AWS AI Practitioner Practice Test

AWS Certified AI Practitioner

AWS AI Practitioner overview

  • 200 practice questions
  • 32 flashcards
  • 5 study guide topics
  • 38 key terms

What you'll study

  • Fundamentals of AI and ML20%
  • Fundamentals of Generative AI24%
  • Applications of Foundation Models28%
  • Guidelines for Responsible AI14%
  • Security, Compliance, and Governance for AI Solutions14%

The AWS Certified AI Practitioner is an entry-level credential from Amazon Web Services that validates foundational knowledge of artificial intelligence, machine learning, and generative AI on AWS. It is designed for people who work with AI-enabled tools and services rather than build models from scratch - product managers, business analysts, sales and marketing staff, developers new to AI, and anyone translating AI capabilities into business outcomes. The exam covers five areas: fundamentals of AI and ML, fundamentals of generative AI, applications of foundation models, guidelines for responsible AI, and security, compliance, and governance for AI solutions. Expect conceptual questions about when to use which AWS service, prompt engineering and model selection trade-offs, bias and fairness concerns, and how shared responsibility applies to AI workloads. Preparation typically means working through AWS's own training materials, hands-on experimentation in an AWS account with services such as Amazon Bedrock, and reviewing responsible-AI guidance. The credential is not a substitute for the associate or professional ML certifications; it is a starting point. This page is independent study material and is not affiliated with, endorsed by, or sponsored by Amazon Web Services. AWS, and all related marks, are trademarks of their respective owner.

Common mistakes

Treating AI, ML, deep learning, and generative AI as interchangeable labels and picking whichever option sounds most advanced, instead of mapping each scenario to the correct level of the hierarchy.

Marketing language blurs the terms, and the exam deliberately offers options where a broader or narrower term is technically wrong for the described system (for example calling a rules-based classifier 'generative AI').

What to do instead

Before answering, restate the scenario in one sentence and classify it: is there learning from data (ML), neural networks with many layers (deep learning), or a model that produces new content such as text or images (generative AI)? Eliminate any option whose term is broader or narrower than that classification, and be able to state that generative AI is a subset of deep learning, which is a subset of ML, which is a subset of AI.

Confusing supervised, unsupervised, and reinforcement learning by focusing on the data format rather than on whether labels exist and what signal drives learning.

Candidates memorise example algorithms (clustering equals unsupervised) but freeze when the scenario describes a business problem in plain language, such as grouping customers without prior categories or rewarding an agent for winning a game.

What to do instead

Ask two questions for every scenario: are the training examples labelled with the correct answer, and is the model learning from trial-and-error reward? Labelled equals supervised (classification or regression); unlabelled structure-finding equals unsupervised (clustering, anomaly detection); reward-driven sequential decisions equal reinforcement learning. Anchor each answer to that decision tree rather than to algorithm names.

Assuming a foundation model already knows proprietary or current facts, and choosing an answer that skips retrieval or fine-tuning when the scenario requires company-specific or up-to-date knowledge.

The word 'foundation' suggests completeness, and candidates underestimate that model knowledge is frozen at training time and limited to public data.

What to do instead

Whenever a scenario mentions internal documents, recent events, or verifiable citations, look for an option that supplies external context at inference time, such as retrieval augmented generation over a knowledge base, rather than relying on the model's parameters. Reserve fine-tuning for changing behaviour, style, or task format, and treat 'the model will already know it' as a distractor.

Mixing up fine-tuning, retrieval augmented generation, and prompt engineering when the question asks for the most appropriate customisation approach under constraints such as cost, latency, or data volume.

All three adapt a model to a task, so candidates pick the most powerful-sounding one instead of matching the technique to the constraint described.

What to do instead

Build a mental table: prompt engineering changes instructions only, needs no training, and is cheapest and fastest to iterate; retrieval augmented generation adds external knowledge without changing weights and suits factual grounding and freshness; fine-tuning updates weights, needs labelled examples and compute, and suits consistent style, format, or domain behaviour. Then reread the constraint words in the stem (no labelled data, must cite sources, must minimise cost) and let them select the row.

Answering temperature and inference-parameter questions with intuition about 'more creativity is better' instead of the specific behaviour each parameter controls.

Parameters such as temperature, top-p, and maximum length sound similar, and candidates have not practised predicting output changes from parameter changes.

What to do instead

Memorise the direction of each effect: higher temperature increases randomness and diversity while lower temperature makes output more deterministic and repeatable; top-p restricts sampling to a cumulative probability mass; maximum length caps output tokens and can truncate responses. For each scenario, identify whether the requirement is consistency, variety, or length control, then choose the single parameter that governs it.

Choosing a model by benchmark reputation alone when the scenario supplies concrete constraints such as latency, cost per token, context window, or modality.

Candidates default to the largest or most famous model, missing that the exam tests matching model characteristics to workload requirements.

What to do instead

Extract every constraint from the stem into a checklist (input modality, output modality, context length, latency budget, cost sensitivity, data residency). Eliminate any option that violates a hard constraint first, then compare survivors on the remaining trade-offs. Remember that smaller models are often correct answers for high-volume, low-complexity tasks, and that multimodal requirements immediately exclude text-only models.

Treating responsible AI as a single compliance checkbox rather than distinguishing bias, fairness, explainability, privacy, safety, and hallucination as separate concerns with separate mitigations.

The principles overlap in casual discussion, so candidates match a mitigation to the wrong principle, for example citing transparency for a harmful-content problem.

What to do instead

For each scenario, name the harm in one phrase before looking at the options: skewed outcomes across groups (bias and fairness), inability to justify a decision (explainability), exposure of personal data (privacy), offensive or dangerous output (safety), fabricated facts (hallucination), or undisclosed AI use (transparency). Then select the mitigation that targets that named harm, and reject options that address a different principle even if they are generally good practice.

Assuming that removing sensitive attributes such as gender or ethnicity from training data automatically eliminates bias.

It feels like an obvious privacy and fairness fix, but correlated proxy features can still encode the excluded attribute, producing discriminatory outcomes.

What to do instead

Recognise proxy-variable questions: when an option says 'simply drop the protected attribute', check whether the stem mentions correlated fields such as postcode or job title. Prefer answers that measure outcomes across groups, evaluate for disparate impact, and use techniques such as rebalancing data, adjusting thresholds, or testing with subgroup metrics, and be ready to explain that fairness requires ongoing monitoring rather than a one-time data edit.

Confusing the shared responsibility model for AI services, attributing data protection, access control, or content filtering entirely to the provider or entirely to the customer.

Cloud security habits from infrastructure exams do not transfer cleanly, and managed AI services blur the line between provider-managed model security and customer-managed data and configuration.

What to do instead

Split every scenario into provider duties (security of the service, underlying infrastructure, and the base model) and customer duties (data classification, encryption choices, identity and access policies, prompt and output safeguards, and monitoring usage). When an option claims the provider will handle customer data governance or that the customer must patch the model, treat it as a shared-responsibility error and choose the option that assigns the duty to the correct party.

Overlooking governance and auditability requirements, such as logging model inputs and outputs, documenting training data provenance, or tracking model versions, when the scenario mentions regulated industries or audits.

Candidates focus on getting the model to work and treat documentation and logging as optional extras, so they miss the option that satisfies traceability.

What to do instead

Scan the stem for trigger words such as audit, regulation, incident investigation, reproducibility, or accountability. When present, prefer answers that enable traceability: capturing invocation logs with metadata, maintaining a model card or data lineage record, versioning models and prompts, and defining human review and approval steps. Pair this with the principle that governance controls are configured by the customer even when the model is provider-managed.

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

AWS AI Practitioner 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 1Fundamentals of AI and ML

    A company wants a system that can identify the sentiment of customer reviews without any labeled training examples. Which approach best fits this requirement?

    1. Applying k-means clustering to the review text
    2. Training a supervised classifier on labeled reviews
    3. Using a pretrained foundation model with zero-shot prompting Correct answer.
    4. Building a decision tree from historical ratings

    Explanation

    Zero-shot prompting lets a pretrained foundation model perform a task from instructions alone, without task-specific labeled data. Supervised training and decision trees require labeled examples, so they do not satisfy the no-label constraint.

  2. Question 2Fundamentals of AI and ML

    Which statement best describes the relationship between artificial intelligence, machine learning, and deep learning?

    1. The three terms describe identical techniques and are interchangeable
    2. Machine learning is a subset of deep learning, which is a subset of artificial intelligence
    3. Deep learning is a subset of machine learning, which is a subset of artificial intelligence Correct answer.
    4. Artificial intelligence is a subset of machine learning, which is a subset of deep learning

    Explanation

    Artificial intelligence is the broad field of building systems that perform tasks associated with human intelligence. Machine learning is one approach within AI, and deep learning is a specialized subset of machine learning using multi-layer neural networks.

  3. Question 3Fundamentals of AI and ML

    A data scientist splits a dataset into training, validation, and test sets. What is the primary purpose of the validation set?

    1. To provide the final unbiased estimate of model performance
    2. To store outliers that are excluded from the training process
    3. To tune hyperparameters and compare model variants during development Correct answer.
    4. To increase the total volume of data available for gradient updates

    Explanation

    The validation set is used during development to select hyperparameters and compare candidate models. The test set, kept untouched until the end, provides the final estimate of generalization performance.

  4. Question 4Fundamentals of AI and ML

    Which type of machine learning uses labeled input-output pairs to learn a mapping from features to a target?

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

    Explanation

    Supervised learning trains on examples where the correct output is provided, allowing the model to learn a mapping from inputs to targets. Unsupervised and self-supervised methods do not rely on explicit human-provided labels for the target task.

  5. Question 5Fundamentals of AI and ML

    What is the main advantage of using a managed AI service such as Amazon SageMaker instead of building training infrastructure from scratch?

    1. It eliminates the need to prepare or clean training data
    2. It automatically guarantees high accuracy on any dataset
    3. It removes the need to choose an algorithm or model type
    4. It reduces undifferentiated infrastructure work so teams can focus on models Correct answer.

    Explanation

    Managed AI platforms handle provisioning, scaling, and operational tasks so practitioners can concentrate on data and modeling. They do not remove the need for data preparation, algorithm selection, or accuracy validation.

  6. Question 6Fundamentals of AI and ML

    A model performs very well on its training data but poorly on new data. Which term describes this behavior?

    1. Underfitting
    2. Data drift
    3. Overfitting Correct answer.
    4. Feature engineering

    Explanation

    Overfitting occurs when a model captures noise and specifics of the training set instead of general patterns, so performance degrades on unseen data. Underfitting is the opposite problem, where the model is too simple to fit even the training data well.

  7. Question 7Fundamentals of AI and ML

    Which task is a classic example of unsupervised learning?

    1. Estimating a reward function from agent actions
    2. Predicting house prices from labeled sales records
    3. Grouping customers into segments based on purchase behavior without labels Correct answer.
    4. Classifying emails as spam or not spam using labeled examples

    Explanation

    Customer segmentation via clustering is unsupervised because the algorithm discovers groupings without predefined labels. Price prediction and spam classification are supervised tasks that rely on labeled targets.

  8. Question 8Fundamentals of AI and ML

    What does the term inference mean in the context of a deployed machine learning model?

    1. Updating model weights using a training dataset
    2. Measuring the size of the training dataset
    3. Converting raw data into numerical features
    4. Using a trained model to generate predictions on new inputs Correct answer.

    Explanation

    Inference is the process of running a trained model on new inputs to produce predictions or outputs. Training, by contrast, is the phase in which model weights are learned from data.

  9. Question 9Fundamentals of AI and ML

    An ML team wants to reduce the risk of bias in a credit-scoring model. Which practice most directly addresses this during development?

    1. Increasing the number of training epochs until loss converges
    2. Removing all features related to income and employment
    3. Evaluating model outcomes across demographic groups and remediating disparities Correct answer.
    4. Deploying the model to production to observe real-world results

    Explanation

    Bias mitigation requires measuring performance and outcomes across subgroups and then correcting disparities through data or modeling changes. Simply training longer or deploying early does not address fairness concerns.

  10. Question 10Fundamentals of AI and ML

    Which metric is most appropriate for evaluating a regression model that predicts delivery times in minutes?

    1. Root mean squared error Correct answer.
    2. Precision
    3. F1 score
    4. Area under the ROC curve

    Explanation

    Regression models predict continuous values, and root mean squared error measures the average magnitude of prediction error in the target's units. ROC AUC, F1, and precision are classification metrics that assume discrete class labels.

  11. Question 11Fundamentals of AI and ML

    A team needs a model that can answer questions about internal documents and cite the source passages. Which technique is designed for this?

    1. Fine-tuning an image classifier on document scans
    2. Retrieval-augmented generation that fetches relevant passages before answering Correct answer.
    3. Training a reinforcement learning agent on user clicks
    4. Clustering documents by length

    Explanation

    Retrieval-augmented generation combines a search step over a document store with a generative model, allowing answers grounded in retrieved passages. This supports source citation and reduces unsupported generation.

  12. Question 12Fundamentals of AI and ML

    What is the primary role of a loss function during model training?

    1. To measure the throughput of the inference endpoint
    2. To select which features are removed from the dataset
    3. To store the training dataset in compressed form
    4. To quantify prediction error so optimization can adjust model parameters Correct answer.

    Explanation

    A loss function assigns a numeric penalty to the difference between predictions and targets. Optimization algorithms minimize this loss by adjusting model parameters during training.

  13. Question 13Fundamentals of AI and ML

    A company must keep sensitive patient data on premises but wants to use a cloud-based foundation model. Which approach best satisfies both needs?

    1. Use a private deployment or a service that processes data within a controlled environment Correct answer.
    2. Upload the full patient records to the model provider for fine-tuning
    3. Use the model only for public marketing content
    4. Abandon AI and rely on manual review

    Explanation

    Private deployment or services that keep data within a controlled environment allow the organization to use foundation models while maintaining data residency and confidentiality. Uploading sensitive records to a third party or avoiding AI entirely does not balance both requirements.

  14. Question 14Fundamentals of AI and ML

    Which statement accurately distinguishes generative AI from traditional predictive ML?

    1. Generative AI only classifies inputs, while predictive ML only creates new content
    2. Generative AI requires no training data, while predictive ML always requires labeled data
    3. Generative AI cannot be used for text, only for numeric forecasting
    4. Generative AI creates new content such as text or images, while predictive ML estimates labels or values Correct answer.

    Explanation

    Generative models learn data distributions and produce new content, whereas traditional predictive models output a label, class, or numeric estimate. Both families rely on training data and can be applied to text or numeric domains.

  15. Question 15Fundamentals of AI and ML

    A recommendation team notices that their model keeps recommending only popular items, reducing discovery for users. Which concept does this illustrate?

    1. Feature leakage
    2. Overfitting to the test set
    3. Feedback loop and popularity bias Correct answer.
    4. Gradient vanishing

    Explanation

    Recommending popular items increases their engagement, which feeds back as training signal and further amplifies their popularity. This self-reinforcing cycle is a feedback loop that produces popularity bias.

  16. Question 16Fundamentals of AI and ML

    Which AWS service provides a fully managed environment for building, training, and deploying machine learning models?

    1. Amazon SageMaker Correct answer.
    2. Amazon CloudWatch
    3. AWS Identity and Access Management
    4. Amazon Route 53

    Explanation

    Amazon SageMaker is a managed platform that covers the machine learning lifecycle, including data preparation, training, tuning, and deployment. The other services provide monitoring, access control, and DNS rather than ML workflows.

  17. Question 17Fundamentals of AI and ML

    A team wants to improve a model's performance by creating new input variables from raw data, such as extracting day-of-week from a timestamp. What is this practice called?

    1. Cross-validation
    2. Model quantization
    3. Feature engineering Correct answer.
    4. Regularization

    Explanation

    Feature engineering is the process of using domain knowledge to derive new input variables that make patterns easier for a model to learn. Regularization, quantization, and cross-validation serve different purposes in training and evaluation.

  18. Question 18Fundamentals of AI and ML

    Which scenario is the best fit for reinforcement learning rather than supervised learning?

    1. Translating documents using paired source and target sentences
    2. Detecting fraudulent transactions using labeled chargebacks
    3. Predicting tomorrow's temperature from historical sensor readings
    4. Teaching an agent to navigate a maze through trial and error with rewards Correct answer.

    Explanation

    Reinforcement learning suits sequential decision problems where an agent learns from rewards and penalties through interaction. The other scenarios involve labeled examples and are better handled by supervised methods.

  19. Question 19Fundamentals of AI and ML

    What does the F1 score balance in a binary classification problem?

    1. Training time and inference latency
    2. Precision and recall Correct answer.
    3. Bias and variance
    4. Model size and accuracy

    Explanation

    The F1 score is the harmonic mean of precision and recall, giving a single number that balances false positives and false negatives. This makes it useful when classes are imbalanced and both error types matter.

  20. Question 20Fundamentals of AI and ML

    A team is choosing between a large foundation model and a small task-specific model for a high-volume, low-latency classification task with abundant labeled data. Which choice is generally most appropriate?

    1. An unsupervised clustering model, because labels are abundant
    2. The large foundation model, because larger models are always more accurate
    3. The small task-specific model, because it can be cheaper and faster while meeting accuracy needs Correct answer.
    4. A reinforcement learning agent, because it learns without labels

    Explanation

    When abundant labeled data exists and latency and cost are critical, a compact task-specific model often delivers sufficient accuracy at lower cost and latency. Large foundation models are better suited to open-ended or low-data tasks.

  21. Question 21Fundamentals of AI and ML

    A model trained on historical loan approvals shows lower approval rates for a protected group. Investigation finds the training labels reflect past human decisions. Which root cause best explains this outcome?

