. Privacy law prohibits the use of any purely numeric data

    Explanation

    Models, especially large ones, can retain enough detail about individual examples to reproduce them when prompted or probed. This means protection must extend beyond access controls on the original dataset.

  41. Question 41Foundations of AI systems, their use cases and impacts

    A dermatology model is trained only on images from a single hospital serving one demographic group. What is the main risk?

    1. The model will store copies of every image it saw
    2. It may perform poorly for under-represented skin types Correct answer.
    3. The model will need fewer images than comparable models
    4. The model will train too slowly to be practical

    Explanation

    Training data that is not representative produces a model tuned to the patterns of the sampled population. Accuracy typically degrades for groups whose appearance or conditions were rare or absent during training.

  42. Question 42Foundations of AI systems, their use cases and impacts

    A radiologist accepts nearly every AI suggestion without independent review, including some confident errors. What does this illustrate?

    1. Automation bias: the reviewer defers to the system Correct answer.
    2. Data drift affecting the imaging pipeline over time
    3. Overfitting of the underlying diagnostic model
    4. A limitation inherent to unsupervised learning methods

    Explanation

    Automation bias is the human tendency to over-trust automated advice, which erodes the independent check that human oversight is meant to provide. Mitigations include requiring the reviewer to record their own assessment before seeing the model output.

  43. Question 43Foundations of AI systems, their use cases and impacts

    In a shadow deployment, a new model:

    1. Receives live inputs but its outputs are only logged Correct answer.
    2. Is trained on live production traffic and never evaluated offline
    3. Is deployed to all users only after each one opts in
    4. Replaces the current model for a random half of all users

    Explanation

    Shadow mode runs the candidate model on real traffic while the incumbent continues to serve decisions, so the team can compare outputs without user risk. It is a common step between offline evaluation and full rollout.

  44. Question 44Foundations of AI systems, their use cases and impacts

    A support assistant reads user messages and also follows any instruction found inside retrieved documents, including ones telling it to reveal its system prompt. Which risk does this describe?

    1. Feature scaling error
    2. Class imbalance
    3. Model underfitting
    4. Indirect prompt injection Correct answer.

    Explanation

    Indirect prompt injection occurs when hostile instructions arrive through content the model ingests, such as retrieved pages or documents, rather than from the user directly. Defences include treating retrieved content as data and validating actions before execution.

  45. Question 45Foundations of AI systems, their use cases and impacts

    What is an embedding?

    1. A label attached to each training example by an annotator
    2. A numeric vector representing an item's meaning Correct answer.
    3. A filter rule that blocks unsafe model outputs
    4. A human-written summary of a model's overall behaviour

    Explanation

    An embedding maps an item such as a word, sentence, or image into a dense numeric vector so that similar items land close together in that space. Distances between embeddings power search, clustering, and recommendation.

  46. Question 46Foundations of AI systems, their use cases and impacts

    A model that detects pneumonia from X-rays works well at one hospital but fails elsewhere. Analysis shows that images containing a chest drain or a portable-scanner marker were nearly always positive in the training set. What best explains the failure?

    1. The model was trained with a learning rate that was too high
    2. The test images had a much higher resolution than the training images
    3. The model was accidentally trained with unsupervised learning
    4. The model learned a spurious cue correlated with the label Correct answer.

    Explanation

    The model latched onto incidental features that happened to correlate with the label in one setting, a shortcut rather than the clinical signal. Such spurious correlates do not transfer to sites where the correlation is absent or reversed.

  47. Question 47Foundations of AI systems, their use cases and impacts

    Why can a model rarely satisfy demographic parity and equalized odds simultaneously when group base rates differ?

    1. Equalized odds requires removing all protected attributes entirely
    2. Fairness can only be measured after a model is deployed
    3. The two criteria conflict when group base rates differ Correct answer.
    4. One of the metrics is defined only for regression models

    Explanation

    Formal results show that, apart from degenerate cases, equalized odds and demographic parity cannot both hold when the groups have different prevalence of the positive outcome. Teams must therefore choose a fairness criterion explicitly and justify the trade-off.

  48. Question 48Foundations of AI systems, their use cases and impacts

    A city deploys a model that sends patrols to predicted crime hotspots. Arrests rise in those areas, which supplies more training data from them. What dynamic does this describe?

    1. Regularization
    2. Catastrophic forgetting
    3. Data leakage
    4. A self-reinforcing feedback loop Correct answer.

    Explanation

    The model's own actions change the data that will train its successor, so predictions and enforcement reinforce each other. This endogenous loop can entrench initial patterns and make measured performance diverge from reality.

  49. Question 49Foundations of AI systems, their use cases and impacts

    A model reports exceptional scores on a public benchmark, but its training corpus contained the benchmark's questions and answers. What is this problem and why does it matter?

    1. Underfitting; the model has learned too little from its data
    2. Regularization; it improves generalization on unseen data
    3. Contamination; the score overstates real capability Correct answer.
    4. Pruning; it removes redundant parameters from the network

    Explanation

    When evaluation items appear in training data, the benchmark measures memorization rather than generalization. Reported results then overstate capability and can mislead procurement, deployment, and safety decisions.

  50. Question 50Foundations of AI systems, their use cases and impacts

    In AI safety, what does the alignment problem refer to?

    1. Aligning the database schemas that feed the training pipeline
    2. Matching the size of the training and test datasets exactly
    3. Synchronizing GPU clocks across distributed training workers
    4. Keeping systems' behaviour consistent with intended goals Correct answer.

    Explanation

    Alignment concerns whether a system actually optimizes what its designers intended, including in situations not anticipated during development. It covers specification errors, reward hacking, and behaviour that looks correct in testing but diverges in deployment.

  51. Question 51AI laws, regulation and frameworks for responsible governance

    Under the EU AI Act, which of these AI practices is prohibited in principle?

    1. Using an AI system to answer customer queries through a chat interface
    2. Using an AI system to infer the emotions of employees in the workplace Correct answer.
    3. Using an AI system to detect defects on a production line
    4. Using an AI system to screen candidates during a recruitment process

    Explanation

    Article 5 bans emotion recognition in the workplace and in education institutions, except where used for medical or safety reasons. Recruitment screening is treated as high-risk and may be deployed subject to obligations, while chatbots are permitted with transparency and industrial quality inspection is minimal risk.

  52. Question 52AI laws, regulation and frameworks for responsible governance

    Which NIST AI Risk Management Framework function is cross-cutting and establishes policies, roles, responsibilities and a culture of risk management?

    1. Map
    2. Manage
    3. Govern Correct answer.
    4. Measure

    Explanation

    Govern is the cross-cutting function that is applied throughout the AI lifecycle and underpins the other three functions by assigning accountability, setting policy and fostering a risk-aware culture. Map frames context and identifies risks, Measure analyses and assesses them, and Manage prioritises and treats them.

  53. Question 53AI laws, regulation and frameworks for responsible governance

    Which principle appears first among the value-based OECD AI Principles adopted in 2019?

    1. Human-centred values and fairness
    2. Transparency and explainability
    3. Inclusive growth, sustainable development and well-being Correct answer.
    4. Accountability

    Explanation

    The OECD's five values-based principles begin with inclusive growth, sustainable development and well-being, followed by human rights and human-centred values, transparency and explainability, robustness, safety and security, and finally accountability.

  54. Question 54AI laws, regulation and frameworks for responsible governance

    Under the GDPR, what protection does a data subject have regarding a decision based solely on automated processing that produces legal effects concerning them?

    1. The right to receive the source code of the algorithm used
    2. An absolute right to have any automated decision reversed by a court
    3. The right not to be subject to the decision, subject to limited exceptions such as contract, legal authorisation or explicit consent Correct answer.
    4. The right to be notified within 24 hours of the decision being taken

    Explanation

    Article 22 prohibits solely automated decisions with legal or similarly significant effects unless they are necessary for a contract, authorised by law, or based on explicit consent, and in the latter two cases the controller must provide safeguards including human intervention and the ability to contest the decision.

  55. Question 55AI laws, regulation and frameworks for responsible governance

    New York City Local Law 144 requires employers using automated employment decision tools to do which of the following?

    1. Retain candidate data for a minimum of ten years
    2. Limit the use of the tool to candidates who have consented to algorithmic assessment
    3. Publish a bias audit conducted within the preceding 12 months together with a summary of results and data distribution Correct answer.
    4. Obtain prior approval from the city's Commission on Human Rights before each use

    Explanation

    Local Law 144 requires an independent bias audit performed no more than 12 months before use, publication of an audit summary including selection and scoring rates, and advance notice to candidates and employees at least ten business days before the tool is used.

  56. Question 56AI laws, regulation and frameworks for responsible governance

    The UNESCO Recommendation on the Ethics of Artificial Intelligence is significant because it was:

    1. A European Union regulation directly applicable in member states
    2. The first global standard-setting instrument on AI ethics, adopted by all 193 member states in 2021 Correct answer.
    3. The first legally binding treaty on AI, opened for signature in 2019
    4. A technical standard issued by ISO and mandatory for certification

    Explanation

    Adopted by acclamation at UNESCO's General Conference in November 2021, the Recommendation is the first globally agreed normative instrument on AI ethics, committing all member states to report periodically on implementation. It is a recommendation rather than a treaty, so it is not directly legally binding.

  57. Question 57AI laws, regulation and frameworks for responsible governance

    What is ISO/IEC 42001?

    1. A benchmark for measuring model accuracy
    2. A certifiable management system standard specifying requirements for an AI management system Correct answer.
    3. A data protection regulation for AI processing in the EU
    4. A risk management framework limited to generative AI systems

    Explanation

    ISO/IEC 42001:2023 specifies requirements for establishing, implementing, maintaining and continually improving an AI management system, and organisations can be independently audited and certified against it. It follows the harmonised structure common to other ISO management system standards, enabling integration with quality and information security systems.

  58. Question 58AI laws, regulation and frameworks for responsible governance

    How does the OECD define an AI system?

    1. A machine-based system that infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions Correct answer.
    2. A database system that stores and retrieves labelled training data
    3. A hardware device that is capable of autonomous physical movement and that interacts with its environment without any human intervention
    4. Any software that executes a fixed sequence of rules written by a programmer

    Explanation

    The OECD definition, adopted in 2019 and revised in 2023, describes an AI system as a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs and that operates with varying levels of autonomy and adaptiveness. This definition underpins the EU AI Act and several national frameworks.

  59. Question 59AI laws, regulation and frameworks for responsible governance

    Under the EU AI Act, a general-purpose AI model is presumed to have systemic risk when the cumulative computation used for its training exceeds which threshold?

    1. 10^20 floating point operations
    2. 10^25 floating point operations Correct answer.
    3. 10^27 floating point operations
    4. 10^23 floating point operations

    Explanation

    Article 51 creates a rebuttable presumption of systemic risk where the cumulative amount of computation used for training exceeds 10^25 floating point operations. The Commission may also designate a model as having systemic risk on the basis of Annex XIII criteria such as the number of users or the model's capabilities.

  60. Question 60AI laws, regulation and frameworks for responsible governance

    What is the maximum administrative fine under the EU AI Act for putting a prohibited AI practice into service?

    1. €10,000,000 or 2% of total worldwide annual turnover
    2. €35,000,000 or 7% of total worldwide annual turnover Correct answer.
    3. €15,000,000 or 3% of total worldwide annual turnover
    4. €20,000,000 or 4% of total worldwide annual turnover

    Explanation

    Article 99 sets three tiers: up to €35 million or 7% of worldwide annual turnover for prohibited practices, up to €15 million or 3% for most other obligations including high-risk requirements and general-purpose AI duties, and up to €7.5 million or 1% for supplying incorrect information.

  61. Question 61AI laws, regulation and frameworks for responsible governance

    Under Article 50 of the EU AI Act, what obligation applies to providers and deployers of systems that generate image, audio or video content constituting a deepfake?

    1. The generated content must be deleted within 30 days of generation
    2. Consent must be obtained from every identifiable person depicted in the output before any release of the content
    3. Use must be restricted to professional media organizations and licensed newsrooms
    4. Outputs must be marked machine-readably and the deepfake disclosed as artificially generated or manipulated Correct answer.

    Explanation

    Article 50 requires providers of synthetic content to mark outputs in a machine-readable format so that they are detectable as artificially generated, and deployers of deepfakes to disclose that the content has been artificially generated or manipulated, subject to limited exceptions such as artistic and law enforcement use.

  62. Question 62AI laws, regulation and frameworks for responsible governance

    Before placing a high-risk AI system on the EU market, a provider must generally take which set of steps?

    1. Complete a conformity assessment, issue an EU declaration of conformity, affix CE marking and register the system Correct answer.
    2. Certify the system to ISO/IEC 27001 through an accredited certification body and file the certificate with a national supervisory authority
    3. Obtain approval from the European Parliament and publish the full training dataset
    4. Notify the European AI Office in advance, pay a registration fee and obtain an operating licence

    Explanation

    Chapter III requires providers to carry out the applicable conformity assessment (usually internal control under Annex VI), draw up an EU declaration of conformity, affix the CE marking, and register the system in the EU database before placing it on the market. Notified body involvement applies only in specified cases such as biometric identification systems.

  63. Question 63AI laws, regulation and frameworks for responsible governance

    What is the Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law?

    1. The first legally binding international treaty on AI, opened for signature in September 2024 Correct answer.
    2. A non-binding set of ethical guidelines adopted by the G20
    3. A United Nations Security Council resolution on autonomous weapons
    4. An EU regulation that supersedes the AI Act for public authorities and that is directly applicable in all member states

    Explanation

    Opened for signature in Vilnius in September 2024, the Framework Convention is the first binding international treaty on AI, requiring parties to adopt measures ensuring that AI activities are consistent with human rights, democracy and the rule of law, with principles covering transparency, oversight, accountability and risk management.

  64. Question 64AI laws, regulation and frameworks for responsible governance

    China's Interim Measures for the Management of Generative Artificial Intelligence Services require providers to do which of the following?

    1. Restrict generative AI services to government agencies and state-owned enterprises
    2. Meet training-data and labelling duties and pass a security assessment where public opinion capacity is involved Correct answer.
    3. Register every model with the US National Institute of Standards and Technology before any public release
    4. Publish model weights openly as a condition of market access and disclose the full pre-training corpus

    Explanation

    The Interim Measures impose duties covering the lawfulness of training data and intellectual property, content aligned with core socialist values, labelling of generated content, real-identity verification of users, and a security assessment plus algorithm filing for services with public opinion attributes or social mobilisation capacity.

  65. Question 65AI laws, regulation and frameworks for responsible governance

    Colorado's Artificial Intelligence Act imposes which core obligation on developers and deployers of high-risk AI systems?

    1. Mandatory open-sourcing of all training data used for the model or service
    2. A strict liability standard for any harm caused by the system, regardless of fault, intent or the size of the deployer
    3. Reasonable care to protect consumers from foreseeable algorithmic discrimination, backed by risk management and impact assessments Correct answer.
    4. Prior approval by the state legislature of each model release and of every subsequent update

    Explanation

    The Colorado statute adopts a reasonable-care duty to protect consumers from algorithmic discrimination in consequential decisions, supported by risk management programmes, impact assessments, disclosure to consumers before a consequential decision, and reporting to the Attorney General.

  66. Question 66AI laws, regulation and frameworks for responsible governance

    Under the GDPR, when is a data protection impact assessment required for an AI system?

    1. When processing is likely to result in high risk to individuals' rights and freedoms Correct answer.
    2. Only when a data subject has formally objected to the processing or requested erasure
    3. Only when the processing is carried out by a public authority or body
    4. Whenever any personal data is processed at all, regardless of the level of risk involved

    Explanation

    Article 35 requires a DPIA where a type of processing, particularly using new technologies, is likely to result in a high risk to rights and freedoms. Supervisory authorities list triggers including systematic and extensive evaluation of personal aspects, large-scale processing of special category data, and large-scale systematic monitoring.

  67. Question 67AI laws, regulation and frameworks for responsible governance

    How does the United Kingdom currently regulate AI at the national level?

    1. Through a sector-based approach in which existing regulators apply five cross-sectoral principles Correct answer.
    2. Through a single comprehensive AI statute modelled closely on the EU AI Act and enforced by a new AI agency
    3. Through voluntary industry codes alone, with no government policy involvement
    4. Through an independent AI authority empowered to license and ban models

    Explanation

    The UK's 2023 AI regulation white paper set out a context-based, sector-led approach in which existing regulators apply five cross-sectoral principles: safety, security and robustness; transparency and explainability; fairness; accountability and governance; and contestability and redress. Government has instead pursued targeted legislation rather than a single AI act.

  68. Question 68AI laws, regulation and frameworks for responsible governance

    Under the EU AI Act, which deployers must carry out a fundamental rights impact assessment before first use of certain high-risk AI systems?

    1. Only small and medium-sized enterprises that employ more than 250 people
    2. All private companies deploying any high-risk AI system, in any sector and of any size
    3. Only providers established outside the European Union
    4. Public bodies, entities providing public services and deployers of credit-scoring or insurance risk systems Correct answer.

    Explanation

    Article 27 requires a fundamental rights impact assessment from bodies governed by public law, private entities providing public services, and deployers of high-risk systems used for creditworthiness assessment or credit scoring and for life and health insurance risk assessment. The assessment must be completed before the first use of the system.

  69. Question 69AI laws, regulation and frameworks for responsible governance

    The US Blueprint for an AI Bill of Rights (2022) set out five principles. Which of the following is one of them?

    1. Maximum automation
    2. Data maximization
    3. Notice and explanation Correct answer.
    4. Algorithmic determinism

    Explanation

    The Blueprint's five principles are safe and effective systems; algorithmic discrimination protections; data privacy; notice and explanation; and human alternatives, consideration and fallback. It is a non-binding statement of principles rather than enforceable law.

  70. Question 70AI laws, regulation and frameworks for responsible governance

    Singapore's Model AI Governance Framework is best characterized as which of the following?

    1. A treaty obligation binding on all ASEAN member states and their agencies
    2. A voluntary, principle-based framework, later extended with generative AI guidance Correct answer.
    3. A certification scheme administered by the European AI Office and notified bodies
    4. A legally binding regulation backed by financial penalties and inspections

    Explanation

    Singapore has pursued a voluntary, framework-based approach, first published in 2020 and supplemented in 2024 by the Model AI Governance Framework for Generative AI, which addresses accountability, data, trusted development, incident reporting, testing and content provenance. Adoption is encouraged rather than legally compelled.

  71. Question 71AI laws, regulation and frameworks for responsible governance

    What does ISO/IEC 23894 provide?

    1. Guidance on managing AI risk, aligned with the ISO 31000 risk management process Correct answer.
    2. Requirements for certifying an AI management system
    3. A mandatory template for reporting AI incidents to EU authorities
    4. A benchmark suite for evaluating large language model accuracy

    Explanation

    ISO/IEC 23894:2023 gives guidance on AI risk management for organisations that develop, produce, deploy or use AI products and services, integrating AI-specific risk sources into the ISO 31000 process. It is guidance rather than certifiable requirements, which are set out in ISO/IEC 42001.

  72. Question 72AI laws, regulation and frameworks for responsible governance

    Under the EU AI Act, an AI system listed in Annex III is not considered high-risk where it poses no significant risk of harm and performs which type of task?

    1. Any task involving personal data for which the data subject has given explicit consent
    2. A narrow procedural task, an improvement to a completed human activity, or pattern detection without replacing human judgement Correct answer.
    3. Any task that has been reviewed in advance by a notified body and recorded in the EU database before the system is deployed
    4. Any task performed exclusively inside a research laboratory or regulatory sandbox

    Explanation

    Article 6(3) disapplies the high-risk classification for Annex III systems that perform a narrow procedural task, improve the result of a previously completed human activity, detect decision-making patterns without replacing human assessment, or perform a preparatory task. The derogation is never available where the system performs profiling of natural persons.

  73. Question 73AI laws, regulation and frameworks for responsible governance

    Which situation falls within the territorial scope of the EU AI Act even though the provider is established outside the Union?

    1. A US provider whose system is used only by customers located in the United States
    2. A US provider whose AI system's output is used in the Union Correct answer.
    3. A US provider that employs EU nationals at its headquarters but serves no EU customers
    4. A US provider that trains its model on EU-language data but offers the system only outside the EU

    Explanation

    Article 2 extends the Act to providers and deployers established in third countries where the output produced by the AI system is used in the Union. Training on EU-sourced data or having EU employees does not by itself bring a provider within scope.

  74. Question 74AI laws, regulation and frameworks for responsible governance

    Under the EU AI Act's staged application, when do the obligations for high-risk AI systems listed in Annex III become applicable?

    1. 2 August 2025
    2. 2 August 2027
    3. 2 August 2026 Correct answer.
    4. 2 February 2025

    Explanation

    The Act entered into force on 1 August 2024; prohibitions applied from 2 February 2025, general-purpose AI obligations from 2 August 2025, and the high-risk regime for Annex III systems from 2 August 2026, with Annex I product-safety systems following on 2 August 2027.

  75. Question 75AI laws, regulation and frameworks for responsible governance

    How does NIST AI 600-1, the Generative AI Profile, relate to the NIST AI Risk Management Framework 1.0?

    1. It creates binding legal obligations for US federal agencies
    2. It adds generative-AI-specific risks and suggested actions mapped to the existing RMF functions Correct answer.
    3. It replaces the four RMF functions with a single generative AI checklist and scoring rubric
    4. It certifies generative AI models against a fixed accuracy benchmark

    Explanation

    AI 600-1 is a cross-sectoral profile that supplements the AI RMF rather than replacing it: it catalogues generative AI risk categories and suggests actions organised under the existing Govern, Map, Measure and Manage functions. It is voluntary guidance and does not itself create legal duties.

  76. Question 76AI laws, regulation and frameworks for responsible governance

    Which set lists the four core functions of the NIST AI Risk Management Framework 1.0?

    1. Identify, Protect, Detect, Respond
    2. Plan, Do, Check, Act
    3. Govern, Map, Measure, Manage Correct answer.
    4. Scope, Assess, Treat, Monitor

    Explanation

    The NIST AI RMF organizes AI risk management around four core functions: Govern, Map, Measure, and Manage. Govern is cross-cutting, while Map, Measure, and Manage address context, assessment, and treatment of risks.

  77. Question 77AI laws, regulation and frameworks for responsible governance

    Which statement best captures the OECD's recommended approach to artificial intelligence?

    1. AI should be deployed only after binding government certification.
    2. AI development should be governed solely by private contracts.
    3. AI should be human-centred, promote inclusive growth and sustainable development, and be transparent, robust, and accountable. Correct answer.
    4. AI should be regulated only through ex post criminal liability.

    Explanation

    The OECD AI Principles recommend that AI be human-centred and promote inclusive growth, sustainable development, and well-being, while respecting human rights and democratic values. They also call for transparency, robustness, safety, security, and accountability.

  78. Question 78AI laws, regulation and frameworks for responsible governance

    The EU Artificial Intelligence Act primarily regulates AI systems through which structure?

    1. Voluntary labeling with no legal obligations
    2. Mandatory open-source licensing for all models
    3. A risk-based tiered system with prohibited practices and high-risk obligations Correct answer.
    4. Sector-specific self-regulation by industry bodies

    Explanation

    The EU AI Act uses a risk-based approach. It prohibits certain unacceptable practices, imposes extensive requirements on high-risk AI systems, and applies transparency or minimal obligations to other systems.

  79. Question 79AI laws, regulation and frameworks for responsible governance

    Under the GDPR, an individual has a right not to be subject to a decision based solely on automated processing when that decision does what?

    1. Is made by any AI system
    2. Involves any personal data
    3. Involves any profiling activity
    4. Produces legal effects or similarly significantly affects the individual Correct answer.

    Explanation

    Article 22 GDPR protects individuals from decisions based solely on automated processing that produce legal effects or similarly significantly affect them. The right is subject to limited exceptions, such as contractual necessity or explicit consent.

  80. Question 80AI laws, regulation and frameworks for responsible governance

    What does New York City Local Law 144 require of employers that use automated employment decision tools?

    1. Conduct a data protection impact assessment for every job applicant.
    2. Commission an annual independent bias audit and provide notice to candidates. Correct answer.
    3. Obtain an EU conformity assessment before use.
    4. Publish the complete source code of the tool.

    Explanation

    NYC Local Law 144 requires covered employers and employment agencies to obtain an annual independent bias audit and provide required notice to candidates and employees. The law targets automated tools used to substantially assist employment decisions.

  81. Question 81AI laws, regulation and frameworks for responsible governance

    What is ISO/IEC 22989?

    1. A standard for AI risk management
    2. A standard providing AI concepts and terminology Correct answer.
    3. An AI management system certification standard
    4. A standard for auditing AI management systems

    Explanation

    ISO/IEC 22989 establishes a vocabulary and concepts for artificial intelligence. It helps ensure consistent terminology across AI standards and governance work.

  82. Question 82AI laws, regulation and frameworks for responsible governance

    What is the G7 Hiroshima Process International Code of Conduct for Organizations Developing Advanced AI Systems?

    1. A certification scheme administered by the United Nations
    2. A voluntary code of conduct for advanced AI developers Correct answer.
    3. A technical standard for neural network robustness
    4. A binding treaty that imposes fines on AI developers

    Explanation

    The G7 Hiroshima Process International Code of Conduct is a voluntary code for organizations developing advanced AI systems. It sets out responsible practices but does not create binding legal obligations.

  83. Question 83AI laws, regulation and frameworks for responsible governance

    Singapore's AI Verify is best described as which of the following?

    1. An international treaty on AI and human rights
    2. A binding statute regulating high-risk AI
    3. A voluntary AI governance testing framework and toolkit Correct answer.
    4. A data protection authority for AI systems

    Explanation

    AI Verify is a voluntary testing framework and toolkit intended to help organizations demonstrate responsible AI governance. It is not a binding law or international treaty.

  84. Question 84AI laws, regulation and frameworks for responsible governance

    Which practice is prohibited under Article 5 of the EU AI Act?

    1. An AI system that answers routine customer-service questions.
    2. An AI system used by public authorities to socially score individuals and impose detrimental treatment unrelated to the context in which the data were originally generated. Correct answer.
    3. An AI system that assesses a consumer's creditworthiness before a loan is granted.
    4. An AI system that assists a physician in interpreting a medical image.

    Explanation

    Article 5 prohibits social scoring by public authorities, or on their behalf, where it leads to detrimental or unfavorable treatment unrelated to the context of the original data or unjustified by the behavior. Credit scoring and medical AI are generally high-risk rather than prohibited.

  85. Question 85AI laws, regulation and frameworks for responsible governance

    What baseline obligations apply to a provider of any general-purpose AI model under the EU AI Act?

    1. Maintain technical documentation, give downstream providers information, and publish a copyright-compliance policy and training-content summary. Correct answer.
    2. Register every model in the EU database before placing it on the market.
    3. Obtain prior authorization from the European AI Office before releasing the model.
    4. Conduct a clinical evaluation and affix a CE marking.

    Explanation

    All providers of general-purpose AI models have baseline obligations under Article 53, including technical documentation, information for downstream providers, a copyright-compliance policy, and a summary of training content. Additional obligations apply if the model presents systemic risk.

  86. Question 86AI laws, regulation and frameworks for responsible governance

    What must a deployer of a high-risk AI system generally do under the EU AI Act?

    1. Carry out the conformity assessment and establish a quality management system.
    2. Register the system in the EU database before first use.
    3. Affix the CE marking and draw up the EU declaration of conformity.
    4. Use the system according to instructions, assign human oversight, ensure input data are relevant, and monitor operation and logs. Correct answer.

    Explanation

    Deployers of high-risk AI systems have obligations under Article 26, including using the system according to instructions, assigning competent human oversight, ensuring input data are relevant, and monitoring operation and logs. Provider obligations such as CE marking and conformity assessment generally fall on the provider.

  87. Question 87AI laws, regulation and frameworks for responsible governance

    What does Article 4 of the EU AI Act require regarding AI literacy?

    1. Take measures to ensure sufficient AI literacy among staff and other persons dealing with AI on their behalf. Correct answer.
    2. Appoint a statutory AI officer in every company.
    3. Publish the training data used by all AI systems.
    4. Certify all employees in AI ethics through a notified body.

    Explanation

    Article 4 requires providers and deployers to take measures to ensure a sufficient level of AI literacy among their staff and other persons dealing with AI systems on their behalf. The obligation is tailored to the context and does not require formal certification.

  88. Question 88AI laws, regulation and frameworks for responsible governance

    What does the EU AI Act require a provider of a high-risk AI system to include in its quality management system?

    1. Only a marketing plan describing intended customers.
    2. Documented risk management, data governance, testing, and post-market monitoring arrangements. Correct answer.
    3. Only an open-source licence for the model weights.
    4. Only a financial audit of the provider's accounts.

    Explanation

    Article 17 requires providers of high-risk AI systems to establish a quality management system covering areas such as risk management, data governance, testing, incident reporting, and post-market monitoring. The system must be documented and maintained.

  89. Question 89AI laws, regulation and frameworks for responsible governance

    For a serious incident under the EU AI Act that does not fall into the shorter reporting categories, by when must a provider report it to the market surveillance authorities?

    1. No later than 15 days after establishing a causal link or reasonable possibility Correct answer.
    2. Within six months
    3. Within 30 days
    4. Within 72 hours

    Explanation

    Article 73 generally requires providers of high-risk AI systems to report serious incidents no later than 15 days after they establish a causal link or a reasonable possibility that the AI system caused the incident. Shorter deadlines apply to specified severe incidents.

  90. Question 90AI laws, regulation and frameworks for responsible governance

    The Govern function in the NIST AI Risk Management Framework is primarily concerned with which of the following?

    1. Treating identified risks and monitoring residual risk
    2. Cultivating a risk-management culture, assigning accountability, and establishing policies and oversight Correct answer.
    3. Measuring the accuracy and robustness of a specific model
    4. Mapping the context and intended use of an AI system

    Explanation

    Govern is a cross-cutting function that establishes the culture, policies, accountability, and oversight needed for AI risk management. It supports and connects the Map, Measure, and Manage functions.

  91. Question 91AI laws, regulation and frameworks for responsible governance

    What does NIST AI 100-2 provide?

    1. A generative AI risk profile
    2. A binding international treaty on AI
    3. A taxonomy of adversarial machine learning attacks and mitigations Correct answer.
    4. An AI management system certification standard

    Explanation

    NIST AI 100-2 is a taxonomy and terminology resource on adversarial machine learning. It describes attack types, attacker goals, and mitigation approaches.

  92. Question 92AI laws, regulation and frameworks for responsible governance

    When personal data are used in solely automated decision-making, what must a GDPR controller provide to the data subject?

    1. A list of every data subject whose data were used.
    2. The identity of the AI vendor's shareholders.
    3. Meaningful information about the logic involved, as well as the significance and envisaged consequences. Correct answer.
    4. The complete source code and model weights.

    Explanation

    GDPR Articles 13 to 15 require the controller to provide meaningful information about the logic involved in solely automated decision-making, along with the significance and envisaged consequences. This supports transparency without requiring disclosure of trade secrets or source code.

  93. Question 93AI laws, regulation and frameworks for responsible governance

    What do China's Algorithmic Recommendation Management Provisions require of service providers in certain cases?