    1. The model has too few parameters to learn the pattern
    2. The training data encodes historical human bias that the model learns to reproduce Correct answer.
    3. The validation set was too small to detect accuracy
    4. The inference endpoint is under-provisioned

    Explanation

    When labels are derived from past human decisions, the model learns and perpetuates the biases present in those decisions. This is a data-provenance problem rather than a capacity, infrastructure, or validation-size issue.

  22. Question 22Fundamentals of AI and ML

    A company wants its chatbot to answer only from an approved knowledge base and refuse unrelated questions. Which combination best achieves this?

    1. Fine-tuning on unrelated internet text to broaden knowledge
    2. Reducing the context window to a single token
    3. Retrieval-augmented generation with guardrails that restrict scope Correct answer.
    4. Increasing the model temperature and removing system instructions

    Explanation

    Retrieval-augmented generation grounds responses in approved documents, while guardrails constrain the model to the intended scope and block out-of-domain requests. Higher temperature or broader fine-tuning would increase, not reduce, off-topic responses.

  23. Question 23Fundamentals of AI and ML

    An ML engineer notices that a deployed model's accuracy degrades gradually over months even though the code has not changed. Which explanation is most likely?

    1. The training set has grown larger automatically
    2. The model weights are being corrupted by inference calls
    3. The loss function has been silently replaced
    4. Data drift has shifted the input distribution away from training data Correct answer.

    Explanation

    Data drift occurs when the statistical properties of real-world inputs change over time, so patterns learned during training no longer match production data. This causes gradual accuracy decline without any code change.

  24. Question 24Fundamentals of AI and ML

    A team must decide whether to fine-tune a foundation model or use prompt engineering for a new text-classification task with only a few hundred labeled examples. Which reasoning is soundest?

    1. Neither approach can work without millions of labeled examples
    2. Fine-tuning is always preferred because it uses more compute
    3. Prompt engineering with a pretrained model is a reasonable first step given the small dataset Correct answer.
    4. The team should train a model from scratch to guarantee control

    Explanation

    With limited labeled data, prompt engineering or few-shot prompting on a pretrained model is a low-cost first step that often performs well. Fine-tuning becomes more attractive as labeled data and the need for specialization grow.

  25. Question 25Fundamentals of AI and ML

    Which statement best captures why human oversight remains important in high-stakes AI systems?

    1. Models can reflect training-data limitations and should be reviewed before consequential decisions Correct answer.
    2. Models are guaranteed to be correct once accuracy exceeds ninety percent
    3. Oversight is unnecessary when the model is hosted on managed services
    4. Human review is only needed for image models

    Explanation

    Models inherit limitations and biases from their training data and can fail in unexpected ways, so human review is essential for consequential decisions. High aggregate accuracy does not guarantee correctness for every case or subgroup.

  26. Question 26Fundamentals of AI and ML

    In supervised learning, what is the essential role of labeled training data?

    1. It supplies input-output pairs so the model can learn a mapping that generalizes to new inputs. Correct answer.
    2. It defines a reward signal that the agent maximizes through trial and error.
    3. It stores the final model weights so predictions can be served without retraining.
    4. It provides unlabeled examples so the model can discover hidden clusters on its own.

    Explanation

    Supervised learning learns a function from inputs to known target outputs, so each training example must pair features with the correct label. The goal is generalization to unseen inputs, not memorization of the training set.

  27. Question 27Fundamentals of AI and ML

    A hospital wants an AI system to flag suspicious regions in radiology images and also show which pixels drove each flag so radiologists can verify the reasoning. Which concept does the second requirement describe?

    1. Data augmentation
    2. Model interpretability Correct answer.
    3. Batch normalization
    4. Hyperparameter tuning

    Explanation

    Interpretability concerns how clearly a human can understand why a model produced a given output. Highlighting influential input regions makes the model's reasoning inspectable, which supports trust and review in clinical settings.

  28. Question 28Fundamentals of AI and ML

    A team needs to split data for a model that predicts next month's sales, and the newest months are the most representative of future conditions. Which splitting strategy is most appropriate?

    1. Randomly shuffle all rows and assign 80 percent to training and 20 percent to testing.
    2. Train on earlier periods and test on later periods in chronological order. Correct answer.
    3. Train on later periods and test on earlier periods to check historical accuracy.
    4. Use the same rows for both training and testing to maximize measured accuracy.

    Explanation

    For time-dependent data, the test set should mimic the future the model will face, so training on past periods and testing on later ones reflects realistic deployment. Random shuffling can leak future information into training and inflate results.

  29. Question 29Fundamentals of AI and ML

    A model must assign each incoming support ticket to exactly one of five categories, and every ticket already has one correct category recorded. Which learning setup fits?

    1. Clustering
    2. Regression
    3. Binary classification
    4. Multiclass classification Correct answer.

    Explanation

    The task predicts one label out of five mutually exclusive categories from labeled examples, which is multiclass classification. Binary classification would apply only with two possible labels.

  30. Question 30Fundamentals of AI and ML

    A data scientist wants to estimate how well a trained model will perform on unseen data before deployment, using data the model has never learned from. Which practice provides this estimate?

    1. Deploying the model and waiting for customer complaints
    2. Increasing the number of training epochs until training loss approaches zero
    3. Evaluating the model on a held-out validation or test set Correct answer.
    4. Removing outliers from the training data before fitting

    Explanation

    A held-out set is not used during training, so its error approximates generalization performance on new data. Training-set metrics tend to be optimistic and cannot reveal overfitting.

  31. Question 31Fundamentals of AI and ML

    A bank wants to know which customer attributes most influenced a declined loan application so staff can explain the decision to the applicant. Which capability is needed?

    1. Feature importance or local explanation of the prediction Correct answer.
    2. Automated hyperparameter search
    3. Data labeling through crowdsourcing
    4. Horizontal scaling of inference endpoints

    Explanation

    Explaining an individual decision requires attributing the output to input features, which feature importance or local explanation methods provide. Operational techniques such as scaling or tuning do not reveal reasons for a specific outcome.

  32. Question 32Fundamentals of AI and ML

    A company has a small labeled dataset of product reviews and wants to know whether a simple model is learning real patterns or just memorizing. Training accuracy is 99 percent while validation accuracy is 62 percent. What does this gap most likely indicate?

    1. The labels in the training data are all correct
    2. The model is overfitting the training data Correct answer.
    3. The model is underfitting the training data
    4. The validation set is too large to be useful

    Explanation

    A large gap with near-perfect training accuracy and much lower validation accuracy is the classic signature of overfitting, where the model captures noise specific to the training rows. Underfitting would show poor accuracy on both sets.

  33. Question 33Fundamentals of AI and ML

    Which description correctly separates a parameter from a hyperparameter in a neural network?

    1. Parameters control only the learning rate, while hyperparameters are the network weights.
    2. Parameters are set by the engineer before training, while hyperparameters are learned from data.
    3. Parameters and hyperparameters are both learned automatically and differ only in name.
    4. Parameters are learned from data during training, while hyperparameters are set before training to control the process. Correct answer.

    Explanation

    Weights and biases are parameters adjusted by the optimization algorithm using training data. Hyperparameters, such as learning rate, batch size, and layer count, are chosen beforehand and govern how training proceeds.

  34. Question 34Fundamentals of AI and ML

    A team must choose an evaluation metric for a fraud detector where missing a fraudulent transaction is far costlier than flagging a legitimate one for review. Which metric should they prioritize?

    1. Precision for the fraud class
    2. Overall accuracy
    3. Mean squared error
    4. Recall for the fraud class Correct answer.

    Explanation

    Recall measures the share of actual fraud cases that were caught, directly reflecting the cost of missed fraud. Accuracy can look high while missing most fraud when fraud is rare, and precision mainly reflects false alarms.

  35. Question 35Fundamentals of AI and ML

    An engineer must explain why a spam filter's output changes when one word is replaced, and wants a method that perturbs inputs and observes output changes without inspecting model internals. Which category of technique applies?

    1. White-box gradient inspection
    2. Model-agnostic explanation methods Correct answer.
    3. Synthetic minority oversampling
    4. Grid search cross-validation

    Explanation

    Methods that probe a model by altering inputs and observing outputs work regardless of the model's internal structure, so they are model-agnostic. They are useful when the model is a black box or access to internals is restricted.

  36. Question 36Fundamentals of AI and ML

    A team must decide whether a candidate model is ready for deployment. They want a single procedure that repeatedly resamples the training data, trains the model on each split, and reports the spread of validation scores. Which approach matches this requirement?

    1. k-fold cross-validation Correct answer.
    2. Confusion matrix analysis on the training set
    3. A single train-test split
    4. Principal component analysis

    Explanation

    k-fold cross-validation rotates which fold is held out, producing multiple validation scores whose spread indicates stability and sensitivity to the data split. A single split gives only one estimate and no measure of variability.

  37. Question 37Fundamentals of AI and ML

    A retail team has millions of unlabeled transaction records and wants to group customers by purchasing behavior without predefined segments. Which approach fits, and why?

    1. Binary classification, because each customer is either a buyer or not
    2. Clustering, because it discovers structure in data without predefined labels Correct answer.
    3. Reinforcement learning, because the model learns from customer rewards
    4. Regression, because it predicts a numeric segment value for each customer

    Explanation

    Clustering is an unsupervised technique that groups similar records based on patterns in the features, requiring no predefined segment labels. The absence of labels rules out supervised classification and regression.

  38. Question 38Fundamentals of AI and ML

    Which statement best describes why foundation models are often described as general-purpose compared with traditional task-specific models?

    1. A foundation model can only ever produce numeric regression outputs.
    2. A foundation model is trained separately from scratch for every downstream task it supports.
    3. A foundation model requires no data at all to perform a new task.
    4. A single pretrained foundation model can be adapted to many downstream tasks through prompting or fine-tuning. Correct answer.

    Explanation

    Foundation models are pretrained on broad data and then adapted to many tasks, which is what makes them general-purpose. Task-specific models are typically trained for one narrow objective from the start.

  39. Question 39Fundamentals of AI and ML

    A team is selecting an algorithm for a tabular dataset with a nonlinear decision boundary and limited training time. They want a model that builds an ensemble of decision trees and usually performs well with little tuning. Which choice fits best?

    1. A rule-based expert system
    2. A gradient-boosted tree ensemble Correct answer.
    3. Linear regression
    4. k-means clustering

    Explanation

    Tree ensembles capture nonlinear relationships and interactions in tabular data and are known for strong performance with modest tuning effort. Linear regression assumes a linear relationship, and k-means is unsupervised.

  40. Question 40Fundamentals of AI and ML

    A product team wants to add a feature that suggests a likely next word as a user types, and they have a large corpus of unlabeled text but no annotated examples. Which learning approach fits this situation?

    1. Unsupervised clustering, where words are grouped by frequency and each cluster becomes a suggestion
    2. Supervised learning, where each sentence must first be labeled by a human with its intended meaning
    3. Reinforcement learning, where the model receives a reward each time a user accepts a suggestion
    4. Self-supervised learning, where the model learns by predicting parts of the text from the rest of the text Correct answer.

    Explanation

    Self-supervised learning creates its own labels from the structure of unlabeled data, for example hiding the next word and asking the model to predict it. This lets a model learn language patterns from raw text without human annotation, which matches a next-word suggestion feature. The other approaches either require labels that do not exist here or do not produce a predictive language model.

  41. Question 41Fundamentals of Generative AI

    Which statement best describes a foundation model in generative AI?

    1. A small model trained only on one company's private documents
    2. A rule-based system that generates responses from hand-written templates
    3. A large model pretrained on broad data that can be adapted to many downstream tasks Correct answer.
    4. A statistical model used only for ranking search results

    Explanation

    Foundation models are trained on very large, broad datasets and learn general patterns that transfer to many tasks. Because of this generality, they can be adapted through prompting or fine-tuning instead of being built from scratch for each use case.

  42. Question 42Fundamentals of Generative AI

    A team wants a model to summarize customer emails in a specific house style. Which approach most directly teaches the model that style using labeled example pairs?

    1. Switching to a smaller tokenizer
    2. Fine-tuning on curated input-output examples Correct answer.
    3. Increasing the temperature setting
    4. Reducing the context window

    Explanation

    Fine-tuning adjusts model weights using labeled examples so the model learns a desired style or task behavior. Providing many curated input-output pairs lets the model internalize the target format and tone beyond what a prompt alone conveys.

  43. Question 43Fundamentals of Generative AI

    In a retrieval-augmented generation (RAG) system, what is the primary role of the retrieval component?

    1. To compress the model weights for cheaper inference
    2. To score the fluency of the generated response
    3. To fetch relevant source passages that are inserted into the prompt Correct answer.
    4. To generate the final answer without any external data

    Explanation

    RAG separates knowledge lookup from generation: a retriever finds relevant passages from an external store and passes them into the prompt. The generator then conditions its answer on those passages, which grounds responses in current or private data.

  44. Question 44Fundamentals of Generative AI

    Which term describes the tendency of a language model to produce confident but factually incorrect statements?

    1. Hallucination Correct answer.
    2. Latency
    3. Tokenization
    4. Overfitting

    Explanation

    Hallucination refers to generated content that sounds plausible but is not supported by facts or source data. It happens because models predict likely text rather than verifying truth, so grounding techniques and evaluation are used to reduce it.

  45. Question 45Fundamentals of Generative AI

    What does a token represent in the context of large language models?

    1. A permission grant in an identity policy
    2. A cached copy of a previous model response
    3. A single byte of network traffic between client and server
    4. A unit of text, such as a word piece or character sequence, processed by the model Correct answer.

    Explanation

    Models do not read raw characters; text is split into tokens that may be whole words, word pieces, or characters. Token counts drive context limits and usage billing, so understanding them matters for prompt design and cost.

  46. Question 46Fundamentals of Generative AI

    Which capability is characteristic of diffusion models rather than autoregressive language models?

    1. Iteratively removing noise to produce an image Correct answer.
    2. Encrypting data at rest with customer keys
    3. Parsing SQL queries against a relational database
    4. Predicting the next token in a sequence

    Explanation

    Diffusion models learn to reverse a noising process, starting from random noise and progressively denoising toward a coherent output such as an image. Autoregressive models instead generate one token at a time conditioned on prior tokens.

  47. Question 47Fundamentals of Generative AI

    A developer sets the temperature parameter to a very low value. What is the most likely effect on generated text?

    1. More deterministic and repeatable outputs Correct answer.
    2. Lower token cost per request
    3. Longer maximum output length
    4. More random and diverse outputs

    Explanation

    Temperature scales the randomness of sampling: lower values concentrate probability on high-likelihood tokens, producing steadier, more repeatable text. Higher values flatten the distribution and increase variety.

  48. Question 48Fundamentals of Generative AI

    Which practice best reduces the risk of exposing sensitive data when using a hosted generative AI service?

    1. Increasing the model temperature for anonymity
    2. Disabling logging on the client application only
    3. Sending full production records to improve answer quality
    4. Redacting or excluding sensitive fields before sending prompts Correct answer.

    Explanation

    Data minimization is the core control: only send what the task requires and remove personal or confidential fields beforehand. This limits what the provider could retain or inadvertently expose, independent of sampling settings.

  49. Question 49Fundamentals of Generative AI

    A support team wants answers grounded in the latest internal policy documents without retraining the model each time a policy changes. Which design fits best?

    1. Use retrieval-augmented generation over an indexed policy store Correct answer.
    2. Fine-tune the model weekly on the full policy corpus
    3. Shorten the system prompt to save tokens
    4. Raise the top-p value so the model recalls policies

    Explanation

    RAG keeps knowledge in an external index that can be updated independently of model weights, so policy edits take effect after reindexing. This avoids repeated, costly fine-tuning cycles and keeps answers traceable to sources.

  50. Question 50Fundamentals of Generative AI

    An application must classify incoming tickets into one of five fixed categories. Which prompting approach is most appropriate?

    1. Open-ended creative writing prompt with high temperature
    2. Prompt that asks the model to invent new categories
    3. Structured prompt that lists the categories and requests one label Correct answer.
    4. Prompt that requests a long free-form essay about the ticket

    Explanation

    Classification benefits from constrained instructions that enumerate the allowed labels and specify the output format. Low randomness and explicit categories make the result easier to validate and parse programmatically.

  51. Question 51Fundamentals of Generative AI

    A team observes that model responses become inconsistent when the prompt grows very long and important instructions appear at the start. Which adjustment is most likely to help?

    1. Remove all examples from the prompt
    2. Increase the temperature to make the model more creative
    3. Move critical instructions closer to the end of the prompt and restate them concisely Correct answer.
    4. Switch the output format to free text

    Explanation

    Models can under-weight information buried in long contexts, so placing key instructions near the end and restating them improves adherence. Keeping instructions concise and salient reduces the chance they are lost among other content.

  52. Question 52Fundamentals of Generative AI

    Which metric is most appropriate for evaluating a RAG system's retriever separately from its generator?

    1. Average response latency in milliseconds
    2. Recall of relevant passages in the retrieved set Correct answer.
    3. BLEU score of the final answer
    4. Number of tokens in the system prompt

    Explanation

    Retriever quality is measured by whether relevant passages appear in the retrieved set, commonly via recall or precision at k. Answer-level metrics conflate retrieval with generation and cannot isolate retriever performance.

  53. Question 53Fundamentals of Generative AI

    What is the main purpose of a guardrail in a generative AI application?