    1. Obtain a UNESCO ethics certificate before launch.
    2. Publish all model weights as open source.
    3. File algorithm-related information with authorities where the service has public-opinion or social-mobilization attributes. Correct answer.
    4. Prohibit any personalized recommendation.

    Explanation

    China's Algorithmic Recommendation Management Provisions require providers of services with public-opinion or social-mobilization attributes to file algorithm-related information with authorities. The provisions also include user notice and opt-out requirements.

  94. Question 94AI laws, regulation and frameworks for responsible governance

    What do China's Deep Synthesis Provisions require of providers?

    1. Obtain prior consent from every person depicted in synthetic media.
    2. Ban all synthetic audio and video.
    3. Register each model with the EU AI Office.
    4. Label generated content so users can identify it as deep synthesis and refrain from producing fake news. Correct answer.

    Explanation

    The Deep Synthesis Provisions require providers to mark synthetic content and to prevent the service from being used to produce or spread false news. They also impose identity verification and related governance duties.

  95. Question 95AI laws, regulation and frameworks for responsible governance

    Under the EU Digital Services Act, what must very large online platforms do regarding systemic risks?

    1. Ban all recommender systems on their platforms.
    2. Obtain prior approval from the European Commission for each algorithm.
    3. Carry out annual risk assessments, independent audits, and transparency reporting on systemic risks, including algorithmic amplification. Correct answer.
    4. Publish their complete source code.

    Explanation

    The DSA requires very large online platforms and search engines to assess and mitigate systemic risks, undergo independent audits, and report transparently. Algorithmic amplification and recommender systems are expressly within the scope of those risk assessments.

  96. Question 96AI laws, regulation and frameworks for responsible governance

    What does Article 25 of the GDPR require controllers to do?

    1. Conduct a data protection impact assessment for every AI system.
    2. Implement appropriate technical and organizational measures and safeguards by design and by default. Correct answer.
    3. Appoint a data protection officer in every organization.
    4. Obtain explicit consent for all processing.

    Explanation

    Article 25 requires data protection by design and by default, meaning controllers must implement appropriate technical and organizational measures and safeguards. These measures must be proportionate to the risks and applied throughout the processing lifecycle.

  97. Question 97AI laws, regulation and frameworks for responsible governance

    What additional obligations apply to a provider of a general-purpose AI model with systemic risk under the EU AI Act?

    1. Perform model evaluation and adversarial testing, assess and mitigate systemic risks, report serious incidents, and ensure cybersecurity. Correct answer.
    2. Only register a trademark in the EU.
    3. Only notify users that the model is generative.
    4. Only publish a public model card.

    Explanation

    Article 55 imposes additional obligations on providers of general-purpose AI models with systemic risk. These include model evaluation, adversarial testing, systemic-risk assessment and mitigation, serious incident reporting, and cybersecurity measures.

  98. Question 98AI laws, regulation and frameworks for responsible governance

    Under the EU AI Act, what happens when a distributor, importer, or deployer substantially modifies a high-risk AI system?

    1. It may continue under the original provider's conformity assessment without any new obligations.
    2. It must immediately withdraw the system from the market.
    3. It becomes a provider and must meet the provider's obligations for that high-risk AI system. Correct answer.
    4. It is always exempt from the EU AI Act.

    Explanation

    Article 25 provides that a distributor, importer, deployer, or other third party that substantially modifies a high-risk AI system, or changes its intended purpose, becomes a provider. It must then comply with the provider obligations for that system.

  99. Question 99AI laws, regulation and frameworks for responsible governance

    How does the revised EU Product Liability Directive address AI?

    1. It limits liability to physical robots.
    2. It creates a fault-based regime only for medical AI.
    3. It exempts AI from product liability because it is intangible.
    4. It treats software, including AI, as a product and introduces presumptions and disclosure duties for complex products. Correct answer.

    Explanation

    The revised Product Liability Directive explicitly covers software, including AI systems, as a product. It also introduces mechanisms such as presumptions of defectiveness or causation and disclosure duties to help claimants in complex technical cases.

  100. Question 100AI laws, regulation and frameworks for responsible governance

    Under the EU AI Act, what is a notified body?

    1. An independent third party designated to assess the conformity of certain high-risk AI systems. Correct answer.
    2. An internal team of the provider that self-certifies all systems.
    3. A national court that imposes administrative fines.
    4. A unit of the European Commission that drafts delegated acts.

    Explanation

    Notified bodies are independent conformity assessment bodies designated under the EU AI Act. They assess certain high-risk AI systems where third-party conformity assessment is required.

  101. Question 101AI life cycle and AI risk management

    The NIST AI Risk Management Framework organizes its activities into four core functions. Which set lists them correctly?

    1. Identify, Protect, Detect, Respond
    2. Govern, Map, Measure, Manage Correct answer.
    3. Scope, Assess, Treat, Monitor
    4. Plan, Do, Check, Act

    Explanation

    The AI RMF defines four functions: Govern (culture, accountability, and oversight), Map (context and risk framing), Measure (analysis, testing, and monitoring), and Manage (prioritization and treatment). Govern is cross-cutting and informs the other three. The functions are meant to be applied iteratively across the AI life cycle.

  102. Question 102AI life cycle and AI risk management

    What is the primary purpose of a model card?

    1. To list the infrastructure requirements for serving the model
    2. To summarize a model's intended use, performance, and limits Correct answer.
    3. To record the licensing terms under which training data was obtained
    4. To store the commit history of the source code repository

    Explanation

    Model cards are short, standardized disclosures that describe a model's intended use, evaluation results (including performance across subgroups), and known limitations. They help downstream users and reviewers judge whether the model is appropriate for a given context. They do not replace data documentation or code repositories.

  103. Question 103AI life cycle and AI risk management

    An organization buys cyber-insurance to cover losses from a generative AI system that could expose customer data. Which risk treatment option does this represent?

    1. Risk acceptance
    2. Risk avoidance
    3. Risk transfer Correct answer.
    4. Risk mitigation

    Explanation

    Risk transfer shifts the financial consequences of a risk to a third party, such as an insurer or a contractual counterparty, without removing the underlying exposure. Avoidance would mean not operating the system at all, and mitigation would mean reducing likelihood or impact through controls.

  104. Question 104AI life cycle and AI risk management

    In AI governance, what does 'human in the loop' describe?

    1. A person who audits the model on an annual cycle
    2. A person who labels the data used to train the model
    3. A person who reviews model output before a decision is final Correct answer.
    4. A person who maintains the infrastructure hosting the model

    Explanation

    Human in the loop refers to a designed point of human review or approval in the decision flow, so that a consequential outcome is not produced autonomously by the model. It is a control over automation, distinct from data labeling, operations, or periodic audit roles.

  105. Question 105AI life cycle and AI risk management

    A model's input features gradually shift in distribution after deployment while the relationship between inputs and outcomes stays the same. What is this called?

    1. Overfitting
    2. Data drift Correct answer.
    3. Concept drift
    4. Label leakage

    Explanation

    Data drift (also called covariate shift) is a change in the distribution of inputs, P(X), without a change in how inputs map to outcomes. Concept drift, by contrast, is a change in the underlying relationship P(Y|X). Both can degrade performance but call for different monitoring and remediation.

  106. Question 106AI life cycle and AI risk management

    What is the main function of a risk register in an AI risk management program?

    1. To track identified risks, owners, ratings, and treatment status Correct answer.
    2. To record an audit trail of every model prediction
    3. To list the regulatory filings required in each jurisdiction
    4. To document the technical architecture of each model in production

    Explanation

    A risk register is the system of record for identified risks, capturing descriptions, ownership, ratings, treatment plans, and status over time. It supports prioritization, reporting, and accountability as risks evolve across the life cycle.

  107. Question 107AI life cycle and AI risk management

    Which life cycle stage covers retiring a model, archiving its documentation, and removing its integrations?

    1. Decommissioning Correct answer.
    2. Validation
    3. Deployment
    4. Scoping

    Explanation

    Decommissioning is the planned end-of-life phase in which a model is withdrawn from service, dependent systems and users are migrated, and records are retained for audit and legal obligations. Treating it as a defined phase prevents orphaned integrations and lost documentation.

  108. Question 108AI life cycle and AI risk management

    After controls have been applied to an AI system, the risk that remains is called:

    1. Risk appetite
    2. Control risk
    3. Residual risk Correct answer.
    4. Inherent risk

    Explanation

    Residual risk is what remains after treatments and controls are in place, and it is the figure compared against the organization's risk appetite. Inherent risk is the exposure before any controls are applied.

  109. Question 109AI life cycle and AI risk management

    A product team plans a model that will prioritize customer service tickets. Before any data is collected, which activity best reflects good life cycle practice?

    1. Training a baseline model to see what accuracy is achievable
    2. Selecting the cloud vendor that will host the model
    3. Documenting intended use, stakeholders, and success criteria Correct answer.
    4. Running a full red-team exercise against the model

    Explanation

    The framing stage establishes purpose, affected parties, constraints, and acceptance criteria before data and modeling work begins. These artifacts anchor later risk assessment, testing design, and oversight decisions. Jumping to modeling or procurement early makes risks harder to identify and remediate.

  110. Question 110AI life cycle and AI risk management

    Which characteristic most strongly justifies classifying an AI system as high risk in a governance program?

    1. It was built with an open-source library
    2. It uses more than one billion parameters
    3. It materially affects access to credit, employment, or safety Correct answer.
    4. It is retrained more than once per quarter

    Explanation

    Risk classification is driven by the severity and nature of potential harm to individuals and society, especially in domains such as credit, employment, health, and physical safety. Technical attributes like parameter count or update cadence are relevant to engineering but are not the primary determinant of harm-based tiering.

  111. Question 111AI life cycle and AI risk management

    An organization must evaluate how a new AI system could affect community trust and accessibility, in addition to personal data handling. Which instrument is most appropriate?

    1. A data processing impact assessment limited to personal data
    2. A penetration test of the model's API endpoints
    3. An AI impact assessment covering societal effects Correct answer.
    4. A vendor security questionnaire

    Explanation

    AI impact assessments examine a broader range of effects than privacy alone, including accessibility, fairness, and societal consequences for affected groups. Privacy-focused assessments and security testing address narrower questions and would not capture those wider impacts.

  112. Question 112AI life cycle and AI risk management

    Which arrangement best satisfies the principle of independent validation for a credit-scoring model?

    1. The vendor that supplied the training data certifies the results
    2. A function independent of development reviews the model and evidence Correct answer.
    3. The product owner approves the model based on business metrics
    4. The lead developer signs off after re-running the training script

    Explanation

    Independent validation requires that the reviewers have no stake in the model's approval and can challenge design choices and evidence objectively. Self-review or approval driven solely by commercial metrics undermines the credibility of the assurance.

  113. Question 113AI life cycle and AI risk management

    What is the main purpose of red teaming an AI system before launch?

    1. To surface failure modes that developers did not anticipate Correct answer.
    2. To estimate the infrastructure cost of serving model traffic
    3. To confirm the model meets its accuracy target on the holdout set
    4. To verify that the source code compiles in production

    Explanation

    Red teaming uses adversarial probing to find harmful, unsafe, or unexpected behaviors that standard testing may miss. It complements, but does not replace, performance evaluation against holdout data or operational readiness checks.

  114. Question 114AI life cycle and AI risk management

    Six months after launch, a recommendation model's accuracy declines because the customer base has changed. Which control would most reliably have detected this?

    1. Encryption of the model's training data at rest
    2. Continuous monitoring of performance and inputs with alert thresholds Correct answer.
    3. An annual third-party audit of the model's documentation
    4. A pre-deployment bias assessment

    Explanation

    Post-deployment degradation is detected through ongoing measurement of live performance and input distributions against defined thresholds that trigger investigation. Pre-deployment assessments and annual reviews are too infrequent or too early to catch a gradual shift.

  115. Question 115AI life cycle and AI risk management

    A company integrates a vendor's hosted language model into a customer-facing product. Which contract provision most directly supports ongoing AI risk management?

    1. A warranty that the model will never produce incorrect output
    2. A clause fixing the subscription price for three years
    3. Notification of material model changes and rights to audit or test Correct answer.
    4. A requirement that vendor staff be located in one country

    Explanation

    Because hosted models change without the customer's involvement, contractual notice of material updates plus audit or testing rights preserve visibility and control over third-party risk. Absolute accuracy warranties are not realistic for probabilistic systems and would not substitute for ongoing oversight.

  116. Question 116AI life cycle and AI risk management

    What does a risk appetite statement define for an AI program?

    1. The technical controls applied to each production model
    2. The maximum number of models the team may deploy
    3. The amount and type of risk the organization is willing to accept Correct answer.
    4. The list of risks identified during the last assessment

    Explanation

    Risk appetite expresses the level and categories of risk the organization will tolerate in pursuit of its objectives, and it sets the benchmark against which residual risk is judged. Risk tolerance then translates appetite into concrete thresholds for specific activities.

  117. Question 117AI life cycle and AI risk management

    A team retrains a production model on new data and wants to redeploy it. Which step is most consistent with sound change management?

    1. Regression and subgroup testing with documented sign-off Correct answer.
    2. Deploying directly to production and monitoring complaints
    3. Skipping documentation because the architecture is unchanged
    4. Retiring the old model without archiving its documentation

    Explanation

    A retrained model is a changed system even when the architecture is identical, so it should pass regression and subgroup testing and receive documented approval before release. This preserves the audit trail and prevents silent degradation of previously validated behavior.

  118. Question 118AI life cycle and AI risk management

    Which practice best supports evaluating whether a model treats groups equitably?

    1. Disaggregating error rates across relevant subgroups Correct answer.
    2. Removing all demographic fields without further analysis
    3. Reporting a single aggregate accuracy figure for the population
    4. Increasing the training set size until losses converge

    Explanation

    Equity evaluation requires comparing error rates and outcomes across relevant subgroups, since a strong aggregate metric can mask poor performance for a smaller group. Aggregate-only reporting and blind removal of demographic fields both conceal disparities rather than reveal them.

  119. Question 119AI life cycle and AI risk management

    Why should an AI program maintain dataset provenance and transformation records?

    1. To support auditability, reproducibility, and defect tracing Correct answer.
    2. To allow the model to be retrained more quickly
    3. To reduce the storage cost of the training data
    4. To satisfy licensing terms of open-source libraries

    Explanation

    Provenance and transformation records show where data came from and how it was altered, enabling auditors to reconstruct decisions, teams to reproduce results, and investigators to trace the source of errors. These are governance and assurance benefits rather than performance optimizations.

  120. Question 120AI life cycle and AI risk management

    Which approach is most effective for identifying AI risks early in the life cycle?

    1. Relying on the development team's internal checklist alone
    2. Scenario workshops with cross-functional stakeholders Correct answer.
    3. Reviewing only the model's accuracy metrics
    4. Waiting until an incident is reported in production

    Explanation

    Early risk identification benefits from diverse perspectives, including legal, security, domain experts, and affected users, applying structured scenarios to the proposed use. Narrow technical review or post-incident discovery misses harms that emerge from context and deployment.

  121. Question 121AI life cycle and AI risk management

    An organization has documented its AI use cases, mapped affected stakeholders, and catalogued risks, but has defined no metrics, test methods, or evaluation results. Which AI RMF function is weakest?

    1. Manage
    2. Map
    3. Govern
    4. Measure Correct answer.

    Explanation

    The Measure function covers the methods, metrics, and results used to analyze, benchmark, and monitor risks, so the absence of metrics and test methods is a gap there. Map has already been performed, since context, stakeholders, and risks are documented; Manage addresses prioritization and response, and Govern addresses policy and accountability.

  122. Question 122AI life cycle and AI risk management

    Loan officers approve nearly every application a scoring model recommends and rarely override it, even when supporting documents conflict with the score. Which mitigation addresses the root cause?

    1. Publish the model's overall accuracy to the public
    2. Retrain the model on a larger historical dataset
    3. Add a second model to average out disagreements
    4. Have reviewers assess cases before seeing the model score Correct answer.

    Explanation

    The behavior described is automation bias: reviewers anchor on the model's recommendation and stop evaluating evidence independently. Structuring the workflow so a human judgment is recorded before the score is revealed counters that anchoring effect. Model-side changes do not address the human reliance pattern.

  123. Question 123AI life cycle and AI risk management

    A model predicts which neighborhoods need extra patrols, and the resulting patrol data becomes training data for the next version. What risk does this pattern create?

    1. A feedback loop that reinforces prior deployment bias Correct answer.
    2. Overfitting to a small training set
    3. Label leakage from the validation split
    4. Concept drift from a changed outcome definition

    Explanation

    When a model's outputs shape the data later used to retrain it, the system can confirm its own prior predictions rather than learn about the underlying reality. This self-reinforcing loop entrenches whatever bias existed at first deployment and grows over successive versions.

  124. Question 124AI life cycle and AI risk management

    After treatment, an AI system's residual risk exceeds the stated risk appetite. What is the most appropriate governance response?

    1. Escalate to the accountable executive or risk committee Correct answer.
    2. Accept the residual risk because treatment has already been applied
    3. Re-run the original risk assessment using the same assumptions
    4. Document the exception and continue operating quietly

    Explanation

    Residual risk above appetite requires a decision by those with authority to hold that level of risk, which means escalation to an accountable executive or risk committee. Treatment completion does not itself legitimize exceeding appetite, and repeating an assessment with unchanged assumptions will not resolve the gap.

  125. Question 125AI life cycle and AI risk management

    An organization wants a certifiable, organization-wide management system for AI that can be audited alongside its other ISO management systems. Which reference is most appropriate?

    1. ISO/IEC 42001 Correct answer.
    2. ISO/IEC 23894
    3. The NIST AI Risk Management Framework
    4. ISO/IEC 25059

    Explanation

    ISO/IEC 42001 specifies requirements for an AI management system, including risk assessment, impact assessment, and continual improvement, and is structured to integrate with other ISO management-system standards for certification. The other references offer guidance, voluntary frameworks, or quality models rather than certifiable management-system requirements.

  126. Question 126AI life cycle and AI risk management

    Which AI life cycle stage involves specifying intended use, affected stakeholders, and success criteria before data is collected?

    1. Problem definition Correct answer.
    2. Model training
    3. Data acquisition
    4. Deployment

    Explanation

    Problem definition establishes why the system is needed, who is affected, and what success looks like. These choices constrain later data, modeling, and deployment decisions.

  127. Question 127AI life cycle and AI risk management

    What is the primary purpose of a model inventory in an AI governance program?

    1. To catalogue AI systems and risks Correct answer.
    2. To replace the need for periodic monitoring
    3. To calculate the salary of data scientists
    4. To store the training data for every model

    Explanation

    A model inventory is a central record of AI systems, owners, purposes, and risk attributes. It supports oversight, reporting, and consistent risk tiering.

  128. Question 128AI life cycle and AI risk management

    In AI risk management, what does inherent risk mean?

    1. The risk before any controls are applied Correct answer.
    2. The risk that remains after controls are applied
    3. The risk that a model will be retrained
    4. The risk that a vendor will breach a contract

    Explanation

    Inherent risk is assessed before mitigating controls so the organization can see the raw exposure. Residual risk is what remains after controls.

  129. Question 129AI life cycle and AI risk management

    Which risk treatment option means declining to deploy an AI use case because its risk is unacceptable?

    1. Risk transfer
    2. Risk mitigation
    3. Risk avoidance Correct answer.
    4. Risk acceptance

    Explanation

    Risk avoidance eliminates the exposure by not undertaking the activity. Mitigation reduces it, transfer shifts it, and acceptance retains it.

  130. Question 130AI life cycle and AI risk management

    Which artifact typically summarizes a model's intended use, limitations, and performance for downstream users?

    1. A data retention schedule
    2. A model card Correct answer.
    3. A service level agreement
    4. A source code repository

    Explanation

    A model card is a concise disclosure of a model's purpose, limitations, performance, and appropriate uses. It helps downstream users interpret and govern the model.

  131. Question 131AI life cycle and AI risk management

    What is the main purpose of a key risk indicator in an AI risk program?

    1. To signal early that a risk is rising Correct answer.
    2. To replace all qualitative risk assessments
    3. To record the final approval of a model
    4. To document the model's source code

    Explanation

    A key risk indicator is a measurable signal that a risk is approaching an undesirable level. It supports timely action before losses occur.

  132. Question 132AI life cycle and AI risk management

    Which life cycle activity checks whether a deployed model still performs as expected?

    1. Procurement
    2. Data labeling
    3. Monitoring Correct answer.
    4. Problem definition

    Explanation

    Monitoring compares deployed behavior and performance against expectations over time. It can trigger investigation, retraining, or rollback.

  133. Question 133AI life cycle and AI risk management

    What is the main purpose of human oversight in an AI system?

    1. To allow meaningful intervention when needed Correct answer.
    2. To increase the model's training speed
    3. To remove all human involvement from decisions
    4. To reduce the need for data governance

    Explanation

    Human oversight provides a check when the system is uncertain, errs, or faces an edge case. It must be meaningful, not merely nominal.

  134. Question 134AI life cycle and AI risk management

    Before contracting a vendor's hosted AI service, which practice best addresses third-party risk?

    1. Rely on the vendor's marketing claims
    2. Wait until after launch to request audit reports
    3. Review the vendor's controls and audit evidence Correct answer.
    4. Skip review because the vendor is large

    Explanation

    Third-party risk management requires evidence about the vendor's security, privacy, and AI controls before reliance. Audit reports and control documentation support that assessment.

  135. Question 135AI life cycle and AI risk management

    Why do risk assessments combine likelihood and impact ratings?

    1. To rank risks by expected severity Correct answer.
    2. To prove that all risks are equally important
    3. To calculate the model's training cost
    4. To eliminate the need for risk treatment

    Explanation

    Combining likelihood and impact produces a severity estimate that helps compare and prioritize risks. Prioritization then guides treatment and oversight.

  136. Question 136AI life cycle and AI risk management

    Which statement best distinguishes verification from validation?

    1. Verification applies only to data; validation only to code
    2. Verification and validation are identical activities
    3. Verification meets specifications; validation meets intended use Correct answer.
    4. Verification confirms user needs; validation confirms specifications

    Explanation

    Verification asks whether the system was built according to its requirements. Validation asks whether it satisfies the intended use and user needs.

  137. Question 137AI life cycle and AI risk management

    A team uses synthetic data to supplement rare cases in training. Which risk should be assessed?

    1. The data will require no privacy review
    2. Synthetic records may distort real distributions Correct answer.
    3. The model will definitely become more accurate
    4. The model will no longer need validation

    Explanation

    Synthetic data can improve coverage but may not faithfully represent real cases. Teams should assess fidelity, provenance, and possible bias amplification.

  138. Question 138AI life cycle and AI risk management

    After an AI system produces harmful outputs in production, what should be done first?

    1. Contain impact by disabling or rolling back Correct answer.
    2. Retrain the model on all available data
    3. Publish a detailed root-cause report
    4. Expand the system's scope to gather more evidence

    Explanation

    Incident response first limits harm to users and the organization. Investigation, root-cause analysis, and remediation follow containment.

  139. Question 139AI life cycle and AI risk management

    Which practice best supports traceability of changes to an AI model?

    1. Versioning code, data, and models Correct answer.
    2. Storing only the final model weights
    3. Relying on oral handover between teams
    4. Deleting old versions to save storage

    Explanation

    Traceability requires linking each model version to the code, data, and configuration that produced it. Version control makes changes reproducible and auditable.

  140. Question 140AI life cycle and AI risk management

    What should a model risk tier determine?

    1. The depth of validation and monitoring Correct answer.
    2. The vendor's marketing budget
    3. The color scheme of the model dashboard
    4. The number of data scientists assigned

    Explanation

    Risk tiering scales governance proportional to potential harm and uncertainty. Higher tiers typically require more validation, documentation, and monitoring.

  141. Question 141AI life cycle and AI risk management

    What is the main purpose of control testing in an AI risk management program?

    1. To calculate the model's final accuracy
    2. To approve the model without further review
    3. To replace the need for stakeholder input
    4. To verify controls work as designed Correct answer.

    Explanation

    Control testing provides evidence that a safeguard operates consistently and effectively. It is distinct from testing model accuracy.

  142. Question 142AI life cycle and AI risk management

    Which mechanism transfers some AI risk to a third party?

    1. A data dictionary
    2. A model card
    3. A monitoring dashboard
    4. An insurance policy Correct answer.

    Explanation

    Risk transfer shifts some consequences to another party, often through insurance or contractual arrangements. The underlying risk may remain but its financial impact is shared.

  143. Question 143AI life cycle and AI risk management

    A control reduces the likelihood of an AI risk but leaves its impact unchanged. What happens to residual risk?

    1. Residual risk increases because controls add complexity
    2. Residual risk becomes zero
    3. Residual risk is unaffected by the control
    4. It falls but keeps the same impact Correct answer.

    Explanation

    Residual risk reflects the remaining likelihood and impact after controls. A likelihood-only control lowers the frequency estimate while the consequence estimate stays the same.

  144. Question 144AI life cycle and AI risk management

    A risk exceeds the organization's risk appetite and no effective control is available. What treatment is most appropriate?

    1. Monitor the risk without any action
    2. Transfer the risk through a model card
    3. Avoid the use case or stop it Correct answer.
    4. Accept the risk without documentation

    Explanation

    When risk exceeds appetite and cannot be reduced or transferred effectively, avoidance is the remaining treatment. Continuing without action would exceed the organization's stated tolerance.

  145. Question 145AI life cycle and AI risk management

    In federated learning, what is a key risk management challenge for the central model owner?

    1. Poor visibility into participant data quality Correct answer.
    2. The guaranteed inability to use any encryption method
    3. The guaranteed loss of all model accuracy
    4. The elimination of all privacy concerns

    Explanation

    Federated learning keeps data distributed, so the central owner has less direct ability to inspect data quality and labeling. Governance must use participant attestations, monitoring, and validation.

  146. Question 146AI life cycle and AI risk management

    An organization runs several AI systems that share data sources and decision pipelines. Which method best assesses their combined risk?

    1. Reviewing each model in isolation only
    2. Assuming risks cannot compound
    3. Ignoring interactions between systems
    4. Portfolio-level scenario analysis Correct answer.

    Explanation

    Shared data and pipelines can create correlated or compounding failures across systems. Portfolio-level scenario analysis examines those interactions rather than treating each model alone.

  147. Question 147AI life cycle and AI risk management

    A model learns from historical hiring decisions that disadvantaged some groups. What risk is most directly illustrated?

    1. A lack of encryption in transit
    2. Excessive model explainability
    3. Label bias repeating past harm Correct answer.
    4. Overfitting to synthetic data

    Explanation

    Historical decisions used as labels can encode past discrimination into the model. The model then learns to repeat that pattern in new predictions.

  148. Question 148AI life cycle and AI risk management

    A vendor updates its hosted model frequently without advance notice. Which governance mechanism is most important?

    1. Relying on the original procurement review forever
    2. Change notice and revalidation triggers Correct answer.
    3. Disabling all monitoring to reduce alerts
    4. Assuming updates always improve performance

    Explanation

    Frequent vendor updates can change behavior, risk, and compliance posture. Contracts and governance should require notice and define when revalidation is triggered.

  149. Question 149AI life cycle and AI risk management

    A model's risk tier is lowered after new controls are validated. What should governance do next?

    1. Delete all prior risk records
    2. Immediately deploy without approval
    3. Record rationale and set review date Correct answer.
    4. Stop monitoring because risk is now low

    Explanation

    Changing a risk tier is a governance decision that should be documented and periodically rechecked. Records preserve accountability and support future audits.

  150. Question 150AI life cycle and AI risk management

    A model's aggregate accuracy is stable, but one subgroup has much higher error rates. What is the most appropriate next step?

    1. Investigate the subgroup gap and plan remediation Correct answer.
    2. Deploy immediately and monitor only aggregate accuracy
    3. Ignore the gap because aggregate accuracy is stable
    4. Remove the subgroup from the evaluation

    Explanation

    Aggregate metrics can hide material harm to a subgroup. The team should investigate causes and apply targeted remediation before or alongside deployment.

  151. Question 151Emerging and debated issues in AI governance

    Under current United States copyright practice, how is material generated by a generative AI system treated?

    1. Copyright vests in the model's weights, which are treated as a literary work.
    2. The developer of the model owns the copyright in all outputs the model generates.
    3. Material generated without sufficient human authorship is not copyrightable. Correct answer.
    4. The person who enters the prompt is automatically treated as the author.

    Explanation

    US copyright requires human authorship; works lacking it cannot be registered. The Copyright Office has stated that prompts alone generally do not supply the authorship needed for protection of the output. Arrangements among developers and users are contractual matters, not allocations of copyright by operation of law.

  152. Question 152Emerging and debated issues in AI governance

    What is the primary purpose of content provenance standards and watermarking for AI-generated media?

    1. Removing personal data from training datasets before training begins.
    2. Encrypting model weights so that they cannot be copied by competitors.
    3. Compressing model outputs so they are cheaper to serve to users.
    4. Recording origin and edit history so that authenticity can be verified. Correct answer.

    Explanation

    Provenance approaches such as signed content credentials attach verifiable metadata about who created or edited a piece of media. The governance goal is to help viewers and platforms distinguish authentic from synthetic or manipulated content. They do not shrink files, clean datasets, or protect model weights.

  153. Question 153Emerging and debated issues in AI governance

    What does ISO/IEC 42001 set out for organisations that develop or use AI systems?

    1. A treaty obligation that binds signatory states in international law.
    2. A benchmark leaderboard for comparing model accuracy across vendors.
    3. Requirements for establishing and continually improving an AI management system. Correct answer.
    4. A technical specification for how model weights must be encrypted.

    Explanation

    ISO/IEC 42001 is a management system standard, structured like other ISO management standards, covering policy, roles, risk treatment and continual improvement for AI. It is not a model benchmark, a cryptographic specification, or a treaty instrument.

  154. Question 154Emerging and debated issues in AI governance

    Which set lists the four core functions of the NIST AI Risk Management Framework?

    1. Govern, Map, Measure, Manage. Correct answer.
    2. Identify, Protect, Detect, Respond.
    3. Collect, Label, Train, Deploy.
    4. Plan, Do, Check, Act.

    Explanation

    The NIST AI RMF organises AI risk work into Govern, Map, Measure and Manage, with Govern cutting across the other three. The other listed cycles belong to quality management, cybersecurity or machine learning engineering practice rather than the AI RMF itself.

  155. Question 155Emerging and debated issues in AI governance

    Which described use of an AI system falls within the prohibited practices of the EU AI Act?

    1. A factory uses computer vision to detect defects on an assembly line.
    2. A company deploys a chatbot that tells users they are interacting with an AI system.
    3. A public authority socially scores residents, causing detrimental treatment in unrelated contexts. Correct answer.
    4. A mail provider automatically routes suspected spam into a separate folder.

    Explanation

    Social scoring by public authorities that leads to detrimental treatment in unrelated contexts is expressly prohibited. Transparency about AI interaction, spam filtering and industrial quality control are lawful uses subject to ordinary obligations, not prohibitions.

  156. Question 156Emerging and debated issues in AI governance

    What does New York City Local Law 144 require of employers using automated employment decision tools?

    1. Mandatory facial recognition to verify applicant identity.
    2. A complete prohibition on software tools in hiring decisions.
    3. An annual independent bias audit and advance notice to candidates. Correct answer.
    4. Public disclosure of salary ranges in all job advertisements.