    1. To filter or block unsafe or disallowed inputs and outputs Correct answer.
    2. To reduce the number of tokens in the vocabulary
    3. To increase the model's parameter count
    4. To train the base model from scratch

    Explanation

    Guardrails apply policy checks around model calls, screening prompts and responses for harmful, sensitive, or off-policy content. They add a control layer that operates independently of the underlying model's training.

  54. Question 54Fundamentals of Generative AI

    A company wants the model to answer using a strict JSON schema so downstream code can parse it. Which technique is most reliable?

    1. Describing the schema in the prompt and requesting only JSON output Correct answer.
    2. Asking the model to answer in prose and hoping it includes JSON
    3. Setting temperature to the maximum value
    4. Removing the schema from the prompt to save tokens

    Explanation

    Stating the exact schema and constraining the output to JSON greatly improves parseability, especially when combined with validation and retries. Structured output requirements make format errors detectable and correctable.

  55. Question 55Fundamentals of Generative AI

    Which scenario is the strongest candidate for fine-tuning rather than prompt engineering alone?

    1. A high-volume task needing a consistent proprietary tone across thousands of calls Correct answer.
    2. A one-off request to translate a short paragraph
    3. A quick prototype to test whether the model understands a question
    4. A task where the answer must cite retrieved documents

    Explanation

    Fine-tuning pays off when a stable, repeatable behavior is needed at scale and prompting cannot reliably encode it. High volume amortizes training cost, and consistent tone is exactly the kind of learned behavior fine-tuning captures.

  56. Question 56Fundamentals of Generative AI

    What is the effect of a larger context window on application design?

    1. It eliminates the need for retrieval in all cases
    2. It guarantees lower cost per request
    3. It removes the possibility of hallucination
    4. It allows more source text and conversation history to be supplied in a single request Correct answer.

    Explanation

    A context window bounds how many tokens a request may contain, so a larger window permits longer documents and histories. It does not guarantee cheaper calls, since cost scales with tokens processed, nor does it remove accuracy risks.

  57. Question 57Fundamentals of Generative AI

    Which statement correctly distinguishes zero-shot from few-shot prompting?

    1. Zero-shot supplies no task examples, while few-shot includes a small number of demonstrations Correct answer.
    2. Few-shot always uses a larger model than zero-shot
    3. Few-shot supplies no examples, while zero-shot includes demonstrations
    4. Zero-shot requires fine-tuning, while few-shot does not

    Explanation

    Zero-shot relies on instructions alone, whereas few-shot adds a handful of example input-output pairs in the prompt. Demonstrations help the model infer format and task expectations without changing model weights.

  58. Question 58Fundamentals of Generative AI

    A generative model produces fluent text but repeatedly ignores a required output length. Which change is most targeted?

    1. Switch to a model with a smaller vocabulary
    2. Remove the length instruction to avoid conflict
    3. Increase temperature so lengths vary more
    4. Add an explicit length constraint with an example and validate output length programmatically Correct answer.

    Explanation

    Explicit constraints plus a concrete example clarify the requirement, and programmatic validation catches violations. Automated checks let the application retry or truncate rather than relying on the model alone.

  59. Question 59Fundamentals of Generative AI

    Which pair correctly matches a generative AI risk with a mitigations strategy?

    1. Cost - increasing output length
    2. Hallucination - grounding responses with retrieved sources Correct answer.
    3. Bias - raising the temperature
    4. Latency - adding more few-shot examples

    Explanation

    Grounding with retrieved, verifiable sources gives the model evidence to condition on, which reduces unsupported claims. Other listed pairings either worsen the stated risk or address unrelated dimensions.

  60. Question 60Fundamentals of Generative AI

    What does top-p (nucleus) sampling control during text generation?

    1. The cumulative probability mass of tokens considered at each step Correct answer.
    2. The size of the training dataset
    3. The number of layers in the transformer
    4. The maximum number of requests per second

    Explanation

    Top-p restricts sampling to the smallest set of tokens whose cumulative probability reaches a threshold p. Lower p narrows choices for more focused output, while higher p allows broader selection.

  61. Question 61Fundamentals of Generative AI

    A team must choose between a large model and a small model for a simple extraction task at high volume. Which reasoning is soundest?

    1. Benchmark both on representative samples and weigh quality against latency and cost Correct answer.
    2. Choose based on the number of parameters alone
    3. Pick the smallest model without testing to save time
    4. Always pick the largest model because size guarantees accuracy

    Explanation

    Model selection is an empirical trade-off among quality, latency, and cost on the actual task. Testing representative samples reveals whether a smaller, cheaper model meets the quality bar for simple extraction.

  62. Question 62Fundamentals of Generative AI

    An analyst notices the model confidently cites a policy number that does not exist. Which combination best addresses this failure mode?

    1. Fine-tuning on unrelated marketing copy
    2. Higher temperature plus longer outputs
    3. Retrieval grounding plus a verification step that checks citations against sources Correct answer.
    4. Removing all context from the prompt

    Explanation

    Grounding supplies authoritative source text, and a verification step confirms that cited identifiers actually appear in those sources. Together they catch fabricated references that fluency alone would hide.

  63. Question 63Fundamentals of Generative AI

    A RAG pipeline returns poor answers even though the generator is strong. Retrieval returns many passages, but the correct passage is usually ranked far down the list. Which change addresses the root cause?

    1. Improve chunking and embedding quality, then rerank retrieved passages before generation Correct answer.
    2. Increase the generator temperature to explore more answers
    3. Reduce the number of retrieved passages to one without reranking
    4. Lengthen the system prompt with unrelated instructions

    Explanation

    When the right passage exists but ranks poorly, the problem lies in retrieval and ranking, not generation. Better chunking and embeddings improve candidate quality, and a reranker promotes the most relevant passages before they reach the model.

  64. Question 64Fundamentals of Generative AI

    A regulated workflow requires that every generated claim be traceable to an approved source and that unsupported claims be blocked. Which architecture satisfies this most directly?

    1. A larger model with a higher temperature for creativity
    2. Fine-tuning on approved documents with no citation requirement
    3. A prompt that asks the model to be accurate without providing sources
    4. RAG with mandatory citations plus a post-generation check that rejects claims lacking source support Correct answer.

    Explanation

    Traceability requires linking each claim to a specific approved source, which retrieval with citations provides. A post-generation verifier that blocks unsupported claims enforces the policy rather than merely encouraging it.

  65. Question 65Fundamentals of Generative AI

    A team wants to reduce inference cost for a summarization workload while preserving quality. Which combination is most defensible?

    1. Disable evaluation to save engineering time
    2. Cache repeated summaries, cap output length, and evaluate a smaller model on representative samples Correct answer.
    3. Send full documents plus all prior conversation on every call
    4. Increase output length and raise temperature for variety

    Explanation

    Cost reduction comes from fewer processed tokens and avoided duplicate work: caching, bounded outputs, and right-sizing the model. Each change must be validated against quality metrics so savings do not silently degrade results.

  66. Question 66Fundamentals of Generative AI

    Which statement best describes what a foundation model is?

    1. A model trained on broad data that can be adapted to many downstream tasks Correct answer.
    2. A rule-based system that follows hand-written decision trees
    3. A model trained only on a single company's proprietary documents
    4. A small model that only performs arithmetic operations

    Explanation

    Foundation models are trained on very large, broad datasets and learn general patterns that transfer to many tasks through prompting or adaptation. They are not limited to one organization's data, and they are learned rather than hand-coded rule systems.

  67. Question 67Fundamentals of Generative AI

    In generative AI, what does the term token refer to?

    1. A unit of text such as a word piece that the model processes and bills against Correct answer.
    2. A human reviewer who grades model output
    3. A security credential issued to API callers
    4. A hardware accelerator used during training

    Explanation

    Tokenization splits text into subword units that the model encodes as numbers, and both context limits and usage pricing are measured in these units. Tokens are a text representation concept, not hardware, credentials, or people.

  68. Question 68Fundamentals of Generative AI

    Which description matches how an autoregressive language model generates text?

    1. It generates the entire response in one step with no sequential dependency
    2. It retrieves a complete stored answer from a database of past conversations
    3. It predicts the next token repeatedly, feeding prior output back as input Correct answer.
    4. It selects the response by majority vote across many separate models

    Explanation

    Autoregressive generation produces one token at a time, conditioning each new token on the prompt plus everything already generated. This sequential feedback loop is what allows coherent long-form output.

  69. Question 69Fundamentals of Generative AI

    What is an embedding in the context of generative AI applications?

    1. A prompt template with placeholders for user input
    2. A numeric vector that represents text so that similar meanings are near each other Correct answer.
    3. A permission policy attached to an API key
    4. A compressed archive of the original training corpus

    Explanation

    Embeddings map text into a high-dimensional vector space where semantic similarity corresponds to geometric closeness. This property underlies semantic search and retrieval-augmented generation.

  70. Question 70Fundamentals of Generative AI

    Which factor most directly determines the maximum amount of conversation history a model can consider at once?

    1. The number of API keys provisioned
    2. The storage class of the log bucket
    3. The size of the model's context window Correct answer.
    4. The geographic region of the endpoint

    Explanation

    The context window is the fixed budget of tokens the model can attend to in a single request, covering instructions, retrieved content, and prior turns. Exceeding it forces truncation or summarization regardless of other infrastructure choices.

  71. Question 71Fundamentals of Generative AI

    Which outcome is a typical benefit of using a foundation model instead of building a task-specific model from scratch?

    1. Guaranteed factual accuracy without any evaluation
    2. Complete elimination of inference costs
    3. Automatic compliance with every industry regulation
    4. Faster time to a working prototype with far less labeled data Correct answer.

    Explanation

    Because foundation models already encode broad knowledge, teams can reach a useful prototype with prompting and small amounts of task data. They still require evaluation, cost management, and compliance work.

  72. Question 72Fundamentals of Generative AI

    Which term describes a model generating a confident but fabricated statement?

    1. Tokenization
    2. Hallucination Correct answer.
    3. Quantization
    4. Distillation

    Explanation

    Hallucination is fluent output that is not grounded in fact or in the provided source material. It is a core risk that grounding and evaluation techniques are designed to reduce.

  73. Question 73Fundamentals of Generative AI

    A team wants the model to summarize each incoming email in a fixed three-bullet format for a downstream report. Which approach most directly shapes the output structure?

    1. Removing the system message entirely
    2. Raising the temperature to its maximum
    3. Specifying the required format and an example in the prompt Correct answer.
    4. Increasing the model's parameter count

    Explanation

    Output structure is controlled through explicit instructions and demonstrations in the prompt, which tell the model exactly what shape to produce. Model size and sampling settings influence quality and variety but do not encode a required layout.

  74. Question 74Fundamentals of Generative AI

    A developer needs the model to reason through a multi-step arithmetic word problem before giving a final number. Which prompting technique is most suitable?

    1. Shortening the prompt to a single keyword
    2. Requesting only the final number with no explanation
    3. Asking the model to show intermediate reasoning steps before the answer Correct answer.
    4. Setting top-k to one to force determinism

    Explanation

    Eliciting intermediate reasoning steps gives the model room to decompose the problem, which generally improves accuracy on multi-step tasks. Suppressing the reasoning or over-constraining sampling removes that benefit.

  75. Question 75Fundamentals of Generative AI

    A team wants the model to adopt a consistent persona and follow safety rules across every turn of a chat session. Where should these persistent instructions be placed?

    1. In the final user turn only
    2. In the system message that precedes the conversation Correct answer.
    3. In a comment inside the application source code
    4. In the HTTP response headers

    Explanation

    System-level instructions are applied across the session and take precedence over individual turns, making them the right place for persona and policy constraints. Placing them in a single user turn makes them easy to lose or override later.

  76. Question 76Fundamentals of Generative AI

    Which sampling change makes generated text more deterministic and repeatable?

    1. Increasing the maximum output length
    2. Raising top-k to include more candidates
    3. Lowering the temperature toward zero Correct answer.
    4. Adding more random seed variation per request

    Explanation

    Temperature scales the randomness of the next-token distribution, so values near zero concentrate probability on the most likely token and produce stable output. Widening candidate pools or varying seeds does the opposite.

  77. Question 77Fundamentals of Generative AI

    A product team must choose between a hosted API model and a self-hosted open-weight model. Which consideration most strongly favors the self-hosted option?

    1. A need for the largest possible context window
    2. A requirement that model weights and inference stay inside the organization's own network Correct answer.
    3. A desire to avoid writing any evaluation prompts
    4. A preference for paying strictly per token

    Explanation

    Self-hosting gives the organization direct control over where weights and inference run, which matters when data cannot leave a controlled environment. The other listed goals are either unrelated to hosting choice or actually favor hosted APIs.

  78. Question 78Fundamentals of Generative AI

    What is the primary role of a vector database in a retrieval-augmented generation system?

    1. Fine-tuning the generator on each new document
    2. Enforcing rate limits on API calls
    3. Storing embeddings and returning passages whose vectors are closest to the query Correct answer.
    4. Rendering the final answer to the end user

    Explanation

    A vector store indexes embeddings so that a query embedding can be matched to semantically similar chunks at low latency. It supplies context to the generator but does not train it or handle presentation and throttling.

  79. Question 79Fundamentals of Generative AI

    A RAG application answers questions about a 400-page manual. Which chunking choice is generally most effective?

    1. Chunk randomly without regard to sentence boundaries
    2. Split into overlapping passages aligned to logical sections and keep metadata Correct answer.
    3. Index the entire manual as one single chunk
    4. Split at every individual character

    Explanation

    Section-aligned chunks with modest overlap keep related ideas together while preserving enough context at the edges, and metadata enables filtering. Extremely large or extremely fragmented chunks both hurt retrieval precision.

  80. Question 80Fundamentals of Generative AI

    Which statement best describes the role of an agent in a generative AI system?

    1. It permanently retrains the base model after every user request
    2. It guarantees that tool outputs are factually correct
    3. It replaces the need for any prompt engineering
    4. It uses a model to decide which tools or actions to invoke toward a goal Correct answer.

    Explanation

    Agents combine a model with tool access and a control loop so the model can plan and call functions such as search or calculators. They orchestrate existing capabilities rather than retraining weights or guaranteeing correctness of external results.

  81. Question 81Fundamentals of Generative AI

    A company must show that a generated marketing claim was supported by an approved document. Which design element most directly provides this evidence?

    1. Using a larger batch size during evaluation
    2. Shortening the prompt to reduce token usage
    3. Returning citations that link each claim to the retrieved source passage Correct answer.
    4. Increasing the model's temperature for creative variety

    Explanation

    Citations create an auditable link between generated statements and the source passages the model was given. This traceability is what reviewers need to verify claims, whereas sampling or batching choices do not produce evidence.

  82. Question 82Fundamentals of Generative AI

    A team needs to compare two prompt variants fairly across 300 test cases. Which practice makes the comparison most trustworthy?

    1. Change the model and the prompt at the same time
    2. Run both variants on identical inputs with fixed decoding settings and score with defined criteria Correct answer.
    3. Run each variant on different inputs chosen by each reviewer
    4. Evaluate only the three examples that look best

    Explanation

    Controlled comparison requires holding inputs and decoding settings constant so the prompt is the only variable, plus explicit scoring criteria applied consistently. Changing multiple factors or cherry-picking examples confounds the result.

  83. Question 83Fundamentals of Generative AI

    Which situation is the clearest sign that prompt engineering alone is insufficient and parameter-efficient fine-tuning should be considered?

    1. The team wants to try a different sampling temperature
    2. The model occasionally needs a shorter answer
    3. The model must reliably imitate a specialized house style across thousands of examples that cannot fit in a prompt Correct answer.
    4. The prompt uses more than one paragraph of instructions

    Explanation

    When desired behavior depends on a large, consistent pattern of style or format that cannot be expressed compactly in a prompt, adapting weights through fine-tuning is the appropriate lever. Smaller adjustments such as length or sampling are usually handled by prompt changes.

  84. Question 84Fundamentals of Generative AI

    A generative feature must be blocked from producing disallowed content while still answering legitimate questions. Which design is most appropriate?

    1. Combine input filtering, output classification, and a documented escalation path for edge cases Correct answer.
    2. Route every request to a human reviewer before any response is shown
    3. Rely solely on the model's pretraining to refuse harmful requests
    4. Disable logging so blocked attempts leave no trace

    Explanation

    Layered controls catch different failure modes: input filters stop obvious abuse, output classifiers catch problematic generations, and escalation handles ambiguous cases. Relying on a single mechanism leaves gaps, while full manual review defeats the purpose of automation.

  85. Question 85Fundamentals of Generative AI

    A team must produce verifiable numeric answers from a corpus of financial filings. Which architecture best supports both accuracy and auditability?

    1. Ask the model to answer from memory and trust the highest-confidence output
    2. Raise temperature so the model explores more possible numbers
    3. Remove retrieved context to keep prompts short and cheap
    4. Retrieve relevant passages, require the model to cite them, and validate extracted figures with a deterministic tool Correct answer.

    Explanation

    Grounding in retrieved passages plus citations makes answers checkable, and delegating arithmetic to deterministic code avoids the model's unreliable mental computation. Memory-only answers and higher randomness both increase the risk of fabricated figures.

  86. Question 86Fundamentals of Generative AI

    A RAG assistant must answer questions about data the organization is legally barred from sending to a third-party provider. Which approach satisfies the constraint while preserving grounded answers?

    1. Ask the provider to disable logging and proceed with the public endpoint
    2. Encrypt the documents before sending them to the public API in plaintext-readable form
    3. Send the data to the public API but add a prompt instruction not to retain it
    4. Host the embedding model and generator inside the organization's controlled environment with retrieval over internal indexes Correct answer.

    Explanation

    If data may not leave the controlled environment, the embeddings, retrieval store, and generation must all run inside that boundary. Prompt-level requests and provider settings do not change where the data is actually processed.

  87. Question 87Fundamentals of Generative AI

    A team notices that adding many retrieved passages improves recall but degrades answer quality. Which adjustment best addresses the root cause?

    1. Raise the temperature to help the model cope with noise
    2. Increase the number of passages retrieved even further
    3. Rerank and trim retrieved passages so only the most relevant fit the context budget Correct answer.
    4. Remove the instructions that tell the model to use the sources

    Explanation

    Excess or low-relevance context crowds the window and distracts the generator, so reranking and trimming restores signal density. Adding more passages or randomness worsens the distraction, and removing grounding instructions discards the intended behavior.

  88. Question 88Fundamentals of Generative AI

    A company wants to adapt a foundation model to its own labeled support tickets so the model becomes better at classifying ticket categories, without retraining the entire model. Which approach fits this goal?

    1. Increasing the temperature setting for each request
    2. Adding more few-shot examples to every prompt
    3. Expanding the context window to hold more tickets
    4. Fine-tuning the model on the labeled ticket data Correct answer.

    Explanation

    Fine-tuning continues training an existing foundation model on domain-specific labeled examples, which adjusts its weights so it performs better on that task. It is the standard way to specialize a model when prompt-level guidance is not enough, and parameter-efficient variants can do this while updating only a small portion of the weights.

  89. Question 89Applications of Foundation Models

    A team wants a foundation model to answer questions using its own internal product manuals without retraining the model. Which approach should they use?

    1. Expanding the model's context window to hold all manuals at once
    2. Fine-tuning the model on the full text of every manual
    3. Increasing the model's temperature setting
    4. Retrieval Augmented Generation with a vector store of the manuals Correct answer.

    Explanation

    Retrieval Augmented Generation retrieves relevant passages from an external knowledge source at query time and supplies them to the model as context. This grounds answers in current proprietary content without changing model weights. Fine-tuning teaches behavior and style rather than reliably storing large factual corpora.

  90. Question 90Applications of Foundation Models

    Which statement best describes what fine-tuning a foundation model changes?

    1. The network latency between the client and the inference endpoint
    2. The model's weights, adapting its behavior to a specific task or style Correct answer.
    3. The number of tokens the model can accept in a prompt
    4. The pricing tier of the underlying compute instance

    Explanation

    Fine-tuning continues training on task-specific examples so the model's parameters are updated to better fit a target behavior, format, or domain. It does not change context length, latency, or instance pricing. Those are properties of configuration and infrastructure.

  91. Question 91Applications of Foundation Models

    A developer needs the model to consistently return a JSON object with fixed keys for downstream parsing. Which technique is most appropriate?

    1. Prompt engineering with a schema example and output constraints Correct answer.
    2. Raising the temperature to encourage creativity
    3. Switching to a smaller model with fewer parameters
    4. Shortening the system prompt to a single sentence

    Explanation

    Structured output is best achieved by instructing the model with an explicit schema and constraints, optionally combined with supported structured-output features. Clear instructions and examples reduce format drift. Sampling temperature and model size do not guarantee schema compliance.

  92. Question 92Applications of Foundation Models

    Which metric is most directly used to evaluate a retrieval system inside a RAG pipeline?

    1. Model training loss on the pretraining corpus
    2. Number of parameters in the embedding model
    3. Context relevance of retrieved passages to the query Correct answer.
    4. Cost per thousand input tokens

    Explanation

    RAG evaluation separates retrieval quality from generation quality. Retrieval is judged by whether the passages returned are relevant and sufficient for the query. Training loss and parameter counts describe the model, not retrieval accuracy.

  93. Question 93Applications of Foundation Models

    What is the primary role of embeddings in a semantic search application?

    1. To translate documents into multiple languages
    2. To encrypt sensitive content before indexing
    3. To compress documents so they use less storage
    4. To represent text as vectors so similarity can be computed numerically Correct answer.

    Explanation

    Embeddings map text into a high-dimensional vector space where semantically similar content is close together. Search then ranks items by vector similarity such as cosine distance. They are not a compression, translation, or encryption mechanism.

  94. Question 94Applications of Foundation Models

    Which Amazon Bedrock capability lets an application call its own business logic or external APIs during a model conversation?

    1. Bedrock Guardrails
    2. Agents for Amazon Bedrock with action groups Correct answer.
    3. Provisioned throughput
    4. Bedrock model evaluation jobs

    Explanation

    Agents orchestrate a model with action groups that invoke APIs or Lambda functions to fulfill user requests. Guardrails filter content, evaluation jobs measure quality, and provisioned throughput reserves capacity. Only agents provide tool invocation for business logic.

  95. Question 95Applications of Foundation Models

    A company must keep prompts and responses from leaving its AWS account boundary while using a managed foundation model. Which Bedrock feature supports this?

    1. Private connectivity and data-handling settings that keep data within the account's configured boundary Correct answer.
    2. Sending prompts to a public third-party chatbot for comparison
    3. Storing prompt logs in a public S3 bucket
    4. Disabling encryption on the model endpoint

    Explanation

    Bedrock supports private network connectivity and documented data-handling practices so customer content is not used to train base models and stays within the customer's control. Public sharing or unencrypted storage undermines that boundary. Security controls must be enabled, not disabled.

  96. Question 96Applications of Foundation Models

    Which prompt component is intended to set persistent behavior and constraints for a model across a conversation?

    1. The stop sequence parameter
    2. A single user follow-up question
    3. The top-p sampling value
    4. The system prompt Correct answer.

    Explanation

    The system prompt carries durable instructions such as role, tone, and safety constraints that apply across turns. User turns carry the immediate request. Stop sequences and top-p are decoding parameters, not behavioral instructions.

  97. Question 97Applications of Foundation Models

    A support assistant must cite the exact source document for each answer. Which design most directly enables verifiable citations?

    1. Increase the maximum output tokens
    2. Use a higher temperature for varied phrasing
    3. Return retrieved chunk identifiers and document metadata alongside the generated answer Correct answer.
    4. Remove the retrieval step to reduce latency

    Explanation

    Citations require traceability from the answer back to the retrieved source. Passing chunk IDs and metadata through the pipeline lets the application display verifiable references. Longer outputs, randomness, or removing retrieval provide no traceability.

  98. Question 98Applications of Foundation Models

    A team fine-tunes a model on 200 examples and observes it memorizing training answers while performing poorly on new inputs. Which action best addresses this?

    1. Reduce the evaluation dataset to the training examples
    2. Add more diverse training examples and evaluate on a held-out set Correct answer.
    3. Increase the number of training epochs substantially
    4. Deploy the model immediately since training loss is low

    Explanation

    Memorization with weak generalization signals overfitting to a small dataset. Expanding and diversifying examples plus honest held-out evaluation improves generalization. More epochs on the same small set worsens overfitting.

  99. Question 99Applications of Foundation Models

    An application sends long documents that exceed the model's context window. What is the most appropriate first step?

    1. Chunk the documents and retrieve only the relevant sections per query Correct answer.
    2. Convert the documents to images
    3. Remove the system prompt to save tokens
    4. Increase the temperature so the model focuses

    Explanation

    When input exceeds the context window, splitting content into chunks and retrieving only relevant pieces keeps each request within limits while preserving needed information. Temperature and prompt removal do not reduce input size meaningfully and can harm quality.

  100. Question 100Applications of Foundation Models

    Which situation is the strongest candidate for using a smaller, cheaper model instead of a large frontier model?

    1. High-volume classification of short support tickets into fixed categories Correct answer.
    2. Complex multi-step legal reasoning over long contracts
    3. Creative long-form narrative generation
    4. Open-ended research synthesis across many documents

    Explanation

    Narrow, high-volume tasks with a fixed label set often reach acceptable accuracy with smaller models at much lower cost and latency. Complex reasoning, synthesis, and creative generation benefit more from larger model capability.

  101. Question 101Applications of Foundation Models

    A RAG system returns irrelevant passages even though the correct document exists in the knowledge base. Which change most directly improves retrieval quality?

    1. Switch the output format to XML
    2. Shorten the final answer length
    3. Tune chunk size and embedding model, then re-rank results Correct answer.
    4. Raise the generation temperature

    Explanation

    Retrieval quality depends on how content is chunked, how well the embedding model captures meaning, and how candidates are ranked. Adjusting chunking, embeddings, and adding a re-ranker targets the retrieval stage directly. Generation settings do not affect which passages are found.

  102. Question 102Applications of Foundation Models

    What does a guardrail configured on a foundation model application primarily do?

    1. Automatically fine-tune the model on user data
    2. Filter or block content that violates defined safety and policy rules Correct answer.
    3. Guarantee deterministic model outputs
    4. Reduce the model's parameter count

    Explanation

    Guardrails evaluate inputs and outputs against configured policies such as harmful content, denied topics, or sensitive information filters. They do not make outputs deterministic, shrink models, or trigger training. They add a control layer around inference.

  103. Question 103Applications of Foundation Models

    A developer wants the model to call a weather API only when the user asks about current conditions. Which pattern describes this behavior?

    1. Model distillation into a smaller student model
    2. Tool use, where the model selects a function and supplies arguments Correct answer.
    3. Static few-shot prompting with no external calls
    4. Batch inference over a stored dataset

    Explanation

    Tool use lets the model decide when an external function is needed, choose it, and produce structured arguments for invocation. Batch inference, distillation, and static prompting do not provide conditional runtime API calls.

  104. Question 104Applications of Foundation Models

    Which practice best reduces hallucination in a question-answering assistant?

    1. Remove all reference documents to avoid bias
    2. Increase top-k sampling to widen the candidate pool
    3. Ask the model to answer as confidently as possible
    4. Instruct the model to answer only from provided context and to say when it does not know Correct answer.

    Explanation

    Grounding answers in supplied context and permitting an explicit abstention reduces fabricated content. Confidence framing, wider sampling, and removing references all increase the chance of unsupported claims.

  105. Question 105Applications of Foundation Models

    A team must compare two foundation models on their own labeled dataset before choosing one. Which approach fits?

    1. Choose the model with the lowest token price
    2. Rely on the models' parameter counts
    3. Run a Bedrock model evaluation job using the custom dataset and task metrics Correct answer.
    4. Use the model with the longest context window

    Explanation

    Custom evaluation runs candidate models against your own data and computes task-appropriate metrics, giving evidence for selection. Parameter count, context length, and price are useful constraints but do not measure accuracy on your task.

  106. Question 106Applications of Foundation Models

    A pipeline needs to process 500,000 product descriptions overnight without interactive latency requirements. Which approach is most cost-effective?

    1. Submit a batch inference job over the stored records Correct answer.
    2. Use provisioned throughput sized for interactive traffic
    3. Call the model synchronously from a user-facing web form
    4. Send requests one at a time with retries during peak hours

    Explanation

    Batch inference is designed for large offline workloads and typically costs less than on-demand interactive calls. Interactive patterns add latency management and capacity cost that this workload does not need.

  107. Question 107Applications of Foundation Models

    Which factor most directly increases the cost of an on-demand foundation model request?

    1. The number of input and output tokens processed Correct answer.
    2. The region's availability zone count
    3. The number of IAM roles in the account
    4. The size of the local development machine

    Explanation

    On-demand pricing for foundation models is based primarily on tokens consumed for input and generated output. IAM roles, availability zone counts, and developer hardware are unrelated to per-request model billing.

  108. Question 108Applications of Foundation Models

    A team wants deterministic, repeatable outputs for a classification task where the same input must always map to the same label. Which setting combination is most appropriate?

    1. High top-p with no label constraints
    2. Random stop sequences per request
    3. High temperature and open-ended generation
    4. Low temperature and constrained output labels Correct answer.

    Explanation

    Low temperature narrows sampling toward the most likely tokens, and constraining outputs to a fixed label set removes ambiguity. High randomness and unconstrained formats work against repeatability.

  109. Question 109Applications of Foundation Models

    An agent must complete a multi-step refund workflow that requires checking order status, verifying policy, and issuing a refund. Which design element is essential?

    1. A single one-shot prompt with no tools
    2. An orchestration plan that sequences tool calls and handles intermediate results Correct answer.
    3. A higher temperature to explore options
    4. A longer maximum output length

    Explanation

    Multi-step workflows require orchestration that decides the order of actions, passes results between steps, and handles failures. A one-shot prompt without tools cannot perform the required actions, and sampling or length settings do not supply orchestration.

  110. Question 110Applications of Foundation Models

    A RAG system answers correctly when the right chunk is retrieved but produces wrong answers when retrieval returns near-duplicate chunks that crowd out the best passage. Which remedy targets this failure?

    1. Apply diversity-aware selection or re-ranking before building the prompt context Correct answer.
    2. Increase the generation temperature
    3. Reduce the number of documents in the knowledge base
    4. Shorten the system prompt

    Explanation

    Near-duplicate chunks consume limited context slots and push out more useful evidence. Diversity-aware selection or a re-ranker ensures the most informative passages reach the prompt. Generation settings and corpus reduction do not address redundancy in retrieved context.

  111. Question 111Applications of Foundation Models

    A regulated workload requires that the model never respond on a specific restricted topic, and the control must apply consistently to both prompts and completions. Which design satisfies this?

    1. Filter only the final response with a client-side keyword list
    2. Rely on the base model's default behavior without configuration
    3. Add the restriction to a single user message in one conversation
    4. Configure guardrails with a denied-topic policy evaluated on input and output Correct answer.

    Explanation

    A guardrail policy applied to both input and output enforces the restriction at every turn regardless of prompt phrasing. A one-off message, default behavior, or output-only filtering leaves gaps, especially for paraphrased inputs.

  112. Question 112Applications of Foundation Models

    A team compares two approaches for a domain-specific assistant: fine-tuning versus RAG. The knowledge changes weekly and answers must reflect the latest policy. Which reasoning is soundest?

    1. Favor RAG because it eliminates the need for any evaluation
    2. Favor fine-tuning because weights update automatically each week
    3. Favor RAG because updating the index reflects new policy without retraining weights Correct answer.
    4. Favor fine-tuning because it always produces fewer hallucinations

    Explanation

    When facts change frequently, updating a retrieval index propagates new knowledge quickly, whereas fine-tuning requires repeated training cycles. Neither approach eliminates evaluation, and fine-tuning does not update itself or guarantee fewer hallucinations.

  113. Question 113Applications of Foundation Models

    An evaluation shows a model's answers are fluent but frequently unsupported by the retrieved context. Which pair of metrics best diagnoses this?

    1. Parameter count and context window size
    2. Cost per request and throughput
    3. Token count and latency
    4. Faithfulness to context and answer correctness Correct answer.

    Explanation

    Faithfulness measures whether claims are grounded in the supplied context, and correctness measures agreement with the expected answer. Together they separate fluent but unsupported output from genuinely accurate output. Operational metrics do not describe grounding quality.

  114. Question 114Applications of Foundation Models

    Which Amazon Bedrock feature provides a single API to send prompts to multiple foundation model providers without writing provider-specific integration code?

    1. Knowledge Bases
    2. The Converse API Correct answer.
    3. Provisioned Throughput
    4. Model evaluation jobs

    Explanation

    The Converse API offers one consistent request format across supported models, so switching providers requires minimal code change. Provisioned Throughput and Knowledge Bases serve capacity and retrieval purposes rather than unified invocation.

  115. Question 115Applications of Foundation Models

    A developer wants the model to return output that strictly follows a JSON schema so downstream code can parse it without cleanup. Which Amazon Bedrock capability addresses this?

    1. Raising the maximum token count
    2. Increasing the temperature setting
    3. Structured output enforcement through tool use or schema-constrained prompting Correct answer.
    4. Enabling streaming responses

    Explanation

    Constraining the model with a declared schema, often via tool definitions, forces output to match the expected structure. Temperature and token limits affect variability and length, not structural conformance.

  116. Question 116Applications of Foundation Models

    Which Amazon Bedrock feature stores document embeddings and retrieves relevant passages to ground a model's answer in enterprise content?

    1. Agents
    2. Guardrails
    3. Knowledge Bases Correct answer.
    4. Model customization

    Explanation

    Knowledge Bases handles ingestion, chunking, embedding, and vector retrieval so responses can be grounded in private documents. Guardrails filters content, customization adapts weights, and Agents orchestrates actions.

  117. Question 117Applications of Foundation Models

    What is the primary purpose of a system prompt in a foundation model application?

    1. To store conversation history across sessions
    2. To define the assistant's role, tone, and operating constraints Correct answer.
    3. To select which AWS Region handles the request
    4. To reduce the per-token price of inference

    Explanation

    The system prompt establishes persistent instructions such as persona, boundaries, and response style that apply to the whole interaction. Pricing, persistence, and routing are handled by infrastructure and application code.

  118. Question 118Applications of Foundation Models

    A team wants a Bedrock model to always produce nearly identical answers for the same factual question. Which inference parameter should they lower?

    1. Maximum output tokens
    2. Stop sequences
    3. Temperature Correct answer.
    4. Top-p

    Explanation

    Temperature scales sampling randomness, so a low value makes the model favor the highest-probability tokens and yields more consistent output. Token limits and stop sequences control length and termination instead.

  119. Question 119Applications of Foundation Models

    Which Amazon Bedrock concept allows a model to invoke a function defined by the developer so the application can supply live data mid-conversation?

    1. Model distillation
    2. Batch inference
    3. Provisioned Throughput
    4. Tool use Correct answer.

    Explanation

    Tool use lets the model emit a structured request for a developer-defined function, and the application returns the result for the model to continue. Batch inference, throughput reservation, and distillation concern processing mode and model economics.