    Explanation

    Local Law 144 requires an independent bias audit within the prior year, publication of summary results, and notice to candidates at least ten business days before use. It regulates rather than bans automated hiring tools and says nothing about salary transparency or biometric verification.

  157. Question 157Emerging and debated issues in AI governance

    What is the status of the Council of Europe Framework Convention on Artificial Intelligence?

    1. It is an EU regulation directly applicable in all member states.
    2. It is a UN Security Council resolution binding on all member states.
    3. It is a set of voluntary principles issued by the G7.
    4. It is the first legally binding international treaty on AI, opened for signature in 2024. Correct answer.

    Explanation

    The Council of Europe Framework Convention, opened for signature in September 2024, is widely described as the first binding international treaty addressing AI, framed around human rights, democracy and the rule of law. It is not a UN resolution, a voluntary set of principles, or EU secondary legislation.

  158. Question 158Emerging and debated issues in AI governance

    Which statement best distinguishes open-weight models from open-source AI systems?

    1. Any model offered through a public API is considered open source.
    2. Releasing weights alone does not make a system open source. Correct answer.
    3. Publishing model weights requires publishing the full training dataset as well.
    4. Open-weight models are barred from commercial use under most licences.

    Explanation

    Open source in AI is contested: releasing parameters gives inspection and local use, but training data, code and licences may remain restricted. Many licences also impose use restrictions that depart from conventional open-source definitions. Weight availability is therefore necessary but not sufficient.

  159. Question 159Emerging and debated issues in AI governance

    How do US courts currently approach the question of whether training a model on copyrighted works is fair use?

    1. Federal law exempts machine learning training from copyright liability entirely.
    2. They apply the four statutory factors case by case; no exemption covers AI training. Correct answer.
    3. Fair use depends only on whether the model's outputs carry a watermark.
    4. Training on copyrighted works infringes regardless of the purpose of the use.

    Explanation

    Fair use is an open-ended, fact-specific test weighing purpose and character of the use, nature of the work, amount taken and market effect. Earlier transformative-use cases such as book scanning inform the analysis, but no statute or definitive ruling settles generative training. Market dilution from competing outputs is a central, contested factor.

  160. Question 160Emerging and debated issues in AI governance

    An EU company plans to scrape publicly accessible web pages to build a training corpus containing personal data. Which approach is consistent with GDPR requirements?

    1. A documented balancing assessment is needed, and individuals keep objection and erasure rights. Correct answer.
    2. Scraping is lawful only where every individual has given explicit consent in advance.
    3. GDPR does not apply to data collected automatically by machines.
    4. Publicly accessible personal data may be scraped without identifying any legal basis.

    Explanation

    Legitimate interests can support scraping only after a three-part assessment of interest, necessity and balancing against individual rights, including reasonable expectations. Controllers must still honour objections and erasure requests and provide transparency. Consent is one route but not the only one.

  161. Question 161Emerging and debated issues in AI governance

    What protection does GDPR Article 22 give to individuals subject to solely automated decisions with legal or similarly significant effects?

    1. It prohibits all automated processing of personal data.
    2. It requires controllers to publish the source code of their models.
    3. It provides safeguards such as human intervention and the ability to contest the decision. Correct answer.
    4. It applies only to decisions taken by public authorities.

    Explanation

    Article 22 restricts solely automated decision-making with legal or similarly significant effects unless an exception applies, and in those cases requires safeguards including human intervention, the ability to express a view and to contest. It does not ban automated processing generally, nor require code disclosure. Private controllers are covered, not just public bodies.

  162. Question 162Emerging and debated issues in AI governance

    What would the proposed EU AI Liability Directive change for claimants harmed by AI systems?

    1. It would create a rebuttable presumption of causality and a right to disclosure of evidence for high-risk AI. Correct answer.
    2. It would replace the GDPR with a single unified AI liability regime.
    3. It would impose strict liability on developers for every AI-related harm regardless of fault.
    4. It would cap total damages for AI incidents at a fixed statutory amount.

    Explanation

    The proposal targets evidential difficulty: courts could presume a causal link where a high-risk system contributed to harm and the claimant faced excessive difficulty proving it, alongside powers to order disclosure of evidence from providers. Liability would remain fault-based rather than strict. It is complementary to, not a replacement for, data protection law.

  163. Question 163Emerging and debated issues in AI governance

    Why are advanced semiconductors and training compute central to proposals for governing frontier AI?

    1. Compute governance requires developers to publish their model weights openly.
    2. Compute and chips are treated as a measurable chokepoint that can be monitored and limited. Correct answer.
    3. Compute thresholds determine whether a model's outputs qualify for copyright.
    4. Compute controls restrict how many users are permitted to call a model's API.

    Explanation

    Because frontier training depends on scarce, traceable hardware, compute is one of the few inputs that can be counted, verified and restricted at the supply-chain level. Export controls and compute thresholds exploit this property. API rate limits, copyright and weight release are unrelated policy levers.

  164. Question 164Emerging and debated issues in AI governance

    What do voluntary frontier safety commitments typically involve?

    1. They are legally binding treaties concluded between states.
    2. Developers evaluate frontier models for dangerous capabilities and publish safety information before release. Correct answer.
    3. They apply only to systems placed on the EU market.
    4. They require the release of training data to the general public.

    Explanation

    Voluntary commitments made by leading developers generally cover pre-deployment testing, red teaming, external expert review and publication of safety and capability information for the most capable models. Being voluntary, they are not treaties and do not determine market access in a single jurisdiction. They do not mandate data release.

  165. Question 165Emerging and debated issues in AI governance

    What is the primary function of AI incident reporting regimes and incident databases?

    1. Assigning criminal liability automatically once an incident is recorded.
    2. Recording harms and near misses so patterns can be analysed and recurrence reduced. Correct answer.
    3. Benchmarking model accuracy against standardised test sets.
    4. Satisfying copyright registration formalities for generated works.

    Explanation

    Incident reporting collects structured accounts of harms and near misses, as in the OECD AI Incidents Monitor, so regulators and developers can identify recurring failure modes and target mitigations. Recording an incident does not determine fault or liability. The function is learning and oversight, not benchmarking or copyright administration.

  166. Question 166Emerging and debated issues in AI governance

    What is the purpose of an algorithmic impact assessment before an AI system is deployed?

    1. Documenting intended use, affected stakeholders and planned mitigations. Correct answer.
    2. Ranking prospective vendors by price and performance.
    3. Estimating the carbon cost of the training run.
    4. Certifying that a system contains no bias whatsoever.

    Explanation

    Impact assessments are structured, prospective documents describing the system's purpose, who may be affected, potential rights impacts and mitigations, often with review and publication commitments. They support accountability and contestability. They do not certify the absence of bias, and they are not procurement or emissions tools.

  167. Question 167Emerging and debated issues in AI governance

    What would the proposed NO FAKES Act do at the federal level in the United States?

    1. Ban the training of generative models on copyrighted works.
    2. Grant model developers copyright in the outputs their systems generate.
    3. Create a federal right against unauthorised digital replicas of a person's voice or likeness. Correct answer.
    4. Require watermarking of every audio recording released in the United States.

    Explanation

    The proposal would establish a transferable, licensable federal digital replication right, with allowances and post-mortem protections, addressing voice and likeness clones. It targets replicas rather than model training, output ownership or blanket watermarking mandates.

  168. Question 168Emerging and debated issues in AI governance

    What does the Illinois Biometric Information Privacy Act require?

    1. Informed written consent before collection, backed by a private right of action. Correct answer.
    2. It establishes a single federal biometric standard for the United States.
    3. It permits collection whenever the biometric data is encrypted at rest.
    4. It applies only to government agencies and public bodies.

    Explanation

    BIPA requires notice and written consent before collecting biometric identifiers, a written retention and destruction policy, and no sale or profit from the data. It is enforced through a private right of action with statutory damages. It is an Illinois statute applying to private entities, not a federal standard.

  169. Question 169Emerging and debated issues in AI governance

    Which assessment of synthetic training data is most accurate?

    1. It reduces exposure of real personal data but can degrade model quality and carry over source bias. Correct answer.
    2. Synthetic data is prohibited for model training under GDPR.
    3. Synthetic data counts as personal data only when it is encrypted.
    4. Synthetic data eliminates bias because it contains no real people.

    Explanation

    Synthetic data can limit the use of genuine personal records and fill gaps in rare cases, but models trained repeatedly on generated data can suffer quality collapse, and artefacts of the source data can persist. Whether it is personal data depends on identifiability, not encryption. GDPR does not ban it.

  170. Question 170Emerging and debated issues in AI governance

    Under the US Children's Online Privacy Protection Act, what is required before a covered service collects personal information from a child under 13?

    1. Notification to the school the child attends.
    2. Verifiable parental consent obtained in advance. Correct answer.
    3. Nothing additional, provided the service is free to use.
    4. Age verification performed through facial recognition.

    Explanation

    COPPA requires operators of services directed to children, or with actual knowledge of child users, to obtain verifiable parental consent before collecting personal information, and to provide notice and deletion rights. Biometric age verification and school notification are not mandated by the statute.

  171. Question 171Emerging and debated issues in AI governance

    Why do erasure requests create difficult governance problems for trained models?

    1. Removing records from a training set does not reliably remove their influence on a trained model. Correct answer.
    2. Controllers must retrain every model from scratch within 24 hours of a request.
    3. Erasure rights do not apply to data used to train AI systems.
    4. Erasure is satisfied once the individual's account and login history are deleted.

    Explanation

    Model parameters encode statistical influence from training data, so deleting source records does not by itself remove that influence from a deployed model. Machine unlearning is an active research area without settled guarantees. Erasure rights do apply to training data, but no fixed retraining deadline exists in law.

  172. Question 172Emerging and debated issues in AI governance

    What is the current state of foundation model transparency requirements?

    1. Disclosure of data sources and evaluation results is largely voluntary, and third-party indices score it. Correct answer.
    2. Developers are legally required to publish their full training corpora.
    3. Transparency duties apply only to open-weight models.
    4. Transparency is generally understood as releasing model weights publicly.

    Explanation

    Transparency indices and model cards document disclosed practices such as data sources, compute, evaluation results and risk mitigations, but in most jurisdictions disclosure remains voluntary or partial. No general legal duty requires publishing full training corpora. Releasing weights is one form of openness, not the meaning of transparency.

  173. Question 173Emerging and debated issues in AI governance

    What is the central governance question in debates about lethal autonomous weapons systems?

    1. Whether every autonomous weapon must be registered with the United Nations.
    2. How the EU AI Act's high-risk regime applies to military systems.
    3. Whether meaningful human control can be preserved so use complies with international humanitarian law. Correct answer.
    4. Whether the UN Charter already prohibits all autonomous weapons.

    Explanation

    The debate at the UN Convention on Certain Conventional Weapons focuses on maintaining meaningful human control over the use of force so that distinction, proportionality and precaution can be applied. No registration requirement or blanket Charter prohibition exists. The EU AI Act excludes military, defence and national security uses from its scope.

  174. Question 174Emerging and debated issues in AI governance

    Which statement reflects the emerging debate about AI's environmental footprint?

    1. AI systems are excluded from corporate sustainability reporting.
    2. Efficiency gains fully offset AI energy use, so reporting is unnecessary.
    3. There are growing calls for standardised reporting of energy and water use across training and inference. Correct answer.
    4. Training consumes no measurable energy once the model has been deployed.

    Explanation

    Inference, not just training, drives ongoing electricity and water demand, and measurement methodologies vary widely. Researchers and regulators have therefore pressed for standardised disclosure so that environmental claims can be compared. Offsetting claims and exemptions are not settled policy.

  175. Question 175Emerging and debated issues in AI governance

    What is the current legal position on granting personhood to AI systems?

    1. No major jurisdiction grants AI legal personhood, and an EU 'electronic personhood' proposal was not adopted. Correct answer.
    2. The Council of Europe Framework Convention establishes personhood for autonomous systems.
    3. The United Nations has recognised AI agents as legal persons able to enter contracts.
    4. The EU AI Act grants legal personhood to general-purpose AI models.

    Explanation

    A 2017 European Parliament resolution floated electronic personhood for robots, but it was widely criticised and not carried into legislation. Existing regimes place duties on developers, deployers and operators, who remain the accountable legal persons. No international instrument confers personhood on AI.

  176. Question 176Emerging and debated issues in AI governance

    Under the EU AI Act, which of the following practices is prohibited outright?

    1. Disclosing to users that they are interacting with a chatbot
    2. Government social scoring of individuals Correct answer.
    3. Deploying a certified high-risk recruitment tool
    4. Summarising customer service calls for internal quality review

    Explanation

    Article 5 of the EU AI Act bans a small set of practices considered unacceptable regardless of safeguards, including social scoring by public authorities that leads to detrimental treatment out of proportion to the context. Most other AI uses are permitted subject to transparency or high-risk obligations.

  177. Question 177Emerging and debated issues in AI governance

    The NIST AI Risk Management Framework is best described as which of the following?

    1. A certification scheme that replaces ISO/IEC 42001
    2. A voluntary framework organised around Govern, Map, Measure and Manage Correct answer.
    3. A binding federal regulation that imposes fines for unsafe models
    4. A statute that preempts state artificial intelligence laws

    Explanation

    The NIST AI RMF is a voluntary, use-oriented framework structured around four core functions: Govern, Map, Measure and Manage. It is intended to be paired with sector-specific profiles rather than to create legal duties.

  178. Question 178Emerging and debated issues in AI governance

    Under the EU AI Act, a general-purpose AI model is presumed to have systemic risk when cumulative training compute exceeds which threshold?

    1. 10^23 FLOP
    2. 10^26 FLOP
    3. 10^24 FLOP
    4. 10^25 FLOP Correct answer.

    Explanation

    Annex XIII of the EU AI Act sets the presumption at 10^25 floating point operations of cumulative training compute. The Commission can adjust the threshold as technology develops.

  179. Question 179Emerging and debated issues in AI governance

    The UNESCO Recommendation on the Ethics of Artificial Intelligence is which kind of instrument?

    1. A non-binding normative instrument adopted by UNESCO member states Correct answer.
    2. A technical standard issued by ISO
    3. A treaty binding ratifying states under international law
    4. A code of practice issued by the EU AI Office

    Explanation

    UNESCO adopted the Recommendation in 2021 as a normative but non-binding instrument, and member states are expected to report on implementation. It supplies shared ethical principles rather than enforceable legal duties.

  180. Question 180Emerging and debated issues in AI governance

    Content Credentials built on the C2PA standard primarily do which of the following?

    1. Strip personal data from training corpora before model training
    2. Encrypt media so that only paying subscribers can view it
    3. Detect whether an image was produced by a particular model vendor
    4. Attach cryptographically signed provenance metadata to media Correct answer.

    Explanation

    C2PA provenance records attach signed, tamper-evident metadata describing how and by whom content was created or edited. They support verification of provenance but do not by themselves prove that content is true or benign.

  181. Question 181Emerging and debated issues in AI governance

    In Thaler v. Perlmutter, US courts addressed whether copyright can subsist in machine-generated output; the holding was that copyright protection does which of the following?

    1. Depends on whether the model was trained on public domain works
    2. Is granted to the model developer by default
    3. Extends to any work containing minimal creative input, including purely machine output
    4. Requires human authorship and cannot vest in an AI system Correct answer.

    Explanation

    US copyright law requires an author who is a human being, so a work generated autonomously by an AI system without human creative contribution cannot be registered. Human-authored contributions arranged or edited with AI tools may still be protectable.

  182. Question 182Emerging and debated issues in AI governance

    Article 4 of the EU AI Act requires providers and deployers of AI systems to do which of the following?

    1. Register every model in an EU database before release
    2. Obtain user consent before every model inference
    3. Publish their full training datasets
    4. Take measures to ensure sufficient AI literacy among their staff Correct answer.