  120. Question 120Applications of Foundation Models

    A company wants to route simple FAQ questions to a cheap model and complex reasoning requests to a stronger model. Which design pattern does this describe?

    1. Prompt caching
    2. Guardrail filtering
    3. Fine-tuning
    4. Model routing Correct answer.

    Explanation

    Model routing classifies or scores each request and dispatches it to the model whose cost and capability best fit it. Caching reuses prior results, fine-tuning changes behavior, and guardrails filter content.

  121. Question 121Applications of Foundation Models

    Which statement best describes an Amazon Bedrock guardrail's effect on model input and output?

    1. It retrains the model to remove unsafe knowledge.
    2. It guarantees factual accuracy of every response.
    3. It evaluates content against configured policies and can block or mask disallowed material. Correct answer.
    4. It encrypts prompts at rest using customer-managed keys.

    Explanation

    Guardrails apply policy checks such as denied topics, content filters, and sensitive-information redaction to requests and responses. They do not alter model weights, manage encryption, or verify factual correctness.

  122. Question 122Applications of Foundation Models

    A team needs to evaluate whether a RAG assistant's answers are grounded in retrieved passages. Which metric pair is most appropriate?

    1. Token count and cost
    2. Model size and parameter count
    3. Latency and throughput
    4. Faithfulness and answer relevance Correct answer.

    Explanation

    Faithfulness measures whether claims are supported by the retrieved context, and answer relevance measures whether the response addresses the question. Latency, cost, and size are operational attributes rather than grounding quality.

  123. Question 123Applications of Foundation Models

    An application must summarize 2,000 archived support tickets each night with no user waiting for results. Which Amazon Bedrock approach is most cost-effective?

    1. Streaming responses to an interactive UI
    2. Batch inference jobs Correct answer.
    3. Provisioned Throughput with a large model
    4. Synchronous on-demand calls in a tight loop

    Explanation

    Batch inference submits large volumes of records for asynchronous processing at lower cost, which suits overnight workloads without latency requirements. Synchronous loops, reserved capacity, and streaming are designed for interactive or steady-traffic patterns.

  124. Question 124Applications of Foundation Models

    A developer notices that a model ignores instructions placed in the middle of a very long prompt but follows them when repeated at the end. Which technique addresses this?

    1. Increasing the maximum token limit
    2. Lowering the temperature
    3. Restating key instructions near the end of the prompt Correct answer.
    4. Switching to streaming output

    Explanation

    Models can under-weight information in the middle of long contexts, so placing critical instructions at the start and repeating them near the end improves adherence. Temperature, token limits, and streaming do not change positional attention.

  125. Question 125Applications of Foundation Models

    A company wants a model to answer in its own brand voice and terminology, and has thousands of high-quality example responses. Which approach most directly teaches that style?

    1. Fine-tuning on the curated example responses Correct answer.
    2. Raising top-p to 1.0
    3. Adding a glossary to every request
    4. Enabling a content filter

    Explanation

    Fine-tuning adjusts model weights using demonstration data, which is effective for consistently reproducing a specific style and vocabulary. Prompt glossaries are limited by context length, and sampling or filtering settings do not teach style.

  126. Question 126Applications of Foundation Models

    An agent workflow must pause and ask a human to approve a high-value refund before continuing. Which design element enables this?

    1. A larger context window
    2. A higher temperature for the planning model
    3. A human-in-the-loop confirmation step between agent actions Correct answer.
    4. Provisioned Throughput

    Explanation

    Human-in-the-loop checkpoints let the orchestration layer halt before a sensitive action and resume only after approval. Sampling settings, context size, and capacity reservation do not introduce approval gates.

  127. Question 127Applications of Foundation Models

    A team wants to reduce latency and cost for a chatbot whose system prompt is long and identical on every request. Which Bedrock feature helps most?

    1. Disabling guardrails
    2. Prompt caching of the shared prefix Correct answer.
    3. Increasing the output token limit
    4. Using a larger model

    Explanation

    Prompt caching stores the processed shared prefix so repeated requests avoid reprocessing it, cutting both latency and input cost. Larger models and higher output limits increase expense, and removing guardrails does not address prompt reuse.

  128. Question 128Applications of Foundation Models

    Which chunking strategy is most likely to improve retrieval for documents with clearly separated sections such as policies with numbered clauses?

    1. Splitting every 50 characters regardless of structure
    2. Chunking along section boundaries with some overlap Correct answer.
    3. Embedding the entire document as one chunk
    4. Removing all headings before embedding

    Explanation

    Respecting natural section boundaries keeps each chunk semantically coherent, and modest overlap preserves context across splits. Fixed tiny splits fragment meaning, whole-document chunks dilute relevance, and stripping headings discards useful signals.

  129. Question 129Applications of Foundation Models

    A RAG pipeline retrieves passages that are semantically similar but from an outdated document version. Which change best addresses this?

    1. Switching to a larger embedding dimension
    2. Raising the model temperature
    3. Increasing the number of retrieved passages
    4. Adding metadata filters on document version and date Correct answer.

    Explanation

    Metadata filtering restricts retrieval to current documents so stale versions cannot be returned. Retrieving more passages, changing temperature, or altering embedding size does not exclude outdated sources.

  130. Question 130Applications of Foundation Models

    An application must run several foundation model calls in parallel and aggregate the results before responding. Which implementation choice fits best?

    1. Caching one model's answer for all requests
    2. Sequential blocking calls in a single thread
    3. Asynchronous concurrent invocation of the model endpoints Correct answer.
    4. A single call with a much larger token limit

    Explanation

    Independent model calls can be issued concurrently so total latency approaches the slowest call rather than the sum. Sequential calls add latency, larger token limits do not parallelize work, and reusing one answer sacrifices correctness.

  131. Question 131Applications of Foundation Models

    A team wants to confirm that a fine-tuned model still performs well on general tasks after domain adaptation. Which evaluation practice is most appropriate?

    1. Testing only on the fine-tuning examples
    2. Measuring inference latency only
    3. Comparing token counts of responses
    4. Holding out a domain test set plus a general benchmark set Correct answer.

    Explanation

    Evaluating on unseen domain data shows whether specialization generalizes, and a general benchmark detects regressions in broader capability. Training data, latency, and token counts do not reveal quality trade-offs.

  132. Question 132Applications of Foundation Models

    A developer wants the model to answer only from provided context and to reply that it does not know otherwise. Which prompt design best achieves this?

    1. Increasing the temperature to encourage exploration
    2. Requiring answers to cite supplied passages and to decline when evidence is absent Correct answer.
    3. Instructing the model to be helpful and creative
    4. Adding many few-shot examples unrelated to the task

    Explanation

    Explicit grounding instructions with a fallback behavior constrain the model to the supplied evidence and reduce fabrication. Creative framing, irrelevant examples, and higher randomness all encourage unsupported output.

  133. Question 133Applications of Foundation Models

    A company needs to compare two Bedrock models on its own labeled dataset with automatic scoring of accuracy and toxicity. Which Bedrock capability supports this directly?

    1. Provisioned Throughput
    2. Guardrail versions
    3. Model evaluation jobs Correct answer.
    4. Knowledge Bases sync

    Explanation

    Model evaluation jobs run automated or human review of models against a supplied dataset and chosen metrics, enabling side-by-side comparison. Throughput, knowledge base sync, and guardrail versions serve other purposes.

  134. Question 134Applications of Foundation Models

    An agent must choose among several tools and may need multiple rounds of reasoning before answering. Which capability is essential in the underlying model?

    1. Support for image input
    2. A very large maximum output length
    3. Reliable multi-step reasoning and tool-selection ability Correct answer.
    4. Very high throughput

    Explanation

    Agents depend on the model to plan, select the right tool, and interpret results across turns, so reasoning and tool-selection quality matter most. Throughput, output length, and multimodality are useful but not essential to the reasoning loop.

  135. Question 135Applications of Foundation Models

    A RAG assistant gives correct answers for common questions but fails on rare entity names that the embedding model represents poorly. Which improvement addresses the root cause?

    1. Hybrid retrieval combining keyword search with vector search Correct answer.
    2. Lowering the generation temperature
    3. Reducing the number of retrieved chunks to one
    4. Shortening the system prompt

    Explanation

    Hybrid retrieval adds lexical matching so exact rare terms can be found even when embeddings are weak, then merges results with vector search. Temperature and prompt length affect generation, and retrieving fewer chunks removes evidence.

  136. Question 136Applications of Foundation Models

    A team finds that adding more retrieved passages steadily improves recall but degrades answer quality because irrelevant passages distract the model. Which approach best balances these effects?

    1. Increase temperature to diversify answers
    2. Remove retrieval and rely on model knowledge
    3. Retrieve as many passages as the context window allows
    4. Rerank candidates and keep only the highest-scoring passages Correct answer.

    Explanation

    Retrieving a broad candidate set and then reranking keeps recall high while passing only the most relevant passages to the model. Filling the context with marginal passages, dropping retrieval, or raising randomness all reduce answer quality.

  137. Question 137Applications of Foundation Models

    A regulated assistant must refuse requests on a defined restricted subject even when users rephrase the request indirectly, and the control must apply uniformly across several models. Which design meets this?

    1. A guardrail with a denied-topic policy attached to every model invocation Correct answer.
    2. A post-processing keyword filter on final answers only
    3. A system prompt on each model asking it to avoid the subject
    4. A fine-tuned model per provider

    Explanation

    A guardrail denied-topic policy evaluates inputs and outputs against a defined subject and can be applied consistently across models through a shared configuration. Prompt-only instructions are unreliable, output-only filters miss indirect requests, and per-provider tuning is inconsistent.

  138. Question 138Applications of Foundation Models

    A team must choose between fine-tuning and RAG for a domain assistant whose underlying facts change weekly and whose answers must cite sources. Which choice is best justified?

    1. Fine-tuning, because it eliminates the need for retrieval infrastructure
    2. Neither, because citation is impossible with foundation models
    3. RAG, because the index can be refreshed frequently and passages can be cited Correct answer.
    4. Fine-tuning, because weights encode facts permanently

    Explanation

    RAG separates knowledge from model weights, so updating the index propagates new facts without retraining and retrieved passages supply citations. Fine-tuning bakes in facts that go stale and does not naturally provide source attribution.

  139. Question 139Applications of Foundation Models

    Which Amazon Bedrock feature lets a foundation model call external APIs or developer-defined functions during a conversation so the application can return live data to the model?

    1. Provisioned throughput
    2. Tool use (function calling) Correct answer.
    3. Model evaluation jobs
    4. Knowledge bases

    Explanation

    Tool use, also called function calling, lets the model emit a structured request for a named tool with arguments; the application executes that tool and returns the result so the model can continue reasoning with fresh data. Provisioned throughput only reserves capacity, evaluation jobs score quality, and knowledge bases handle retrieval from indexed documents.

  140. Question 140Applications of Foundation Models

    A developer is writing a prompt for a Bedrock model and wants the model to return a response that can be parsed directly by application code without additional natural-language cleanup. Which prompting technique is most appropriate?

    1. Asking the model to respond in a friendly, conversational tone
    2. Providing several paragraphs of background context before the instruction
    3. Instructing the model to output JSON that conforms to a specified schema Correct answer.
    4. Increasing the maximum token count for the response

    Explanation

    Constraining the output format to a documented schema such as JSON makes the response machine-parseable and reduces the need for post-processing. Tone, extra context, and a larger token limit affect style or length but do not guarantee a structured, code-friendly result.

  141. Question 141Applications of Foundation Models

    A team is building an agent on Amazon Bedrock that must decide which of several tools to call and may need several reasoning steps before answering. Which capability is essential for this design?

    1. A single-shot prompt with no intermediate reasoning
    2. A larger context window only
    3. A fixed decision tree hard-coded by the developer
    4. Multi-step reasoning with tool selection and observation of results Correct answer.

    Explanation

    Agents that must choose among tools and iterate require a reasoning loop in which the model plans, selects a tool, observes the result, and decides the next step. A single-shot prompt or a fixed decision tree cannot adapt tool choice to intermediate observations, and a larger context window alone does not add decision-making.

  142. Question 142Applications of Foundation Models

    A RAG application retrieves relevant passages, but answers sometimes contradict the retrieved text because the model relies on its pretrained knowledge. Which prompt change most directly improves grounding?

    1. Increase the temperature so the model explores more phrasing
    2. Add more few-shot examples of fluent writing style
    3. Shorten the system prompt to reduce token usage
    4. Instruct the model to answer only from the provided passages and to say it does not know when the passages are insufficient Correct answer.

    Explanation

    Grounding improves when the prompt explicitly restricts the answer to the supplied context and defines a fallback for missing information, which discourages the model from filling gaps with parametric knowledge. Style examples, higher temperature, and a shorter prompt do not constrain the model to the retrieved evidence.

  143. Question 143Applications of Foundation Models

    A company wants a Bedrock model to follow a specific response template with fixed section headings for every customer reply. Which approach best enforces the template consistently?

    1. Increase the model's temperature to encourage creativity
    2. Describe the required headings and order explicitly in the system prompt and validate the output Correct answer.
    3. Remove all formatting instructions to avoid confusing the model
    4. Rely on the model to infer the template from a single example

    Explanation

    Explicit instructions that name each required section and its order, combined with output validation, give the model a clear contract and let the application catch deviations. Inferring from one example is unreliable, higher temperature adds variance, and removing formatting guidance leaves the structure unspecified.

  144. Question 144Applications of Foundation Models

    A team must decide between a single large prompt containing all retrieved documents and a two-stage design that first extracts only the sentences relevant to the question and then answers from those sentences. The documents are long and only a small fraction is relevant. Which consideration best supports the two-stage design?

    1. It reduces irrelevant context and token cost while focusing the model on salient evidence Correct answer.
    2. It guarantees the model will never hallucinate
    3. It removes the need to evaluate answer quality
    4. It eliminates the need for any retrieval step

    Explanation

    Extracting relevant sentences before answering shrinks the context, lowers token cost, and reduces distraction from irrelevant passages, which often improves answer focus. It does not replace retrieval, guarantee freedom from hallucination, or remove the need for evaluation.

  145. Question 145Guidelines for Responsible AI

    Which statement best describes the purpose of AWS's responsible AI approach?

    1. It focuses only on maximizing model accuracy on benchmarks.
    2. It replaces human oversight with automated decision-making.
    3. It restricts AI use to a single AWS service.
    4. It aims to build AI systems that are fair, explainable, and governed across their lifecycle. Correct answer.

    Explanation

    Responsible AI at AWS centers on designing, deploying, and operating AI systems with fairness, explainability, privacy, safety, and governance considered throughout the lifecycle. It is broader than raw accuracy and does not eliminate human oversight.

  146. Question 146Guidelines for Responsible AI

    A team wants to detect statistical bias in a trained model before deployment. Which Amazon SageMaker capability is designed for this?

    1. SageMaker Debugger profiling
    2. SageMaker Feature Store
    3. SageMaker Clarify bias detection Correct answer.
    4. SageMaker Model Monitor data capture

    Explanation

    SageMaker Clarify provides tools to detect potential bias in data and models and to explain model predictions. It is the purpose-built component for bias and explainability analysis in SageMaker.

  147. Question 147Guidelines for Responsible AI

    What is the main purpose of a model card in responsible AI practice?

    1. To encrypt model artifacts at rest.
    2. To store the training dataset in its original raw form.
    3. To document a model's intended uses, limitations, and performance characteristics. Correct answer.
    4. To schedule automatic retraining of the model.

    Explanation

    Model cards are transparency documents that summarize what a model is for, how it was evaluated, and where it may fail. They help downstream users judge whether a model is appropriate for their context.

  148. Question 148Guidelines for Responsible AI

    Which practice best supports human oversight of a generative AI assistant used in a customer service workflow?

    1. Routing low-confidence responses to a human agent for review Correct answer.
    2. Disabling content filters to reduce latency
    3. Removing logging to protect customer privacy
    4. Allowing the assistant to issue refunds without review

    Explanation

    Human oversight means people retain meaningful control over consequential outputs. Escalating uncertain or high-impact responses to a human keeps a person in the loop where errors matter most.

  149. Question 149Guidelines for Responsible AI

    Which AWS service feature helps filter harmful or inappropriate content in generative AI applications?

    1. Amazon Guardrails for Amazon Bedrock Correct answer.
    2. Amazon CloudWatch
    3. AWS Identity and Access Management
    4. AWS CloudTrail

    Explanation

    Guardrails for Amazon Bedrock lets teams define policies that filter harmful content and restrict topics in generative AI applications. The other services provide logging, monitoring, and access control rather than content safety policies.

  150. Question 150Guidelines for Responsible AI

    What does the principle of explainability require of an AI system used in a lending decision?

    1. That stakeholders can understand why a particular decision was reached Correct answer.
    2. That the model always achieves perfect accuracy
    3. That the model's source code is published publicly
    4. That no training data is stored anywhere

    Explanation

    Explainability means the factors driving a decision can be understood and communicated to affected parties and reviewers. It does not demand perfect accuracy, open sourcing, or the absence of stored data.

  151. Question 151Guidelines for Responsible AI

    Which action reflects the responsible AI principle of privacy and data protection when building a model?

    1. Minimizing and de-identifying personal data used for training Correct answer.
    2. Sharing raw datasets with all team members by default
    3. Storing credentials inside notebook code
    4. Training on all available customer records regardless of consent

    Explanation

    Privacy-aware AI development limits collection to what is needed, removes or masks personal identifiers, and respects consent. These practices reduce the risk of exposing sensitive information through models or workflows.

  152. Question 152Guidelines for Responsible AI

    Which statement about fairness in machine learning is accurate?

    1. Fairness can be assessed using multiple metrics, and the appropriate metric depends on context. Correct answer.
    2. Fairness is guaranteed once bias is removed from the training labels.
    3. A model is fair if it has high accuracy on the test set.
    4. Fairness only matters for models that process images.

    Explanation

    Fairness has several formal definitions that can conflict, so teams choose metrics based on the use case and affected groups. Removing label bias is helpful but does not by itself guarantee fair outcomes.

  153. Question 153Guidelines for Responsible AI

    A bank must explain declined loan applications to applicants. Which combination best supports this requirement?

    1. Storing only aggregate accuracy metrics
    2. A deep neural network with no interpretation tooling
    3. Batch retraining every hour without documentation
    4. SageMaker Clarify feature attribution plus human review of explanations Correct answer.

    Explanation

    Feature attribution from SageMaker Clarify identifies which inputs most influenced a prediction, and human review ensures the explanation is accurate and appropriate. Together they provide understandable, defensible reasons for a decision.

  154. Question 154Guidelines for Responsible AI

    Which risk does prompt injection primarily introduce in a generative AI application?

    1. It can cause the model to ignore its intended instructions and follow attacker-supplied ones. Correct answer.
    2. It automatically retrains the model on new data.
    3. It permanently deletes the model's training data.
    4. It reduces the storage cost of embeddings.

    Explanation

    Prompt injection occurs when untrusted input manipulates the model into disregarding its system instructions or performing unintended actions. Mitigations include input validation, least-privilege tool access, and output filtering.

  155. Question 155Guidelines for Responsible AI

    How does Amazon Bedrock support transparency about which foundation models are available and their characteristics?

    1. By preventing any logging of model invocations
    2. By requiring customers to reverse-engineer model behavior
    3. By hiding model providers from users
    4. By providing model information and documentation for available foundation models Correct answer.

    Explanation

    Amazon Bedrock exposes information about the foundation models it offers so customers can select models suited to their needs. Documentation of model characteristics supports informed, transparent model selection.

  156. Question 156Guidelines for Responsible AI

    A team deploys a model and later finds its accuracy drops for one demographic group after a policy change. Which practice most directly addresses this?

    1. Removing the affected group from reporting
    2. Increasing the endpoint instance size
    3. Monitoring model performance by subgroup and investigating drift Correct answer.
    4. Disabling monitoring to avoid alerts

    Explanation

    Subgroup monitoring reveals performance differences that aggregate metrics hide, and drift analysis helps explain why they emerged. This enables targeted remediation such as retraining or feature review.

  157. Question 157Guidelines for Responsible AI

    What is the role of an AI governance framework in an organization?

    1. It applies only to external vendors.
    2. It defines roles, policies, and review processes for managing AI risk. Correct answer.
    3. It replaces the need for any technical testing.
    4. It guarantees that no AI system will ever fail.

    Explanation

    Governance establishes accountability through defined roles, policies, risk reviews, and approval gates across the AI lifecycle. It complements technical controls rather than substituting for them.

  158. Question 158Guidelines for Responsible AI

    Which approach best reduces the risk of exposing sensitive information through a generative AI application's responses?

    1. Prompting the model to never reveal secrets
    2. Applying redaction and access controls to data before it reaches the model Correct answer.
    3. Increasing the model's temperature setting
    4. Using a larger context window

    Explanation

    Technical controls such as data redaction, least-privilege retrieval, and access boundaries prevent sensitive data from entering prompts or being returned. Relying on instructions alone is not a reliable safeguard.

  159. Question 159Guidelines for Responsible AI

    Which statement best characterizes the shared responsibility model as it applies to responsible AI?

    1. AWS secures the cloud infrastructure while customers govern their data, models, and use cases. Correct answer.
    2. Responsibility ends once a model is deployed.
    3. AWS is solely responsible for the fairness of customer models.
    4. Customers are solely responsible for the security of the underlying cloud infrastructure.

    Explanation

    Under the shared responsibility model, AWS secures the cloud while customers remain accountable for their data, configurations, model choices, and how outputs are used. Responsible AI obligations therefore continue after deployment.

  160. Question 160Guidelines for Responsible AI

    What does SageMaker Clarify's SHAP-based analysis help a team understand?

    1. How much the endpoint costs per hour
    2. How each input feature contributes to a model's prediction Correct answer.
    3. How many records are stored in the feature store
    4. Which IAM role invoked the endpoint

    Explanation

    SHAP-based feature attribution estimates the contribution of each input feature to individual predictions. This supports explainability and helps teams detect whether sensitive features drive outcomes.

  161. Question 161Guidelines for Responsible AI

    A company wants to restrict a generative AI assistant from discussing competitors. Which control is most appropriate?

    1. A larger foundation model
    2. A longer system prompt with no enforcement
    3. A denied topic policy in Guardrails for Amazon Bedrock Correct answer.
    4. A higher request rate limit

    Explanation

    Denied topic policies in Guardrails block responses that fall within defined restricted subjects, providing enforceable control rather than advisory guidance. This directly limits discussion of disallowed topics.

  162. Question 162Guidelines for Responsible AI

    Which practice best supports accountability when an AI system produces a harmful output?

    1. Maintaining audit logs and a documented incident review process Correct answer.
    2. Deleting invocation logs immediately
    3. Disabling user feedback channels
    4. Blaming the foundation model provider only

    Explanation

    Audit trails and a defined incident process allow an organization to reconstruct what happened, identify root causes, and demonstrate accountability. Logging and review are core governance controls.

  163. Question 163Guidelines for Responsible AI

    Which consideration is most important when deciding whether a high-stakes decision may be fully automated?

    1. The size of the training dataset
    2. The potential impact on individuals and the need for human review Correct answer.
    3. The number of available API endpoints
    4. The color scheme of the user interface

    Explanation

    High-stakes decisions affecting people's rights or opportunities call for meaningful human review and escalation paths. Impact on individuals, not dataset size or interface details, should drive the automation decision.

  164. Question 164Guidelines for Responsible AI

    A team finds that its model satisfies equal opportunity across groups but violates demographic parity. What does this illustrate?

    1. That different fairness definitions can conflict, requiring a justified choice Correct answer.
    2. That the model must be discarded
    3. That bias cannot be measured
    4. That fairness metrics are mutually compatible in all cases

    Explanation

    Fairness definitions encode different values and can be mathematically incompatible, so teams must select and justify metrics aligned with the use case and legal context. Satisfying one metric does not guarantee satisfying another.

  165. Question 165Guidelines for Responsible AI

    Which design choice best mitigates hallucination risk in a retrieval-augmented generative AI application?

    1. Removing the retrieval step to reduce latency
    2. Requiring the model to cite retrieved sources and declining to answer when evidence is insufficient Correct answer.
    3. Increasing temperature to encourage creativity
    4. Allowing the model to answer any question without constraints

    Explanation

    Grounding responses in retrieved evidence and instructing the model to abstain when support is weak reduces fabricated content. Citations also let users verify claims.

  166. Question 166Guidelines for Responsible AI

    An organization wants to document known limitations and evaluation results of a foundation model before approving it for a regulated use case. Which artifact and process best fit?

    1. A model card reviewed through a governance approval gate Correct answer.
    2. A marketing brochure reviewed by sales
    3. A network diagram reviewed by facilities
    4. A billing report reviewed by finance

    Explanation

    Model cards capture intended use, limitations, and evaluation results, and governance approval gates ensure risks are assessed before deployment. Together they provide documented, reviewable evidence for regulated contexts.

  167. Question 167Guidelines for Responsible AI

    A team must choose between two models with similar accuracy, but one shows large performance gaps across subgroups. Which decision best reflects responsible AI practice?

    1. Select the cheaper model without evaluating fairness
    2. Select the model with the highest overall accuracy without further analysis
    3. Select the model with smaller subgroup gaps and document the trade-off Correct answer.
    4. Select the newer model regardless of subgroup performance

    Explanation

    When overall accuracy is comparable, smaller disparities across subgroups indicate more equitable behavior and lower risk of harm. Documenting the trade-off supports accountability and informed approval.

  168. Question 168Guidelines for Responsible AI

    Which control most directly limits a generative AI agent from taking unintended actions through connected tools?

    1. Caching responses for longer periods
    2. Increasing the maximum token limit
    3. Granting the agent broad administrative permissions
    4. Applying least-privilege permissions and requiring confirmation for consequential actions Correct answer.

    Explanation

    Least-privilege permissions constrain what an agent can do, and confirmation steps keep humans in control of consequential operations. These controls reduce the blast radius of manipulated or erroneous agent behavior.

  169. Question 169Guidelines for Responsible AI

    Which statement best explains why continuous monitoring is a responsible AI requirement rather than a one-time task?

    1. Monitoring is only needed to reduce cloud costs.
    2. Model behavior and data distributions can change after deployment, creating new risks. Correct answer.
    3. Once validated, a model cannot produce harmful outputs.
    4. Regulations forbid evaluating deployed models.

    Explanation

    Data drift, changing user behavior, and evolving contexts can degrade performance or introduce bias after launch. Continuous monitoring detects these shifts so teams can remediate before harm occurs.

  170. Question 170Guidelines for Responsible AI

    Which statement best describes the purpose of a model card in responsible AI practice?

    1. It records the individual user prompts submitted to a model during production traffic.
    2. It lists the hourly price of each inference endpoint the model is deployed to.
    3. It stores the encrypted credentials a model uses to reach downstream data sources.
    4. It documents a model's intended use, training approach, evaluation results, and known limitations. Correct answer.

    Explanation

    A model card is a transparency artifact that summarizes what a model is for, how it was built and evaluated, and where it should not be used. Publishing this information helps downstream consumers judge whether the model fits their context and supports accountability for its use.

  171. Question 171Guidelines for Responsible AI

    A retail company plans to deploy a generative AI assistant that recommends products. Leadership wants assurance that the assistant's answers stay grounded in approved catalog content and that harmful or off-brand responses are caught before customers see them. Which combination of controls best addresses this goal?

    1. Granting the assistant broad write access to the catalog so it can correct outdated entries itself.
    2. Increasing the model's temperature so responses sound more natural, and reviewing transcripts weekly.
    3. Fine-tuning the model on marketing copy and disabling all content filters to avoid false positives.
    4. Retrieval from the approved catalog with guardrails that filter inputs and outputs, plus logging of blocked responses for review. Correct answer.

    Explanation

    Grounding responses in a curated knowledge source constrains the model to approved content, while input and output guardrails block harmful or off-brand material. Logging blocked interactions closes the loop by giving reviewers evidence to tune filters and prompts over time.

  172. Question 172Guidelines for Responsible AI

    A health insurer's model performs well overall, but an internal review shows it recommends additional manual case review far more often for one language group than for others, even when clinical inputs are equivalent. The team must decide what to do before launch. Which analysis best reflects responsible AI reasoning?

    1. Proceed because aggregate accuracy is high and the disparity affects only a small share of applicants.
    2. Remove the language attribute from the model's inputs and treat the disparity as resolved.
    3. Deploy the model but add a disclaimer stating that outputs are advisory and may vary by group.
    4. Investigate whether the disparity stems from label quality or proxy features, then mitigate and re-evaluate before launch. Correct answer.

    Explanation

    A subgroup disparity in a high-stakes setting signals a possible fairness defect that must be diagnosed at its source, whether through biased labels, proxy variables, or sampling. Mitigation followed by subgroup re-evaluation verifies that the fix actually narrows the gap rather than hiding it.

  173. Question 173Security, Compliance, and Governance for AI Solutions

    Which AWS service is designed to continuously evaluate your AI workloads against compliance frameworks and produce evidence reports?

    1. Amazon Inspector
    2. AWS Audit Manager Correct answer.
    3. Amazon Macie
    4. AWS Shield Advanced

    Explanation

    AWS Audit Manager automates evidence collection and maps it to control frameworks such as SOC 2 and ISO standards, which supports audit readiness for AI workloads. It gathers configuration and usage data from other AWS services rather than scanning data content itself.

  174. Question 174Security, Compliance, and Governance for AI Solutions

    A team wants to keep foundation model prompts and responses encrypted with keys they control and rotate independently of AWS. Which AWS KMS key type should they choose?

    1. A KMS alias
    2. A customer managed key Correct answer.
    3. An AWS managed key
    4. An AWS owned key

    Explanation

    Customer managed keys are created and governed by the customer, who controls the key policy, rotation schedule, and grants. AWS managed and AWS owned keys are controlled by AWS, so the customer cannot set a custom rotation schedule or key policy.

  175. Question 175Security, Compliance, and Governance for AI Solutions

    Which AWS service centrally manages permissions for users and provides temporary credentials that applications use to call Amazon Bedrock?

    1. AWS Resource Access Manager
    2. AWS IAM Identity Center Correct answer.
    3. AWS Directory Service
    4. Amazon Cognito user pools

    Explanation

    IAM Identity Center centralizes workforce identities and issues short-lived credentials tied to permission sets, which is the recommended pattern for human access to AI services. Long-term access keys stored in code are discouraged because they are harder to rotate and easier to leak.

  176. Question 176Security, Compliance, and Governance for AI Solutions

    Which AWS service should be used to discover and classify personally identifiable information stored in an S3 bucket that feeds a retrieval corpus?

    1. AWS Config
    2. Amazon GuardDuty
    3. Amazon Macie Correct answer.
    4. AWS CloudTrail

    Explanation

    Macie uses managed and custom data identifiers to detect sensitive data such as names, account numbers, and credentials inside S3 objects. This helps teams decide what to exclude from retrieval corpora before exposing it to a generative AI application.

  177. Question 177Security, Compliance, and Governance for AI Solutions

    What is the primary purpose of a guardrail configured on an Amazon Bedrock model?

    1. To automatically fine-tune the model on new data
    2. To filter harmful content and block disallowed topics Correct answer.
    3. To reduce inference latency for large prompts
    4. To replicate the model across multiple Regions

    Explanation

    Bedrock guardrails apply content filters, denied topics, word filters, and sensitive-information filters to both prompts and responses. They enforce safety policy at the application boundary without retraining the underlying model.

  178. Question 178Security, Compliance, and Governance for AI Solutions

    Which IAM feature lets you limit what an Amazon Bedrock agent can do by attaching a scoped permissions document to the agent's execution role?

    1. Service control policies
    2. Identity-based policies Correct answer.
    3. Permissions boundaries
    4. Session tags

    Explanation

    Identity-based policies attached to the execution role define which actions and resources the agent may use, following least privilege. The role's policy is the primary control over what the agent can call at runtime.

  179. Question 179Security, Compliance, and Governance for AI Solutions

    Which AWS service provides a managed catalog of machine learning models and datasets with governance features and can be used to share models across accounts?

    1. Amazon SageMaker Model Registry Correct answer.
    2. AWS Glue Data Catalog
    3. AWS Data Exchange
    4. Amazon Bedrock Model Evaluation

    Explanation

    SageMaker Model Registry tracks model versions, approval status, and metadata, and it supports cross-account model sharing with governance controls. This gives auditors a traceable record of which model version was approved for production.

  180. Question 180Security, Compliance, and Governance for AI Solutions

    Which practice best supports data residency requirements when using a generative AI service in multiple countries?

    1. Enabling verbose prompt logging in every Region
    2. Storing all data in a single global bucket
    3. Choosing Regional endpoints and cross-Region inference settings deliberately Correct answer.
    4. Relying on the model provider's default routing

    Explanation

    Data residency is controlled by where requests are processed and where data at rest is stored. Selecting Regional endpoints and reviewing cross-Region inference behavior keeps processing inside approved geographies.

  181. Question 181Security, Compliance, and Governance for AI Solutions

    A developer stores an Amazon Bedrock API key in a source code repository. Which control most directly prevents long-term credential exposure going forward?

    1. Replacing static keys with IAM role temporary credentials Correct answer.
    2. Enabling AWS CloudTrail data events
    3. Enabling S3 versioning on the repository bucket
    4. Increasing the KMS key rotation period

    Explanation

    Temporary credentials from IAM roles expire automatically and remove the need to embed long-lived secrets in code. This eliminates the class of exposure where a committed key remains valid indefinitely.

  182. Question 182Security, Compliance, and Governance for AI Solutions

    An organization needs to prove that only approved foundation models are invoked in a production account. Which combination gives the strongest ongoing assurance?

    1. AWS CloudTrail logging of model invocation events plus an AWS Config rule or automated alert on unapproved model IDs Correct answer.
    2. Amazon CloudWatch dashboards showing invocation counts per hour
    3. AWS Trusted Advisor checks for service limits
    4. Amazon S3 access logs on the model artifact bucket

    Explanation

    CloudTrail captures the API calls that identify which model was invoked, and an AWS Config rule or event-driven alert detects deviations from the approved list. Together they provide continuous, auditable enforcement evidence rather than a point-in-time view.

  183. Question 183Security, Compliance, and Governance for AI Solutions

    A company must ensure that no prompt containing customer account numbers is sent to a third-party model. Which Amazon Bedrock guardrail component addresses this requirement?

    1. A denied topics filter
    2. A word filter for profanity
    3. A contextual grounding check
    4. A sensitive information filter with a custom regex Correct answer.

    Explanation

    Sensitive information filters detect and block or mask patterns such as account numbers, including custom regular expressions. This stops regulated identifiers from leaving the application boundary.

  184. Question 184Security, Compliance, and Governance for AI Solutions

    Which AWS service should be enabled to detect unusual API activity, such as Amazon Bedrock calls from an unexpected geographic location or compromised credentials?