    Explanation

    Article 4 imposes an AI literacy duty: providers and deployers must take measures to ensure that staff and others dealing with AI systems on their behalf have sufficient understanding of the technology and its risks. The obligation is calibrated to context, prior knowledge and the people affected.

  183. Question 183Emerging and debated issues in AI governance

    California's SB 1047, the Safe and Secure Innovation for Frontier AI Models Act, ultimately met which fate?

    1. It was struck down by the state supreme court
    2. It was signed into law and is now in force
    3. It was withdrawn before reaching the legislature
    4. It was vetoed by the governor in 2024 Correct answer.

    Explanation

    Governor Newsom vetoed SB 1047 in September 2024, citing concerns about the bill's focus on large models and thresholds. The episode is often cited in debates about prescriptive frontier-model statutes versus narrower, harm-specific rules.

  184. Question 184Emerging and debated issues in AI governance

    Colorado's AI Act imposes which duty on developers and deployers of high-risk AI systems used for consequential decisions?

    1. A duty to publish complete model weights to the public
    2. A duty of reasonable care to protect consumers from foreseeable algorithmic discrimination Correct answer.
    3. A duty to obtain federal pre-approval before any deployment
    4. A duty to pay a per-inference levy into a state fund

    Explanation

    Colorado's framework requires developers and deployers to use reasonable care to protect consumers from known or reasonably foreseeable risks of algorithmic discrimination. It is paired with impact assessment, disclosure and consumer notice obligations.

  185. Question 185Emerging and debated issues in AI governance

    The G7 Hiroshima Process International Code of Conduct for Advanced AI Systems is best characterised as which of the following?

    1. A procurement rule binding the suppliers of G7 governments
    2. A voluntary code that organisations may choose to follow Correct answer.
    3. A UN Security Council resolution
    4. A binding treaty open only to G7 member states

    Explanation

    The Hiroshima Process produced a voluntary code of conduct that advanced AI developers are encouraged to apply, alongside a reporting framework. It is not legally binding and operates through political commitment and peer pressure.

  186. Question 186Emerging and debated issues in AI governance

    California's privacy regulations on automated decision-making technology require businesses to do which of the following?

    1. Offer every consumer a copy of the model weights
    2. Notify the attorney general 30 days before training any model
    3. Complete and document a risk assessment before using ADMT for significant decisions Correct answer.
    4. Obtain opt-in consent for all uses of machine learning

    Explanation

    The CPPA regulations treat automated decision-making technology used for significant decisions as high-risk processing, requiring a documented risk assessment before use and periodic review. Consumers gain access, opt-out and related rights calibrated to that use.

  187. Question 187Emerging and debated issues in AI governance

    The US Federal Trade Commission's principal authority over deceptive claims about AI products derives from which source?

    1. The Digital Services Act
    2. The extraterritorial provisions of the EU AI Act
    3. Executive Order 14110
    4. Section 5 of the FTC Act, which prohibits unfair or deceptive practices Correct answer.

    Explanation

    The FTC polices AI marketing and use through its general Section 5 authority over unfair or deceptive acts or practices, applying existing consumer protection law to new technology. Enforcement actions have targeted overstated capability claims and unsubstantiated performance promises.

  188. Question 188Emerging and debated issues in AI governance

    Article 50 of the EU AI Act requires deployers of deepfake content to do which of the following?

    1. Disclose that the content has been artificially generated or manipulated Correct answer.
    2. Delete the content within 24 hours of publication
    3. Register the content with the AI Office before publication
    4. Pay a licence fee to the rights holders of the training data

    Explanation

    Article 50 imposes transparency duties so that people can recognise synthetic or manipulated content, with limited carve-outs for evidently artistic or satirical works. Providers of generative systems must also mark outputs in a machine-readable way.

  189. Question 189Emerging and debated issues in AI governance

    After Executive Order 14110 was rescinded in January 2025, the subsequent US federal policy direction on AI was articulated in which document?

    1. America's AI Action Plan Correct answer.
    2. The OECD AI Principles
    3. The National AI Initiative Act
    4. The Blueprint for an AI Bill of Rights

    Explanation

    The 2025 America's AI Action Plan set out the administration's approach, emphasising innovation, infrastructure and export promotion while replacing the earlier safety-focused executive order. Federal agencies were directed to revise policies accordingly.

  190. Question 190Emerging and debated issues in AI governance

    The International Network of AI Safety Institutes is best described as which of the following?

    1. A UN agency that licenses frontier models
    2. A network of government-backed technical institutes that share testing and evaluation practices Correct answer.
    3. A treaty organisation with binding enforcement powers
    4. A private standards body funded by model developers

    Explanation

    The network connects state-backed technical bodies that conduct and share evaluations, research and best practices on advanced AI safety. It coordinates methodologically rather than exercising regulatory authority.

  191. Question 191Emerging and debated issues in AI governance

    A model card is primarily intended to do which of the following?

    1. Encrypt model weights for secure distribution
    2. Document a model's intended use, performance and limitations Correct answer.
    3. Certify that a model complies with the EU AI Act
    4. Record the licence terms of the training data

    Explanation

    Model cards are short disclosure documents summarising intended use, evaluation results, limitations and ethical considerations. They support informed downstream use but are self-reported and not a compliance certificate.

  192. Question 192Emerging and debated issues in AI governance

    Tennessee's ELVIS Act was notable in AI governance because it did which of the following?

    1. Extended right-of-publicity protection to a person's voice Correct answer.
    2. Banned all generative AI use in the music industry
    3. Required labels on AI-generated albums only
    4. Created a federal registry of AI-generated songs

    Explanation

    The statute updated Tennessee's right of publicity to cover voice explicitly, giving performers a civil route against unauthorised voice clones. It addresses likeness-based harms rather than regulating generative AI generally.

  193. Question 193Emerging and debated issues in AI governance

    The FCC's 2024 declaratory ruling on AI-generated voices in robocalls did which of the following?

    1. Confirmed that AI-generated voices count as 'artificial' under the TCPA Correct answer.
    2. Banned all AI voice technology nationwide
    3. Required carriers to block all international calls
    4. Created a licensing regime for voice cloning

    Explanation

    The ruling clarified that the TCPA's restriction on artificial or prerecorded voices covers AI-generated speech, so such calls require prior express consent where the statute applies. It extended existing telephony rules rather than creating a new technology-specific regime.

  194. Question 194Emerging and debated issues in AI governance

    Under the GDPR, a data protection impact assessment is required when processing, including AI processing, does which of the following?

    1. Involves any personal data used to train a model
    2. Uses a model with more than one billion parameters
    3. Is carried out by an organisation based outside the EU
    4. Is likely to result in a high risk to individuals' rights and freedoms Correct answer.

    Explanation

    Article 35 triggers a DPIA where a processing operation, including profiling and large-scale monitoring, is likely to produce high risk to individuals. Supervisory authorities publish lists of processing that always requires one.

  195. Question 195Emerging and debated issues in AI governance

    The UK's approach to AI regulation has generally favoured which of the following?

    1. A moratorium on AI deployment until 2030
    2. Sector-specific regulators applying existing law alongside cross-cutting principles Correct answer.
    3. A single comprehensive AI statute modelled on the EU AI Act
    4. Delegating all AI oversight to local authorities

    Explanation

    The UK has pursued a pro-innovation, context-based model in which existing regulators apply five cross-sectoral principles within their own sectors. Proposals for targeted statutory duties on frontier developers have been debated but the default remains decentralised.

  196. Question 196Emerging and debated issues in AI governance

    Under the EU AI Act, a serious incident involving a high-risk AI system must generally be reported to the market surveillance authorities within which period after the provider becomes aware of it?

    1. 2 days
    2. 10 days
    3. 15 days Correct answer.
    4. 30 days

    Explanation

    Article 73 sets a general reporting window of 15 days from awareness, shortened to 2 days where there is widespread infringement or a serious incident involving a general-purpose model, and 10 days where a person has died. Reports feed into market surveillance and corrective action.

  197. Question 197Emerging and debated issues in AI governance

    Under the EU AI Act, most Annex III high-risk AI systems may have their conformity assessed through which route?

    1. Internal control procedures set out in Annex VI Correct answer.
    2. Third-party assessment by a notified body in every case
    3. Self-certification with no supporting documentation
    4. A discretionary hearing before the European Commission

    Explanation

    Article 43 allows internal control assessment for most Annex III systems, with third-party notified body involvement required for biometric identification and certain other categories. All routes require technical documentation, risk management and post-market monitoring.

  198. Question 198Emerging and debated issues in AI governance

    Under the Digital Services Act, very large online platforms are required to do which of the following?

    1. Assess and mitigate systemic risks, including those stemming from recommender systems Correct answer.
    2. Publish their full source code to the public
    3. Remove all AI-generated content within 24 hours
    4. Obtain user consent before applying any content ranking

    Explanation

    Designated very large platforms must conduct annual systemic risk assessments covering their algorithms, content moderation and societal impacts, then adopt mitigation measures subject to independent audit. The regime is risk-based rather than technology-specific.

  199. Question 199Emerging and debated issues in AI governance

    The EU's commercial text and data mining exception in the DSM Directive applies in which circumstances?

    1. Unless rightholders have reserved their rights in a machine-readable way Correct answer.
    2. To all works regardless of any reservation of rights by the rightholder
    3. Only to non-commercial academic research projects
    4. Only where a licence has been negotiated for every work used

    Explanation

    Article 4 permits commercial text and data mining unless rightholders have expressly opted out through machine-readable means such as robots.txt or metadata. Article 3 separately covers research organisations and cultural heritage institutions.

  200. Question 200Emerging and debated issues in AI governance

    The debate over federal preemption in US AI policy centres on which question?

    1. Whether the Commerce Department should set standards that states must adopt
    2. Whether a federal statute should displace the growing patchwork of state AI rules Correct answer.
    3. Whether state attorneys general may enforce federal consumer protection law against AI developers
    4. Whether the FTC must approve state AI legislation before it takes effect

    Explanation

    With many states enacting or considering AI laws, proposals have circulated to set a single federal standard that would preempt inconsistent state requirements. The debate pits harmonisation and compliance certainty against concerns about lowering protection and losing state experimentation.

AIGP flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Supervised learning
A machine-learning approach that trains a model on labeled input-output pairs so it can predict the label for new, unseen inputs.
Reinforcement learning
A training paradigm in which an agent learns a policy by acting in an environment and receiving rewards or penalties, optimizing long-run cumulative reward rather than fitting labeled examples.
F1 score
The harmonic mean of precision and recall, computed as 2PR/(P+R); it balances false positives and false negatives and is useful when class distributions are imbalanced.
Overfitting
A model fits noise and idiosyncrasies in training data, producing low training error but poor performance on new data; mitigated with more data, regularization, and cross-validation.
Training versus inference
Training estimates model parameters from data using substantial compute; inference applies the trained model to new inputs and scales with the number of requests served.
Generative AI
Models that learn a data distribution and produce new content such as text, images, audio, or code, typically built on transformer or diffusion architectures.
Narrow AI
An AI system designed for a bounded task or domain, such as translation or image classification; all deployed commercial systems today fall into this category.
Automation bias
The human tendency to over-trust algorithmic outputs and reduce independent scrutiny, which can amplify system errors and weaken the effectiveness of human oversight.
EU AI Act risk tiers
Four-level classification - unacceptable (prohibited), high, limited (transparency duties), and minimal risk - that sets obligations for providers and deployers of AI systems in the EU.
NIST AI RMF functions
Govern, Map, Measure, and Manage: the four voluntary functions structuring the US NIST AI Risk Management Framework, with Govern as the cross-cutting culture and accountability function.
High-risk AI system
Under the EU AI Act, a system meeting Annex I product-safety rules or an Annex III listed use such as biometrics, employment, or credit, triggering conformity assessment and documentation duties.
OECD AI Principles
Five value-based principles covering inclusive growth, human rights, transparency, robustness and safety, and accountability; they underpin many national AI policies and the EU definition.
ISO/IEC 42001
An international standard specifying requirements for an AI management system, certifiable by third parties, covering governance, risk treatment, and continual improvement of AI processes.
GDPR Article 22
Grants individuals the right not to be subject to decisions based solely on automated processing, including profiling, that produce legal or similarly significant effects on them.
NYC Local Law 144
Requires employers and employment agencies using automated employment decision tools in New York City to conduct annual independent bias audits and notify candidates and employees.
General-purpose AI obligations
EU AI Act duties on providers of models usable for many tasks: technical documentation, copyright policy, training-data summary, and added systemic-risk evaluation above a compute threshold.
Model drift
Degradation of a deployed model's performance over time as conditions change; detected by monitoring input distributions, prediction distributions, and live outcome metrics against baselines.
Concept drift
A change in the statistical relationship between inputs and the target outcome, so a previously accurate model becomes wrong even when inputs look familiar.
Residual risk
The risk remaining after mitigation controls are applied; acceptance decisions compare residual risk against the organization's defined risk appetite and tolerance thresholds.
Model card
A short document reporting a model's intended use, training data, evaluation metrics by subgroup, ethical considerations, and known limitations for downstream users.
Data sheet for datasets
Documentation of a dataset's motivation, composition, collection process, preprocessing, recommended uses, and maintenance, enabling users to judge suitability and provenance.
Human-in-the-loop
An oversight mode in which a person reviews and approves each consequential model output before action, appropriate for high-severity or low-confidence decisions.
Red teaming
Structured adversarial testing in which specialists attempt to elicit harmful, unsafe, or incorrect model behavior to surface risks that standard benchmark evaluation misses.
Shadow deployment
Running a new model alongside the production model on live traffic without acting on its outputs, allowing risk-free comparison before a staged rollout.
Alignment problem
The difficulty of specifying objectives so an AI system's behavior reliably matches human intent and values, including risks of specification gaming and reward hacking.
Post-hoc explainability
Techniques such as SHAP, LIME, and counterfactuals that approximate why a trained model produced an output; the explanations are estimates and may not be faithful.
Prompt injection
An attack in which untrusted text, such as content from a web page or uploaded document, is interpreted by a language model as instructions and hijacks its behavior.
C2PA Content Credentials
An open technical standard for cryptographically signed provenance metadata attached to media, indicating how and by which tool content was created or edited.
Open-weight model
A model whose trained parameters are publicly downloadable, enabling local deployment and inspection but limiting the developer's ability to restrict or revoke its use.
Compute governance
Policy approaches that regulate AI development through control of computational resources, such as chip export controls, training-compute thresholds, and cloud reporting requirements.
Synthetic media
Image, audio, or video content generated or altered by AI, raising concerns about fraud, non-consensual intimate imagery, and election misinformation.
AI liability gap
The difficulty of assigning legal responsibility when harm emerges from complex multi-party AI supply chains in which no single actor controlled the full system.

AIGP study guide

Foundations of AI Systems, Their Use Cases and Impacts

Get the taxonomy right

The single most common source of missed points is confusing the nested categories. Artificial intelligence is the broad field of systems performing tasks associated with human cognition. Machine learning (ML) is a subset in which behaviour is learned from data rather than explicitly coded. Deep learning is a subset of ML using multi-layer neural networks. Generative AI is a subset of deep learning that produces new content (text, images, audio, code), typically via foundation models trained with self-supervision and adapted by fine-tuning, prompting or retrieval augmentation.

Learning paradigms and task types

Supervised learning uses labelled examples (classification, regression); unsupervised learning finds structure without labels (clustering, dimensionality reduction); self-supervised learning derives labels from the data itself (the basis of large language models); reinforcement learning optimises actions against a reward signal, including RLHF for alignment. Know the difference between training (fitting parameters) and inference (serving predictions), and between deterministic rule-based systems and probabilistic models.

Use cases and impacts

Candidates should map capabilities to sectors: fraud detection, credit scoring, medical imaging triage, predictive maintenance, recommendation, conversational agents, document review. Impacts are two-sided: efficiency, personalisation and discovery versus bias and allocative harm, opacity, privacy erosion, safety failures, automation-driven labour displacement, energy and water consumption, and error amplification at scale. Evaluation vocabulary matters: confusion matrix, precision, recall, F1, false positive versus false negative costs, overfitting, data drift, and human-in-the-loop oversight.

Worked example

A bank builds a model predicting loan default from ten years of labelled repayment history. That is supervised classification, not generative AI and not reinforcement learning. Because the output is a probability, the threshold is a policy choice: lowering it raises recall (fewer missed defaults) but lowers precision (more good applicants rejected) - so the governance question is who owns that trade-off, not merely how accurate the model is.

Trap: options that treat AI, ML, deep learning and generative AI as synonyms, or that assume more data always improves outcomes.

AI Laws, Regulation and Frameworks for Responsible Governance

Distinguish binding from voluntary

Questions frequently test whether an instrument is hard law, a standard, or a framework. Binding: EU AI Act, GDPR, China's Interim Measures for Generative AI Services (2023), algorithm recommendation and deep synthesis rules, Colorado's AI Act (reasonable care and impact assessments for high-risk systems, with a delayed effective date), and sectoral US laws such as ECOA, FCRA, Fair Housing Act and HIPAA. Voluntary or standards-based: OECD AI Principles, UNESCO Recommendation on the Ethics of AI (2021), the Council of Europe Framework Convention on AI (2024), NIST AI RMF (Govern, Map, Measure, Manage) with its Generative AI Profile, ISO/IEC 42001 (certifiable AI management system) and ISO/IEC 23894 (AI risk guidance).

The EU AI Act's risk pyramid

Prohibited practices (e.g., social scoring, certain manipulation and biometric categorisation) sit at the top; then high-risk systems (conformity assessment, risk management, data governance, logging, human oversight, accuracy and cybersecurity duties); then limited-risk systems subject to transparency obligations (chatbots, deepfake disclosure); then minimal risk. General-purpose AI models carry separate transparency and, for systemic-risk models, evaluation and incident-reporting duties. Application is staged, and the Act has extraterritorial reach where outputs are used in the EU.

Traps to avoid

Assuming GDPR Article 22 bans automated decisions - it does not; it conditions solely automated decisions with legal or similarly significant effects on exceptions and safeguards. Conflating the AI Act's 'high-risk' with GDPR 'high-risk processing'. Assuming NIST AI RMF is mandatory. Believing frameworks and laws are interchangeable.

Rule of thumb

Ask three questions in order: Is it binding here? What risk tier or legal basis applies? What documentation and oversight does the instrument demand?

AI Life Cycle and AI Risk Management

Life cycle stages

A defensible life cycle runs: problem definition and business understanding → data sourcing and assessment → data preparation → model development and selection → validation and testing → deployment → monitoring and maintenance → decommissioning. Governance hooks attach to each stage: data provenance and lawful basis at sourcing; documentation and approval gates before deployment; drift detection, incident response and periodic revalidation after deployment; archival and model retirement at the end.

Risk management mechanics

Know the NIST AI RMF functions - Govern (culture, accountability, policies), Map (context, stakeholders, affected parties), Measure (testing, metrics, bias and safety evaluation), Manage (prioritise, treat, monitor) - and that Govern is continuous, not a phase. Risks are characterised as likelihood × impact, distinguished as inherent versus residual after controls, and held against a stated risk appetite and tolerance. Controls include human review, red teaming, bias and adversarial testing, logging and audit trails, access controls, model cards and datasheets for datasets. Escalation typically follows a three-lines-of-defence model: product teams own risk, risk/compliance functions oversee, internal audit assures. Financial services candidates should recognise SR 11-7-style model risk management: inventory, validation, ongoing monitoring.

Assessments

An AI impact assessment examines rights, safety, fairness and societal effects; a DPIA examines personal-data processing; a fundamental rights impact assessment is the AI Act's high-risk variant. A vendor assessment covers third-party and supply-chain risk - using a vendor model does not transfer accountability.

Worked example

A team proposes an LLM-based HR screening tool. Map: affected parties are applicants. Measure: test for disparate impact across protected groups and for hallucinated justifications. Manage: human review before rejection, logging, and a quarterly re-test. Trap: approving once and never re-testing - drift and prompt or model changes invalidate prior assurance.

Emerging and Debated Issues in AI Governance

Where the debate actually sits

Many topics in this domain are contested rather than settled, and questions often reward recognising that a proposal is a proposal. Understand the main fault lines: capability versus existential risk framing; open-weight versus closed models (argued as both a safety risk and a transparency, competition and auditability benefit); compute governance and training-run thresholds; agentic systems that act autonomously across tools; and alignment and evaluation science, including benchmark contamination, red-teaming standards and the difficulty of measuring capability before deployment.

Data, content and rights

Training-data debates cover copyright and fair use, licensing and opt-out, consent and notice, and memorisation of personal data. Synthetic content raises deepfakes, fraud, non-consensual imagery, election integrity, and the limits of provenance tools such as watermarking and C2PA content credentials - watermarking is useful but defeatable, so it supports rather than replaces verification. Model collapse from training on synthetic data is a research concern, not an established certainty.

Structural and societal questions

Labour market effects and the augmentation-versus-automation distinction; environmental cost of training and inference; concentration of compute and model power; liability allocation across developer, deployer and user; insurance and redress; and global fragmentation versus harmonisation of rules. Treat each as a live policy choice with competing evidence.

Worked example

A question asks whether a company must watermark AI-generated images. If the jurisdiction is the EU, AI Act transparency obligations push toward machine-readable marking of synthetic content; elsewhere it may be voluntary or absent. The defensible answer states the obligation is jurisdiction- and context-dependent, and that provenance marking is a control, not a compliance guarantee.

Trap: treating a research finding, a draft bill or a corporate pledge as binding law.

AIGP cheat sheet

Quick reference

Foundations of AI Systems, Use Cases, and Impacts

  • AI ⊃ ML ⊃ deep learning: AI is the broadest field, ML learns patterns from data, deep learning uses multi-layer neural networks, and generative AI produces new content from learned distributions.

  • Supervised learning maps labeled inputs to outputs (classification predicts discrete labels, regression predicts continuous values); unsupervised learning finds structure in unlabeled data (clustering, dimensionality reduction); reinforcement learning optimizes a policy via reward signals from an environment.

  • Training fits parameters to data and is a one-time, compute-heavy cost; inference applies the fitted model to new inputs and its cost scales with request volume.

  • Precision = TP/(TP+FP); Recall = TP/(TP+FN); F1 = harmonic mean of precision and recall. Moving the decision threshold trades one for the other, so metric choice must match the cost of each error type.

  • Overfitting = low training error with high generalization error; mitigations include more or better data, regularization, cross-validation, and early stopping.

  • Bias enters at data (sampling, labeling, historical), model (objective function, feature choice), and deployment (population shift, human use); mitigation must target the specific stage.

  • Narrow AI (ANI) is task-bounded; artificial general intelligence (AGI) is hypothetical; no deployed commercial system meets an AGI definition.

  • Common enterprise use cases: predictive scoring, recommender systems, computer-vision inspection, NLP extraction and generation, anomaly detection, and process automation - each with distinct accuracy and failure-mode profiles.

  • Impact classes: labor substitution vs. augmentation, allocative effects of scoring, accessibility gains, opacity that reduces contestability, and energy/water cost of training and inference.

Quick reference

AI Laws, Regulation, and Responsible Governance Frameworks

  • EU AI Act (Regulation (EU) 2024/1689) entered into force 1 August 2024 and uses a four-tier risk pyramid: unacceptable (prohibited), high, limited-risk transparency duties, and minimal risk.

  • EU AI Act application dates: prohibitions 2 February 2025; general-purpose AI obligations 2 August 2025; high-risk Annex III systems 2 August 2026; Annex I product-embedded systems 2 August 2027.

  • EU AI Act prohibited practices include public-authority social scoring, subliminal or manipulative techniques, exploiting vulnerabilities, untargeted facial-image scraping to build databases, emotion recognition in workplaces and schools, and most real-time remote biometric identification in public spaces.

  • NIST AI RMF (AI 100-1, January 2023) is voluntary and organized around four functions - Govern, Map, Measure, Manage - and is paired with the Generative AI Profile (NIST AI 600-1).

  • OECD AI Principles (2019, updated 2024): inclusive growth, human rights and democratic values, transparency and explainability, robustness/safety/security, and accountability; they underpin the G20 and EU AI definitions.

  • ISO/IEC 42001 specifies a certifiable AI management system; ISO/IEC 23894 provides AI risk-management guidance; ISO/IEC TR 24028 addresses AI trustworthiness.

  • GDPR Article 22 gives a right not to be subject to solely automated decisions with legal or similarly significant effects; Article 35 requires a DPIA for high-risk processing.

  • Sub-national and state examples: NYC Local Law 144 requires annual independent bias audits and candidate notice for automated employment decision tools; Colorado's AI Act imposes reasonable-care duties on developers and deployers of high-risk systems.

  • Sector regulators still apply: FDA for software as a medical device, SEC for robo-advisory, EEOC/Title VII for hiring algorithms, and FTC Act §5 for unfair or deceptive AI claims.

Quick reference

AI Life Cycle and AI Risk Management

  • Life-cycle stages: problem framing → data acquisition and preparation → model development → validation and evaluation → deployment → monitoring → retirement; governance controls attach at each stage.

  • Risk = likelihood × impact (consequence). Inherent risk is scored before controls and residual risk after; acceptance decisions compare residual risk against a defined risk appetite.

  • Data drift = input distribution changes; concept drift = the input-to-output relationship changes; model drift = observed performance decay. Each requires a different monitoring signal.

  • Bias mitigation timing: pre-processing (data rebalancing, relabeling), in-processing (fairness constraints in the objective), post-processing (group-specific threshold adjustment).

  • Testing layers: unit and functional tests, held-out and out-of-distribution evaluation, red teaming for adversarial and misuse scenarios, shadow or canary deployment, and A/B tests with guardrail metrics.

  • Required artifacts: model card (intended use, training data, subgroup metrics, limits), datasheet for datasets (provenance, collection, preprocessing), system card, and an AI impact assessment.

  • Human oversight modes: human-in-the-loop (approve each action), human-on-the-loop (monitor and intervene), and human-out-of-the-loop; oversight depth should scale with autonomy and harm severity.

  • Third-party controls: vendor due diligence, contractual audit and incident-notice rights, model-version disclosure, and exit or portability plans for foundation-model dependencies.

  • Deployment safeguards: rate limits, output filters, confidence thresholds, kill switch, and a documented rollback path; production incidents feed back into the Map and Measure functions.

Quick reference

Emerging and Debated Issues in AI Governance

  • The alignment problem: optimizing a specified objective can diverge from operator intent through specification gaming or reward hacking; RLHF and rule-based training are mitigations, not guarantees.

  • Interpretability split: inherently interpretable models (linear, tree, GAM) versus post-hoc explanation (SHAP, LIME, counterfactuals); post-hoc explanations are approximations and can be unfaithful.

  • Generative AI copyright debates: whether training on copyrighted works is fair use or fair dealing is jurisdiction-specific and largely unresolved; EU law grants text-and-data-mining opt-outs to rights holders.

  • Synthetic media: watermarking and provenance standards such as C2PA Content Credentials are voluntary and can be stripped; detection is probabilistic and degrades as generators improve.

  • Open-weight vs. closed models: open weights increase external scrutiny and local control but complicate revocation of harmful capabilities and allocation of liability.

  • Security debates: prompt injection and jailbreaks are unsolved at the model layer; data poisoning attacks training pipelines; agentic systems widen the attack surface through tool and system access.

  • Concentration-of-power concerns: compute, data, and talent bottlenecks favor a few firms; compute governance (chip export controls, compute thresholds, cloud reporting) is a proposed policy lever.

  • Accountability gaps: the many-hands problem in multi-party AI supply chains complicates liability; proposed fixes include deployer-centric duties, mandatory audit rights, and AI insurance.

  • Environmental and labor impacts: training and inference energy and water use, plus hidden data-labeling labor, are increasingly included in AI impact assessments.

AIGP key terms

Artificial intelligence (AI)

A field of computing concerned with systems that carry out tasks associated with human intelligence, such as recognising patterns, reasoning, translating languages, and generating text or images.

Related: Machine learning, Generative AI
Machine learning

A branch of AI in which a system learns patterns from examples rather than following only hand-written rules, so its performance on a task improves with exposure to data.

Related: Training data, Foundation model
Training data

The examples used to teach a model statistical patterns. Its quality, representativeness, licensing and provenance strongly shape how the resulting system behaves and which risks it carries.

Related: Machine learning, Algorithmic bias, Copyright and AI training data
Foundation model

A large model trained on broad, general data that can be adapted to many downstream tasks through fine-tuning or prompting instead of being built for a single purpose.

Related: Large language model (LLM), General-purpose AI model
Large language model (LLM)

A foundation model trained on very large bodies of text to predict and produce language, enabling tasks such as summarising, translating, answering questions and writing code.

Related: Generative AI, Hallucination, Foundation model
Generative AI

AI that creates new content, such as text, images, audio, video or code, by reproducing patterns learned from training data rather than only labelling or scoring existing items.

Related: Large language model (LLM), Deepfake
Hallucination

An output in which a generative model states something false or unsupported as if it were true, because it predicts plausible-sounding content instead of checking facts.

Related: Large language model (LLM), Human-in-the-loop
Algorithmic bias

Systematic and unfair differences in how an AI system treats people or groups, often caused by unrepresentative data, poor labels or design choices, which can entrench discrimination.

Related: Training data, Red teaming
Explainability

The extent to which people can understand why a system produced a given output, including which inputs mattered and how confident the result is.

Related: Transparency, Model card
AI governance

The laws, policies, roles, processes and controls used by an organisation or government to steer the development and use of AI toward lawful, ethical and accountable outcomes.

Related: AI risk management, Accountability
Accountability

The principle that identified people and organisations must answer for the AI systems they build or operate, supported by clear responsibilities, records and consequences.

Related: AI governance, Transparency
Transparency

The principle that providers and users of AI should disclose enough about a system's purpose, data, capabilities and limits for people to understand and challenge how it is used.

Related: Explainability, Model card
EU AI Act

A European Union law that sorts AI systems into risk tiers and attaches duties accordingly, from outright bans on prohibited practices to detailed requirements for high-risk and general-purpose AI.

Related: Risk-based approach, High-risk AI system, General-purpose AI model
Risk-based approach

A way of regulating or governing AI in which obligations grow with the chance and seriousness of harm, so riskier uses face stricter requirements and closer oversight.

Related: EU AI Act, Risk assessment, High-risk AI system
High-risk AI system

An AI use that meets danger criteria set in law, such as safety components, biometric identification, critical infrastructure, hiring, education or access to essential services, and so attracts extra obligations.

Related: EU AI Act, Human-in-the-loop, Risk assessment
General-purpose AI model

A model trained on broad data that can perform many different tasks and be built into numerous downstream products; EU rules require documentation and, for the most capable, extra testing.

Related: Foundation model, EU AI Act
NIST AI Risk Management Framework

A voluntary United States framework organising AI risk work into govern, map, measure and manage functions, accompanied by profiles including one for generative AI.

Related: AI risk management, ISO/IEC 42001
ISO/IEC 42001

An international standard setting requirements for an AI management system, giving organisations a certifiable structure of policy, roles, risk treatment and continual improvement for responsible AI.

Related: NIST AI Risk Management Framework, AI governance
OECD AI Principles

Intergovernmental recommendations for trustworthy AI that call for inclusive growth, human rights protection, transparency, robustness, accountability and sustained investment in research.