    1. Amazon Comprehend
    2. AWS Artifact
    3. AWS Audit Manager
    4. Amazon GuardDuty Correct answer.

    Explanation

    GuardDuty continuously analyzes account and network activity, including anomalous API behavior tied to compromised credentials or unusual locations. It surfaces findings that security teams can route to incident response workflows.

  185. Question 185Security, Compliance, and Governance for AI Solutions

    A team wants to log every request to a generative AI application with the user identity, prompt metadata, and response identifier for later investigation, while minimizing storage of sensitive prompt text. Which approach best meets this goal?

    1. Rely only on default service metrics in CloudWatch
    2. Store full prompts and responses indefinitely in the application database
    3. Emit structured audit records with identifiers and metadata to a centralized logging account, excluding raw prompt text Correct answer.
    4. Disable logging to reduce the risk of exposure

    Explanation

    Auditability requires linking each request to an identity and a traceable identifier, while data minimization limits exposure of sensitive content. Centralizing logs in a separate account also protects them from tampering in the workload account.

  186. Question 186Security, Compliance, and Governance for AI Solutions

    Which statement describes how service control policies contribute to governance of AI workloads across an AWS Organization?

    1. They replace the need for identity-based policies
    2. They grant permissions to individual users in member accounts
    3. They encrypt data at rest in member accounts
    4. They define the maximum permissions available to accounts in the organization Correct answer.

    Explanation

    Service control policies act as permission guardrails that cap what identities in member accounts can do, even if their identity policies allow more. They are a central lever for restricting services or Regions where AI workloads may run.

  187. Question 187Security, Compliance, and Governance for AI Solutions

    A regulated customer requires that their data never be used to train or improve the underlying foundation model. Which Bedrock capability should they verify and rely on?

    1. Model evaluation jobs
    2. Cross-Region inference profiles
    3. Data privacy settings where prompts and completions are not used to train base models Correct answer.
    4. Provisioned throughput commitments

    Explanation

    Bedrock does not use customer prompts or completions to train the base foundation models, and this should be confirmed against the service's data handling documentation. Documenting this control supports the customer's compliance evidence.

  188. Question 188Security, Compliance, and Governance for AI Solutions

    A security team wants to confirm that an Amazon Bedrock knowledge base can only retrieve documents the requesting user is authorized to see. Which design achieves this?

    1. Filtering retrieved chunks by user-specific metadata stored with each document Correct answer.
    2. Encrypting the vector store with an AWS managed key
    3. Increasing the number of retrieved chunks per query
    4. Enabling verbose retrieval logs for all users

    Explanation

    Document-level metadata combined with per-user filters ensures retrieval returns only chunks the user is entitled to see. This enforces row-level style access control inside the retrieval layer of a retrieval-augmented generation application.

  189. Question 189Security, Compliance, and Governance for AI Solutions

    Which AWS service provides self-service access to AWS compliance reports, certifications, and third-party attestations that an AI workload auditor may request?

    1. AWS CloudTrail Lake
    2. AWS Config
    3. AWS Artifact Correct answer.
    4. AWS Security Hub

    Explanation

    Artifact is the portal for downloading AWS compliance reports, certifications, and agreements such as the Business Associate Addendum. Auditors typically request these artifacts as evidence of the provider's controls.

  190. Question 190Security, Compliance, and Governance for AI Solutions

    A company wants a repeatable, automated check that every Amazon SageMaker endpoint has encryption at rest enabled and is tagged with a data-classification label. Which approach fits best?

    1. AWS Trusted Advisor cost checks
    2. Amazon Inspector network reachability findings
    3. Manual quarterly review of the console
    4. AWS Config managed and custom rules with automatic remediation Correct answer.

    Explanation

    AWS Config rules continuously evaluate resource configuration and can trigger remediation when a resource drifts from the desired state. This creates durable, auditable enforcement instead of periodic manual inspection.

  191. Question 191Security, Compliance, and Governance for AI Solutions

    Which control most directly limits the blast radius if an AI application's credentials are compromised and used to call Amazon Bedrock at scale?

    1. Using a larger instance type for the application
    2. Storing the credentials in an encrypted environment variable
    3. Attaching a permissions boundary and per-user quotas to the application role Correct answer.
    4. Enabling S3 server access logging

    Explanation

    A permissions boundary caps the maximum permissions of the role, and quotas limit how much can be consumed even within those permissions. Together they constrain both the scope and the volume of abuse after credential compromise.

  192. Question 192Security, Compliance, and Governance for AI Solutions

    An auditor asks how the organization ensures that a generative AI feature respects user consent for data processing in a specific jurisdiction. Which evidence is most relevant?

    1. Documented consent capture, purpose limitation, and retention configuration mapped to the legal basis Correct answer.
    2. The model's published benchmark scores
    3. The instance family used for inference
    4. The number of tokens processed per day

    Explanation

    Consent and purpose limitation are legal controls that must be evidenced through documented data flows, capture mechanisms, and retention settings. Benchmark scores and infrastructure metrics do not demonstrate lawful processing.

  193. Question 193Security, Compliance, and Governance for AI Solutions

    A team must prevent prompt injection attempts from causing an Amazon Bedrock agent to invoke a destructive action. Which layered control is most effective?

    1. Relying solely on the foundation model's instruction-following ability
    2. Shortening the system prompt
    3. Increasing the model's temperature setting
    4. Combining input filtering, least-privilege action groups, and human approval for high-impact actions Correct answer.

    Explanation

    Prompt injection cannot be fully solved by model behavior alone, so defenses must combine input screening with tightly scoped tools and approval gates for sensitive operations. Least privilege ensures a successful injection still cannot reach destructive APIs.

  194. Question 194Security, Compliance, and Governance for AI Solutions

    An organization wants to share a curated dataset with a partner account for model fine-tuning while keeping control over how it is used. Which AWS mechanism is purpose-built for governed cross-account data sharing?

    1. Emailing a presigned URL with no expiration
    2. Publishing the dataset to a public S3 website
    3. AWS Lake Formation cross-account grants with tag-based access control Correct answer.
    4. Copying the dataset into the partner's S3 bucket

    Explanation

    Lake Formation supports fine-grained, cross-account grants that can be conditioned on LF-Tags, so the producer retains control over columns and rows the partner may access. This avoids duplicating data into an account where governance is lost.

  195. Question 195Security, Compliance, and Governance for AI Solutions

    A compliance officer needs to demonstrate that model outputs in a regulated workflow were checked for grounding against approved source documents. Which Amazon Bedrock feature produces this evidence?

    1. Batch inference
    2. Model distillation
    3. Provisioned throughput
    4. Contextual grounding checks in guardrails Correct answer.

    Explanation

    Contextual grounding checks score whether a response is supported by the supplied source and flag potentially ungrounded claims. Logging those scores creates evidence that outputs were validated against approved references.

  196. Question 196Security, Compliance, and Governance for AI Solutions

    A multinational firm must ensure inference for EU customers is processed only within the EU, but also needs resilience if one Region fails. Which architecture satisfies both requirements?

    1. A single endpoint in one EU Region with no failover
    2. A global endpoint that may route to any Region
    3. Multiple endpoints across several EU Regions with routing restricted to those Regions Correct answer.
    4. Cross-Region inference to the lowest-latency Region worldwide

    Explanation

    Deploying across multiple Regions inside the approved geography provides failover while keeping processing within the jurisdiction. Routing must be explicitly constrained, because global or latency-based routing can leave the approved boundary.

  197. Question 197Security, Compliance, and Governance for AI Solutions

    Which combination best provides defense in depth for sensitive data used in a retrieval-augmented generation pipeline?

    1. KMS encryption at rest, TLS in transit, least-privilege retrieval roles, and PII redaction before indexing Correct answer.
    2. A single shared administrator role for all pipeline components
    3. Public read access restricted by an obscure bucket name
    4. Bucket policies alone

    Explanation

    Defense in depth layers encryption at rest and in transit with authorization scoping and data minimization before indexing. Each layer independently reduces the impact of a failure in another layer.

  198. Question 198Security, Compliance, and Governance for AI Solutions

    A company wants a durable, auditable record of who invoked which Amazon Bedrock model, when they did it, and from where, so that security analysts can investigate incidents months later. Which AWS service is designed to record this API activity?

    1. AWS CloudTrail Correct answer.
    2. AWS Config
    3. AWS Trusted Advisor
    4. Amazon CloudWatch

    Explanation

    CloudTrail records AWS API activity as events, capturing the identity of the caller, the time, the source IP address, and the request parameters, and it can deliver those events to an Amazon S3 bucket for long-term retention. Because model invocation calls are API actions, CloudTrail provides the durable audit trail an investigator needs after the fact.

  199. Question 199Security, Compliance, and Governance for AI Solutions

    A bank runs a generative AI assistant on Amazon Bedrock. Policy requires that every model response be screened for harmful content before it reaches the customer, and that blocked responses be counted so the security team can alert on spikes. Which approach meets both requirements?

    1. Enable Amazon Bedrock Guardrails on the model invocation and publish guardrail metrics to CloudWatch for alarming Correct answer.
    2. Rely on the foundation model provider's built-in moderation and trust that blocked content never appears
    3. Restrict access to the application with IAM policies so only approved users can submit prompts
    4. Add a system prompt instructing the model to refuse harmful requests and review the prompt quarterly

    Explanation

    Guardrails apply configurable content filters and denied-topic policies to both prompts and responses at invocation time, independent of the model's own behavior. Guardrail activity emits metrics that can be sent to CloudWatch, which supplies the counting and alarming the security team needs.

  200. Question 200Security, Compliance, and Governance for AI Solutions

    A healthcare company must show auditors that its Amazon SageMaker domain has continuously disallowed public internet access for the past year, including periods when engineers changed network settings. Which evidence best demonstrates this?

    1. Configuration history and compliance timeline in AWS Config for the SageMaker domain resource Correct answer.
    2. CloudWatch dashboards showing inference latency over the past twelve months
    3. An inventory export from AWS Resource Explorer listing SageMaker resources in the account
    4. A screenshot of the current SageMaker domain settings taken on the audit date

    Explanation

    AWS Config records resource configuration changes over time and evaluates them against rules, so its configuration history and compliance timeline can show whether a setting held continuously and when it changed. That point-in-time history is what supports an assertion about the entire audit period rather than a single moment.

AWS AI Practitioner flashcards

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

Artificial intelligence
The broad discipline of building systems that perform tasks requiring human-like intelligence, such as reasoning, perception, and language understanding. Machine learning is one approach within it.
Machine learning
A method where algorithms learn patterns and decision rules from data rather than following hand-coded logic, then generalize those patterns to new, unseen inputs.
Supervised learning
Training on labeled examples where each input has a known target. Classification predicts discrete categories such as spam or not spam; regression predicts continuous numeric values.
Unsupervised learning
Training on unlabeled data to discover inherent structure, such as grouping similar records through clustering or reducing dimensions while preserving important variance.
Overfitting
A model fits training data noise so closely that it performs well on training data but poorly on new data, showing low training error and high validation error.
Validation versus test set
The validation set guides hyperparameter tuning during development; the test set is held out and used once at the end to estimate real-world generalization.
Amazon SageMaker
AWS managed platform covering the ML lifecycle: data labeling, notebook experimentation, distributed training, hyperparameter tuning, model registry, and endpoint deployment.
Batch transform
A SageMaker inference mode for scoring large datasets offline in a single job, suited to non-interactive workloads where per-request latency is not a concern.
Foundation model
A very large model pretrained on broad, general data that can be adapted through prompting or fine-tuning to many downstream tasks without training from scratch.
Transformer self-attention
A mechanism that lets each token weigh its relevance to every other token in the sequence in parallel, capturing long-range context and enabling scalable training.
Token and context window
A token is a subword text unit the model processes; the context window is the maximum combined tokens of input and output allowed in one request.
Temperature
A sampling parameter controlling output randomness. Lower values make responses more deterministic and focused; higher values increase diversity and creative variation.
Top-p sampling
Nucleus sampling that restricts token selection to the smallest set whose cumulative probability reaches p, balancing coherence against variety in generated text.
Embeddings
Dense numeric vectors representing text or images so that semantically similar items sit close together, measured by cosine similarity and used for search.
Hallucination
Model output that is fluent and confident but factually unsupported or fabricated. Retrieval grounding, citations, and guardrails reduce it but cannot fully eliminate it.
Retrieval-augmented generation
A pattern that retrieves relevant documents from a knowledge store and inserts them into the prompt so the model answers from supplied evidence rather than memory alone.
Few-shot prompting
Including several worked input-output examples in the prompt so the model infers the desired task format and style without any weight updates.
Chain-of-thought prompting
Asking the model to reason through intermediate steps before producing a final answer, which improves accuracy on multi-step arithmetic and logic tasks.
Fine-tuning versus RAG
Fine-tuning adjusts model weights to change style or behavior; retrieval-augmented generation supplies external, updatable facts at inference without retraining the model.
Amazon Bedrock Knowledge Bases
A managed RAG capability that chunks and embeds documents into a vector store, then retrieves relevant passages to ground model responses.
Bedrock Agents
A managed feature that lets a foundation model plan multi-step tasks and call external APIs defined as action groups to fulfill user requests.
Provisioned throughput
A Bedrock purchasing option reserving dedicated model capacity for predictable, high-volume workloads, offering steadier latency than on-demand pay-per-token pricing.
Bedrock Guardrails
Configurable controls that filter harmful content, deny specific topics, mask personally identifiable information, and ground responses to reduce unsafe or unsupported output.
SageMaker Clarify
A tool that detects potential bias in datasets and trained models and generates feature-attribution explanations showing which inputs drove predictions.
SageMaker Model Monitor
A production monitoring capability that detects data drift, model quality drift, and bias drift by comparing live inference data against a captured baseline.
Human-in-the-loop review
A workflow routing low-confidence or high-stakes AI outputs to people for verification, using services such as Amazon Augmented AI before actions are finalized.
Model card
A transparency document describing a model's intended use, training data, performance metrics, and known limitations so downstream users can assess suitability.
Shared responsibility model
AWS secures infrastructure and managed service internals; customers remain responsible for their data, identity permissions, encryption configuration, and application code.
Least privilege for AI workloads
Granting each model or application role only the specific actions and resources it needs, such as scoped Bedrock invoke permissions, instead of broad wildcard access.
VPC endpoint for Bedrock
A PrivateLink interface endpoint that keeps model invocation traffic inside the AWS network, avoiding exposure over the public internet.
Amazon Macie
A managed data security service that uses pattern and machine learning detection to discover and classify sensitive data such as credentials in S3.
Prompt injection
An attack where crafted input causes a model to ignore its instructions or leak data; mitigations include input validation, output sanitization, and least-privilege tool access.

AWS AI Practitioner study guide

Fundamentals of AI and ML

Core vocabulary

AI is the broad goal of machines performing tasks that normally need human intelligence; machine learning (ML) is the technique of fitting models to data instead of hand-coding rules; deep learning is ML using multi-layer neural networks. Distinguish supervised learning (labelled data, e.g. classification and regression), unsupervised learning (unlabelled data, e.g. clustering and dimensionality reduction), and reinforcement learning (an agent learns from reward signals through trial and error). Know that training, validation, and test splits exist to detect overfitting (good on training data, poor on new data) versus underfitting (too simple to capture the pattern).

AWS service map

Amazon SageMaker covers the ML lifecycle: data preparation, training, tuning, deployment, and monitoring. SageMaker Ground Truth handles data labelling; SageMaker Data Wrangler handles feature preparation; SageMaker Model Monitor detects drift. Amazon Rekognition (images/video), Comprehend (text/NLP), Transcribe (speech to text), Polly (text to speech), Translate, and Forecast are pre-trained AI services that need no model training. Amazon Bedrock is the managed route to foundation models.

Traps to expect

Questions blur the line between AI, ML, and deep learning, or between a pre-trained AI service and a service that requires you to train a model. Another trap is confusing overfitting with a data-quality problem, or assuming more data always fixes a model. Watch for "fully managed, no ML expertise required" phrasing, which points to an AI service rather than SageMaker.

Rule of thumb

Ask: does the task need a custom model trained on my own labelled data? If yes, choose SageMaker. If a standard capability (vision, speech, translation, forecasting) already solves it, choose the pre-built AI service.

Worked example

A team wants to tag support emails by sentiment with no ML staff. Sentiment is a standard NLP capability, so Amazon Comprehend is the answer, not SageMaker training.

Fundamentals of Generative AI

How foundation models work

A foundation model (FM) is trained on broad data at scale and adapted to many downstream tasks. Large language models are FMs trained on text to predict the next token. Key terms: tokens (text units the model processes), embeddings (numeric vectors capturing meaning), context window (how much input the model can consider at once), parameters (learned weights), and inference (generating output). Training stages include pre-training on vast corpora, then fine-tuning or instruction tuning, then alignment techniques such as RLHF.

Customization options

Prompt engineering changes the input only. Retrieval-augmented generation (RAG) supplies external documents at inference time so answers stay grounded and current without retraining. Fine-tuning adjusts model weights with labelled examples for style or domain behaviour. Continued pre-training adds domain data. Know the trade-off: RAG is fast, cheap, and good for changing facts; fine-tuning is costlier and better for consistent behaviour or format.

Traps to expect

Questions confuse RAG with fine-tuning, or imply that fine-tuning updates knowledge in real time. Another trap is treating a larger context window as unlimited memory, or assuming generative models are factually reliable. Watch for hallucination (confident but fabricated output) and the idea that temperature and top-p control randomness, not accuracy.

Rule of thumb

If the answer depends on facts that change or live in your own documents, use RAG. If it depends on tone, format, or a specialized task pattern, consider fine-tuning. If it is just wording, use prompt engineering first.