Related: AI ethics, Accountability
Automated decision-making

Decisions about people made by a system with little or no human involvement; under the GDPR such decisions with significant effects carry special safeguards and a right to contest.

Related: Human-in-the-loop, Accountability
AI life cycle

The full sequence a system passes through - problem definition, data work, building, testing, deployment, monitoring and retirement - each stage needing its own governance and controls.

Related: AI risk management, Post-deployment monitoring
AI risk management

The continuing process of finding, assessing, prioritising, treating and watching the risks AI poses to people, the organisation and society, so decisions rest on evidence.

Related: Risk assessment, NIST AI Risk Management Framework
Risk assessment

A structured look at what could go wrong with a system, how likely and serious the harm would be, who is affected and which controls would reduce it.

Related: Risk tolerance, Residual risk, AI risk management
Risk tolerance

The amount and kind of risk an organisation is willing to carry in pursuit of its goals, which decides whether a proposed AI use is approved, modified or refused.

Related: Residual risk, Risk assessment
Residual risk

The risk left once planned controls are in place; leaders must decide whether to accept it or ask for further mitigation before the system goes live.

Related: Risk assessment, Risk tolerance
Human-in-the-loop

A design where a person reviews, approves or can override a system's output before it takes effect, giving meaningful control where errors could seriously harm people.

Related: High-risk AI system, Post-deployment monitoring
Red teaming

Adversarial testing in which specialists deliberately try to make a system fail, behave unsafely or be misused, so weaknesses are found before harm occurs.

Related: Post-deployment monitoring, Hallucination
Model drift

The gradual decline in a model's accuracy as live conditions and data move away from what it learned during training; catching it requires ongoing measurement and retraining.

Related: Post-deployment monitoring, Training data
Post-deployment monitoring

Ongoing observation of a live AI system's performance, outputs and effects so that errors, drift, misuse or emerging harms are detected and fixed quickly.

Related: Model drift, AI life cycle
Model card

A short published document describing a model's intended uses, training data, evaluation results, limits and known risks, helping downstream users judge whether it fits their purpose.

Related: Transparency, Explainability
AI ethics

The study of what is right and fair in building and using AI, covering values such as fairness, autonomy, non-harm, dignity, privacy and sustainability.

Related: OECD AI Principles, Accountability
Alignment

The problem of making an AI system's behaviour match human intent and values, especially as models grow more capable and their outputs become harder to predict.

Related: Existential risk, Human-in-the-loop
Existential risk

The contested possibility that highly capable AI could cause harm severe enough to threaten civilisation or human survival, a claim that shapes precautionary policy proposals.

Related: Alignment, Open-weight model
Deepfake

Images, audio or video produced or altered by AI to show people saying or doing things they never did, creating risks of fraud, defamation and election manipulation.

Related: Generative AI, Content provenance
Content provenance

Methods such as watermarking and signed metadata that record how a piece of media was created, helping people and platforms tell AI-generated content from authentic material.

Related: Deepfake, Transparency
Copyright and AI training data

The unresolved question of whether copying protected works to train models is lawful, and how licensing, attribution and the status of generated outputs should work.

Related: Training data, Generative AI
Open-weight model

A model whose trained parameters are released for others to download, inspect or adapt, which spreads benefits and scrutiny but removes the original developer's control over misuse.

Related: General-purpose AI model, Existential risk
Agentic AI

AI that can plan and carry out sequences of actions toward a goal with limited human involvement, raising questions about control, liability and unintended side effects.

Related: Alignment, Human-in-the-loop
Environmental impact of AI

The energy, water and hardware that training and running AI consume, together with the resulting emissions and land use, now a factor in sustainability decisions.

Related: AI life cycle, Foundation model
Labour market impact

Changes in the number, mix and quality of jobs as AI automates or augments tasks, shifting which skills are valued and who bears the transition costs.

Related: AI ethics, Agentic AI

AIGP study plan

8 weeks · About 7 hours per week

Week 1

AI life cycle and AI risk management

  • Map every stage of the AI life cycle (problem definition, data collection and processing, model development, verification and validation, deployment, operation and monitoring, decommissioning) and identify the governance decision points at each stage

  • Differentiate the core risk management frameworks: NIST AI RMF (Govern, Map, Measure, Manage), ISO/IEC 23894, and ISO/IEC 42001

  • Define an accountability structure for AI governance, assigning decision rights to a governing body, system owners, data stewards, and independent reviewers

  • Build a working vocabulary of at least 25 AI governance terms used across the four domains

Your checkpoint

A one-page life-cycle table with all seven stages and at least two governance decisions per stage (14+ entries), plus a 25-term glossary

Week 2

AI life cycle and AI risk management

  • Classify the principal AI risk categories (performance and accuracy, bias and fairness, privacy, security and adversarial, intellectual property, third-party and supply chain, drift and model decay) and pair each with at least two controls

  • Apply the NIST AI RMF functions to a single hypothetical high-risk use case from scoping through monitoring

  • Score risks on likelihood and impact and justify residual-risk acceptance thresholds

  • Specify the documentation artifacts that evidence risk treatment: model cards, system cards, data sheets, algorithmic impact assessments, and DPIAs

Your checkpoint

A 10-row risk register with likelihood/impact scores, named owners, controls, residual rating, and review cadence

Week 3

AI life cycle and AI risk management & AI laws, regulation and frameworks for responsible governance

  • Outline the EU AI Act risk tiers (prohibited, high-risk, transparency-only, minimal) and list the obligations attached to each

  • Distinguish the duties of providers, deployers, importers, and distributors for high-risk systems

  • Compare the GDPR obligations that most often attach to AI systems: lawful basis, data protection impact assessment, and rights around solely automated decision-making

  • Trace how a risk assessment output feeds into technical documentation and post-market monitoring obligations

Your checkpoint

A obligations matrix with four role columns and at least 12 obligation rows for high-risk AI systems, each row citing the source requirement

Week 4

AI laws, regulation and frameworks for responsible governance

  • Compare the scope and structure of the NIST AI RMF, OECD AI Principles, UNESCO Recommendation, ISO/IEC 42001, and the G7 Hiroshima process, noting where they overlap and diverge

  • Explain conformity assessment, CE marking, technical documentation, and post-market monitoring as they apply to high-risk systems

  • Survey the US landscape: federal agency activity, state-level AI statutes, and sectoral regulators such as the FTC, EEOC, FDA, and CFPB

  • Cross-walk organizational policy controls against at least two external frameworks to expose gaps

Your checkpoint

A crosswalk matrix mapping NIST AI RMF subcategories to ISO/IEC 42001 clauses with at least 15 mapped rows and a gaps column

Week 5

Foundations of AI systems, their use cases and impacts

  • Distinguish model families and learning paradigms: supervised, unsupervised, and reinforcement learning; classical versus deep learning; generative versus discriminative systems; and where agentic patterns fit

  • Interpret performance and evaluation metrics (precision, recall, F1, AUC, calibration, drift indicators) and explain the governance consequence of each failure mode

  • Catalogue the organization's AI use cases and classify each by life-cycle stage, risk tier, and affected stakeholder group

  • Summarize impact categories: labor and workforce, accessibility, environmental and compute footprint, economic concentration, and effects on specific communities

Your checkpoint

A use-case inventory of 8 systems, each tagged with life-cycle stage, risk tier, metric of record, and impacted stakeholder groups

Week 6

Foundations of AI systems, their use cases and impacts & Emerging and debated issues in AI governance

  • Evaluate explainability and interpretability approaches (feature attribution, surrogate models, inherently interpretable models) and state the tradeoffs against accuracy and security

  • Analyze competing fairness definitions and harm types, including group versus individual fairness, disparate impact, allocation harms, and representational harms

  • Debate automation versus augmentation and specify what meaningful human oversight requires in practice

  • Assess transparency obligations to end users, including disclosure of AI interaction and synthetic content

Your checkpoint

An 800-word position memo recommending a specific human-oversight design for one high-risk use case, citing at least three sources

Week 7

Emerging and debated issues in AI governance

  • Examine privacy-enhancing technologies, data provenance, and consent debates in model training and inference

  • Analyze intellectual property and copyright questions arising from training data and generated outputs

  • Assess environmental impact measurement and sustainability reporting expectations for AI systems

  • Evaluate emergent governance debates: frontier model evaluation, open-weight release, agentic autonomy, and allocation of liability across the value chain

Your checkpoint

A debate brief covering four contested issues, each with pro and con positions and a recommended governance position with rationale

Week 8

Foundations of AI systems, their use cases and impacts & AI laws, regulation and frameworks for responsible governance & AI life cycle and AI risk management & Emerging and debated issues in AI governance

  • Complete two full-length timed practice exams under exam conditions

  • Log every missed item by domain and subdomain and write the corrected rationale for each

  • Re-study the three weakest subdomains identified by the error log

  • Consolidate the framework comparison, life-cycle table, and risk register into a single rapid-review sheet

Your checkpoint

An error log of at least 30 missed questions categorized by domain with corrected rationales, plus a two-page rapid-review sheet

Milestones

Week 2: Score at least 75% on a 25-question risk management self-quiz and produce a risk register that a reviewer can trace from risk to control to owner without asking questions

Week 4: Recite the EU AI Act risk tiers and their obligations from memory, and score at least 80% on a 40-question set covering domains 2 and 3

Week 6: Score at least 75% on a 30-question mixed set covering domains 1 and 4, and submit a position memo that names the oversight mechanism, the escalation path, and the failure it prevents

Week 8: Score at least 85% on both timed full-length practice exams with no domain below 75%

AIGP exam day

Your checklist

Government-issued photo ID

Bring the original physical ID (not a photocopy). The name on it must match your exam registration exactly - if your registration uses a different form of your name, fix it with the provider before exam day. Confirm with the provider which ID types and name-matching rules apply, including whether a digital ID is accepted. [needs_review]

Appointment confirmation

Have your appointment/registration number, exam date and time, and either the test-centre address with directions or the remote-proctoring check-in link saved offline (screenshot or printout) so a browser or email outage does not derail you.

Confirmed local start time

Convert the appointment time to your own time zone and note any check-in window. Arrive or log in early - late arrivals may be refused entry or forfeit the fee. Confirm the exact cut-off and reschedule policy with the provider. [needs_review]

Private, quiet testing space (remote sittings)

Book a room you can close off, with a door, a clear desk, good lighting, and no other people or pets entering. Proctors commonly require a room scan, so remove or stow anything you would not want filmed.

Stable internet plus a backup

Wired or strong Wi-Fi, with a phone hotspot or second connection as fallback. Run the provider's system/equipment test in advance and note the technical-support number for exam day.

Working webcam and microphone (remote sittings)

Test both the day before. Have a way to show the proctor your desk, screen, and room if asked. Confirm with the provider whether an external camera or headset is permitted or required. [needs_review]

Desk cleared of prohibited items

Remove phones, smart watches, second screens, notes, books, sticky notes, and loose paper from arm's reach unless the provider explicitly allows them. Silence and stow your phone out of the room for remote sittings.

Approved scratch materials

Many certification providers supply an on-screen or physical noteboard and prohibit bringing your own paper or pen. Confirm the rule with the provider rather than assuming. [needs_review]

Calculator status confirmed

Confirm whether any calculator is permitted. Governance exams are typically conceptual rather than computational, so a calculator is likely unnecessary - verify rather than bring one unprompted. [needs_review]

Water and a light snack

Have water in a clear container and a snack ready for before or after the exam. Confirm whether drinks are allowed at the desk during the exam. [needs_review]

Provider rules read in advance

Read the current candidate handbook or testing-rules page for conduct, break, and scoring policy. Policies change, and the version in force on your exam date governs.

A settled, unhurried start

Plan travel or check-in so you are seated with time to spare, and sleep normally the night before rather than cramming new material in the final hours.

Plan the day

Final 72 hours

Stop adding new material. Review your own summary sheets of the domains you find weakest, re-read the ethics/governance frameworks and terminology you keep confusing, and do a moderate set of practice questions to rehearse pacing rather than to learn new content.

Final 72 hours

Logistics sweep: verify your registration name and appointment details, confirm the test-centre address or remote check-in link, run any system or equipment test the provider offers, and confirm ID requirements.

Final 72 hours

Protect your schedule. Clear the evening before and the morning of, and decide in advance how you will travel or set up your room.

Final 24 hours

Light review only - skim your one-page sheet of frameworks, key definitions, and role/responsibility mappings. Do not start a new practice exam or a new textbook chapter.

Final 24 hours

Prepare physically: lay out your ID and confirmation, charge devices, clear the desk, check the room, and set two alarms. Confirm the start time in your own time zone one last time.

Final 24 hours

Eat normally, avoid heavy alcohol and late caffeine, and aim for a full night's sleep. Fatigue costs more marks than one more hour of revision gains.

Morning of

Eat a normal meal, hydrate, and do a 15–20 minute skim of your summary sheet only. Avoid group chats or forums discussing exam content - they raise anxiety and may touch on confidentiality issues.

Morning of

Arrive at the centre early, or for a remote sitting log in for check-in ahead of the stated window, with ID in hand, desk cleared, room closed, and phone stowed away.

During the test

Read the full stem before the options; these exams reward careful reading of 'best', 'most likely', and 'first step' phrasing. Answer every question - there is no advantage to leaving blanks unless the provider penalises wrong answers, which is unusual.

During the test

Pace in thirds: note the total time and question count, and check your progress at roughly the one-third and two-thirds marks. Flag and move on from anything that stalls you for more than about 90 seconds.

During the test

Reserve the last portion of the time for flagged items and a check that nothing is unanswered. Use breaks sparingly if permitted - the clock generally keeps running or the break time is deducted, so confirm the break rule beforehand. [needs_review]

After the test

Note how and when results are released and what your retake options and waiting period are, so you are not scrambling if you need a second attempt. Confirm with the provider. [needs_review]

Rules to check

  • Bring the original government-issued photo ID in the name used at registration. Admittance rules and accepted ID types vary - confirm with the provider rather than assuming a digital or expired document will work.

  • The exam is confidential and almost certainly covered by a non-disclosure agreement. Do not reproduce, discuss, or post actual exam questions or answers, before or after the sitting.

  • You will be proctored - in person or by remote proctor with camera and screen monitoring. Expect identity verification, a room or desk scan, and observation throughout. Behaviours such as reading aloud, muttering, looking off-screen repeatedly, or reaching out of frame may be flagged.

  • Keep only permitted items at your desk. Phones, smart watches, notes, and secondary screens are typically prohibited; unauthorised materials can void your result and may lead to a misconduct finding. Confirm the specific list with the provider. [needs_review]

  • Do not leave the camera view or the testing room during the exam unless a break is expressly permitted. If something goes wrong technically, stay visible and contact the proctor or support line rather than troubleshooting silently.

  • Scoring is criterion-referenced: you are measured against a pre-set standard, not against other candidates, so a hard-feeling paper is not evidence you failed. Confirm the passing standard, score scale, and reporting timeline with the provider. [needs_review]

  • Know the retake and reschedule rules before you sit. There is usually a fee, a waiting period, and a limit on attempts within a given window - confirm the current terms with the provider. [needs_review]

  • Report any irregularity - a technical failure, an interruption, a proctor issue - to the provider promptly and in writing, so there is a record if you later need to request a remedy.

AIGP FAQ

How many AIGP practice questions are available?

The practice set has 200 AIGP questions. Each one includes the choices, the correct answer, and an explanation, and the set is free to read with no signup.

How many AIGP flashcards are available?

The flashcard set has 32 AIGP cards. Each card shows the term and its definition.

Which topics does the AIGP study material cover?

The AIGP study material has 4 sections, including Foundations of AI Systems, Their Use Cases and Impacts, AI Laws, Regulation and Frameworks for Responsible Governance, and AI Life Cycle and AI Risk Management.