Worked example

A company wants an assistant to answer from its internal policy PDFs that change monthly. Fine-tuning would go stale; RAG retrieves current documents at query time, so RAG is correct.

Applications of Foundation Models

Choosing and using models

Amazon Bedrock provides managed access to foundation models from multiple providers through one API, with serverless scaling and no infrastructure management. Candidates should know Bedrock features: model choice, the Converse API for a consistent interface, Knowledge Bases for managed RAG, Agents for multi-step tasks that call APIs, Guardrails for filtering input and output, and provisioned throughput for predictable capacity. Bedrock supports text, image, and embedding models.

Application patterns

Match the pattern to the need. Summarization, classification, extraction, and generation are common text tasks. RAG grounds answers in a knowledge base. Agents orchestrate tools and actions. Embeddings power semantic search and recommendation. Multimodal models handle images plus text. Also know Amazon Q for business assistants and PartyRock for no-code experimentation, and that SageMaker hosts custom or open models you manage yourself.

Traps to expect

Questions conflate Bedrock with SageMaker: Bedrock is managed FM access via API; SageMaker is the full build-and-train platform. Another trap is assuming Bedrock trains models from scratch or that Knowledge Bases fine-tune the model. Watch for confusion between Guardrails (safety filtering) and Agents (task execution), and between on-demand and provisioned throughput.

Rule of thumb

Start with the simplest pattern: a single prompt for one-off text tasks, RAG when answers need your documents, an agent when the model must take actions, and fine-tuning only when prompting and RAG fall short.

Worked example

A team needs a chatbot that answers from a product manual and refuses off-topic requests. Use Bedrock Knowledge Bases for retrieval and Guardrails to block disallowed content, rather than training a new model.

Guidelines for Responsible AI

Core principles

Responsible AI in AWS guidance centers on fairness, explainability, privacy, safety, controllability, verifiability, and governance. Fairness means avoiding bias against groups; explainability means understanding why a model produced an output; privacy means protecting personal data; safety means avoiding harmful output; controllability means keeping humans able to intervene; verifiability means being able to audit results; governance means documented oversight. AWS tools include SageMaker Clarify for bias detection and explainability, SageMaker Model Monitor for drift, Bedrock Guardrails for filtering harmful or sensitive content, and Amazon A2I for human review.

Human oversight and transparency

Human-in-the-loop review is appropriate for high-stakes decisions. Model cards and data sheets document intended use, limitations, and training data. Transparency means disclosing that AI is in use and how outputs should be interpreted. Bias can enter through training data, feature selection, or labelling, so detection should happen before and after deployment.

Traps to expect

Questions treat responsible AI as a one-time compliance step rather than a lifecycle practice. Another trap is confusing fairness with accuracy, or assuming a model is unbiased because it performs well overall. Watch for answers that remove humans entirely from high-stakes workflows, or that claim Guardrails alone guarantees compliance.

Rule of thumb

For any high-impact decision, ask three questions: can we explain the output, is there human review, and are we monitoring for bias and drift after launch? If any answer is no, the design is incomplete.

Worked example

A hiring model screens resumes. Use Clarify to measure bias across groups, keep a human reviewer for rejections, document limitations in a model card, and monitor post-deployment for drift.

Security, Compliance, and Governance for AI Solutions

Shared responsibility and access control

AWS operates on a shared responsibility model: AWS secures the cloud, customers secure what they put in it, including data, prompts, and model configuration. Use IAM for least-privilege access, roles rather than long-lived keys, and separate permissions for training, deployment, and data access. Encrypt data at rest with AWS KMS and in transit with TLS. Amazon VPC, PrivateLink, and VPC endpoints keep traffic private. AWS CloudTrail logs API activity; Amazon CloudWatch monitors metrics and alarms; AWS Config tracks resource compliance.

Data protection and governance

Know that Amazon Bedrock does not use customer prompts or outputs to train base models, and that data can be kept private within your account. SageMaker supports network isolation and private endpoints for training and inference. For governance, define data classification, retention, and lineage; use AWS Audit Manager and Security Hub for evidence and posture; apply Bedrock Guardrails to block sensitive data such as personally identifiable information from leaving the application.

Traps to expect

Questions misstate the shared responsibility boundary, assuming AWS secures customer data or prompts. Another trap is treating encryption alone as full compliance, or confusing CloudTrail (API audit) with CloudWatch (performance monitoring). Watch for answers that grant broad permissions or store secrets in code.

Rule of thumb

Apply least privilege, encrypt everywhere, log everything, and keep data private to your account. If a question asks who secures the model inputs, the answer is the customer.

Worked example

A regulated team deploys a Bedrock chatbot. They scope IAM roles per function, enable CloudTrail, encrypt with KMS, add Guardrails to block PII, and confirm prompts are not used for model training.

AWS AI Practitioner cheat sheet

Quick reference

Fundamentals of AI and ML

  • AI is the broad field of machines performing tasks associated with human intelligence; ML is the subset where models learn patterns from data instead of being explicitly programmed.

  • Supervised learning trains on labeled data (classification predicts discrete classes, regression predicts continuous values); unsupervised learning finds structure in unlabeled data (clustering, dimensionality reduction).

  • Reinforcement learning trains an agent by reward and penalty over sequential actions; it needs an environment, a reward signal, and an exploration strategy.

  • Training data is split into train/validation/test sets; the validation set tunes hyperparameters and the test set estimates generalization and must not influence training.

  • Overfitting = low training error but high validation error (high variance); underfitting = high error on both (high bias). Regularization, more data, and simpler models reduce variance.

  • Common metrics: accuracy (balanced classes), precision/recall/F1 (imbalanced classes), RMSE/MAE (regression), AUC-ROC (ranking quality).

  • AWS managed ML services: SageMaker for building, training, and deploying custom models; Bedrock for foundation models via API; Comprehend, Rekognition, Transcribe, and Forecast for prebuilt task-specific inference.

  • Feature engineering creates informative inputs; feature store centralizes reusable features; SageMaker Ground Truth produces labeled datasets including human-in-the-loop review.

  • Inference choices trade off latency and cost: real-time endpoints for low-latency requests, serverless inference for intermittent traffic, batch transform for large offline jobs, asynchronous inference for large payloads.

Quick reference

Fundamentals of Generative AI

  • Foundation models are large models pretrained on broad data that can be adapted to many downstream tasks; LLMs are foundation models specialized for text.

  • Transformer architecture uses self-attention to weigh relationships between tokens in parallel, enabling long-context modeling and scalable training.

  • Tokenization splits text into subword units; context window is the maximum token count (prompt plus output) the model can process in one request.

  • Training stages: pretraining (next-token prediction on large corpora), fine-tuning or instruction tuning (task/behavior adaptation), and alignment via RLHF using human preference rankings.

  • Inference sampling parameters: temperature controls randomness, top-p (nucleus) restricts to cumulative probability mass, top-k restricts to the k most likely tokens, max tokens caps output length.

  • Embeddings map text or images to dense vectors; semantic similarity is measured with cosine similarity and powers retrieval-augmented generation and vector search.

  • Hallucination is fluent but unsupported model output; grounding with retrieved documents, citations, and guardrails reduces it but does not eliminate it.

  • Amazon Bedrock provides serverless access to multiple foundation models through a unified API, with Knowledge Bases for RAG, Agents for tool use, and Guardrails for content filtering.

  • Amazon Q is a generative AI assistant family for business (Q Business) and developer (Q Developer) use cases, distinct from building custom applications on Bedrock.

Quick reference

Applications of Foundation Models

  • Prompt engineering patterns: zero-shot instruction, few-shot examples, chain-of-thought reasoning, and role/system prompts that fix persona and constraints.

  • Retrieval-augmented generation (RAG) pipeline: ingest and chunk documents, embed into a vector store, retrieve top-k passages by similarity, then pass them in the prompt for grounded answers.

  • Fine-tuning adjusts model weights for domain style or format; RAG adds external knowledge without retraining. Choose fine-tuning for behavior, RAG for fresh or proprietary facts.

  • Bedrock Knowledge Bases handles chunking, embedding, vector storage (e.g., OpenSearch Serverless, Aurora pgvector), and retrieval; Bedrock Agents orchestrate multi-step tasks by calling APIs via action groups.

  • Model selection criteria: task quality, context length, latency, cost per token, modality support (text, image, audio), and regional availability.

  • Provisioned throughput on Bedrock reserves model capacity for predictable high-volume workloads; on-demand pricing suits variable or low traffic.

  • Bedrock Flows chains prompts, models, and knowledge bases into a visual workflow; Bedrock Model Evaluation compares models on quality, toxicity, and robustness.

  • Common application patterns: summarization, classification/extraction, conversational assistants, code generation, semantic search, and content generation with human review.

Quick reference

Guidelines for Responsible AI

  • Core responsible AI dimensions: fairness, explainability, privacy, safety, controllability, verifiability, and governance.

  • Bias sources include unrepresentative training data, labeler bias, and proxy features; mitigation spans data balancing, fairness metrics, and post-processing adjustments.

  • Amazon Bedrock Guardrails can filter harmful content, deny topics, mask PII, block profanity, and ground responses in a source to reduce hallucinations.

  • Amazon SageMaker Clarify detects bias in data and models and produces feature-attribution explanations for transparency.

  • SageMaker Model Monitor detects data drift and model quality drift in production so degradation is caught before it affects users.

  • Human-in-the-loop review (A2I) is appropriate for high-stakes decisions; fully automated decisions require documented justification and appeal paths.

  • Transparency artifacts: model cards document intended use and limitations, data sheets document dataset provenance, and system cards describe architecture and risk controls.

  • Privacy-preserving techniques: PII detection and redaction, data minimization, encryption, and differential privacy for aggregate statistics.

Quick reference

Security, Compliance, and Governance for AI Solutions

  • Shared responsibility model: AWS secures the cloud (infrastructure, managed service internals); the customer secures data, IAM permissions, encryption choices, and application code.

  • IAM least privilege for AI workloads: scoped roles for Bedrock and SageMaker, separate execution roles, and no long-lived access keys in code.

  • Encryption: data at rest via AWS KMS customer-managed keys, data in transit via TLS; Bedrock and SageMaker support customer-managed KMS keys.

  • Network isolation: VPC endpoints (PrivateLink) keep traffic to Bedrock and SageMaker off the public internet; security groups and private subnets restrict access.

  • Data governance: Amazon Macie discovers and classifies sensitive data in S3; AWS Lake Formation and Glue Data Catalog control and catalog training data.

  • Auditability: AWS CloudTrail logs API activity, CloudWatch captures metrics and logs, and Bedrock model invocation logging records prompts and responses when enabled.

  • Compliance frameworks relevant to AI: GDPR, HIPAA, SOC 2, and ISO 27001; AWS Artifact supplies AWS compliance reports and agreements.

  • Governance controls include data residency (regional model availability), retention policies, approval workflows for model deployment, and documented risk assessments.

  • Generative AI security risks: prompt injection, training data poisoning, insecure output handling, and model denial of service; mitigations include input validation, output sanitization, and guardrails.

AWS AI Practitioner key terms

Artificial Intelligence (AI)

The broad field of building computer systems that perform tasks normally requiring human intelligence, such as reasoning, language understanding, perception, and decision-making. It encompasses machine learning, generative AI, and rule-based approaches.

Related: Machine Learning (ML), Generative AI
Machine Learning (ML)

A subset of AI where algorithms learn patterns from data rather than following hand-written rules, then use those patterns to make predictions or decisions on new, unseen data.

Related: Training Data, Inference, Supervised Learning
Supervised Learning

A training approach that uses labeled examples, where each input is paired with the correct answer, so the model learns to map inputs to outputs. Common for classification and regression tasks.

Related: Training Data, Unsupervised Learning, Model Evaluation
Unsupervised Learning

A training approach that works with unlabeled data, letting the algorithm discover structure such as clusters or groupings on its own. Used for segmentation, anomaly detection, and pattern discovery.

Related: Supervised Learning, Clustering
Clustering

An unsupervised technique that groups similar data points together based on shared characteristics, without predefined labels. Useful for customer segmentation, document grouping, and exploratory analysis.

Related: Unsupervised Learning
Reinforcement Learning (RL)

A training paradigm where an agent learns by taking actions in an environment and receiving rewards or penalties, gradually discovering behavior that maximizes cumulative reward. Often used for control and game-playing tasks.

Related: Reinforcement Learning from Human Feedback (RLHF)
Training Data

The dataset used to teach a model its patterns and parameters. Its quality, size, and representativeness strongly influence model accuracy and fairness.

Related: Machine Learning (ML), Bias, Model Evaluation
Inference

The process of using a trained model to produce predictions or outputs on new input data. It happens after training and is what end users typically trigger in production.

Related: Machine Learning (ML), Foundation Model
Model Evaluation

The practice of measuring how well a trained model performs on data it has not seen, using metrics and held-out test sets, to judge accuracy, generalization, and readiness for deployment.

Related: Training Data, Overfitting
Overfitting

When a model learns its training data too closely, including noise, and performs poorly on new data. It signals weak generalization and is often addressed with more data or regularization.

Related: Model Evaluation, Training Data
Generative AI

A category of AI that creates new content such as text, images, audio, or code in response to a prompt, rather than only classifying or predicting labels from existing data.

Related: Foundation Model, Large Language Model (LLM), Prompt
Foundation Model

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

Related: Large Language Model (LLM), Fine-Tuning, Generative AI
Large Language Model (LLM)

A foundation model trained on vast amounts of text to understand and generate language. It predicts likely token sequences and powers tasks like summarization, translation, and question answering.

Related: Token, Foundation Model, Prompt
Token

A small unit of text, such as a word, word fragment, or character, that a language model processes. Token counts drive context limits and often pricing.

Related: Large Language Model (LLM), Context Window
Prompt

The input instruction or question given to a generative model to guide its output. Clear, specific prompts generally produce more relevant and reliable responses.

Related: Prompt Engineering, Large Language Model (LLM)
Prompt Engineering

The practice of designing and refining prompts, including instructions, examples, and formatting, to obtain accurate, relevant, and well-structured responses from a foundation model.

Related: Prompt, Few-Shot Prompting, Retrieval-Augmented Generation (RAG)
Few-Shot Prompting

A prompting technique that includes a small number of worked examples in the prompt so the model can infer the desired format or reasoning pattern before answering the real request.

Related: Prompt Engineering, Zero-Shot Prompting
Zero-Shot Prompting

Asking a model to perform a task with instructions only and no worked examples. It relies on the model's pretrained knowledge and works best for common, well-understood tasks.

Related: Few-Shot Prompting, Prompt Engineering
Context Window

The maximum amount of text, measured in tokens, that a model can consider at one time, including the prompt and its response. Larger windows allow longer documents and conversations.

Related: Token, Retrieval-Augmented Generation (RAG)
Retrieval-Augmented Generation (RAG)

A pattern that retrieves relevant documents from an external knowledge source and adds them to the prompt, so the model answers using current, domain-specific information without retraining.

Related: Embedding, Vector Database, Hallucination
Embedding

A numeric vector representation of text, images, or other data that captures meaning, so items with similar meaning sit close together. Embeddings enable semantic search and retrieval.

Related: Vector Database, Retrieval-Augmented Generation (RAG)
Vector Database

A data store optimized for storing embeddings and quickly finding the most similar vectors to a query. It is a common backing service for retrieval-augmented generation and semantic search.

Related: Embedding, Retrieval-Augmented Generation (RAG)
Fine-Tuning

Further training of a pretrained foundation model on a smaller, task-specific dataset so it adapts to a particular domain, style, or behavior. It is more costly than prompting alone.

Related: Foundation Model, Training Data
Hallucination

When a generative model produces confident but incorrect or fabricated information. It is a key reliability risk and is often mitigated with retrieval, grounding, and human review.

Related: Retrieval-Augmented Generation (RAG), Human-in-the-Loop
Bias

Systematic unfairness in model outputs, often arising from skewed or unrepresentative training data. It can cause discriminatory outcomes and must be measured and mitigated.

Related: Training Data, Fairness
Fairness

The principle that AI systems should treat individuals and groups equitably and avoid discriminatory outcomes. It is assessed through metrics, testing, and diverse data.

Related: Bias, Responsible AI
Explainability

The degree to which people can understand why an AI system produced a given output. It supports trust, debugging, auditing, and accountability in decisions affecting users.

Related: Responsible AI, Human-in-the-Loop
Responsible AI

A set of principles and practices for building AI that is fair, transparent, safe, accountable, and privacy-respecting. It guides design, testing, and monitoring across the model lifecycle.

Related: Fairness, Explainability, Human-in-the-Loop
Human-in-the-Loop

A design where people review, approve, or correct AI outputs at key points. It reduces risk in high-stakes decisions and supports accountability and quality control.

Related: Hallucination, Responsible AI
Guardrails

Configurable controls that constrain model behavior, such as content filters, topic restrictions, and blocked outputs. They help enforce safety, policy, and brand requirements in applications.

Related: Responsible AI, Content Moderation
Content Moderation

Automated or human review that detects and filters harmful, offensive, or policy-violating content in inputs and outputs. It protects users and supports safe deployment of generative applications.

Related: Guardrails
Data Privacy

Protecting personal and sensitive information throughout collection, storage, and processing, including limiting what is shared with AI services and honoring consent and retention rules.

Related: Encryption, Access Control
Encryption

Encoding data so only authorized parties can read it, both while stored and while in transit. It is a core safeguard for protecting sensitive data used by AI systems.

Related: Data Privacy, Access Control
Access Control

Policies and mechanisms that determine who or what can view or use specific resources. Least-privilege access limits exposure of sensitive data and model endpoints.

Related: Data Privacy, Encryption, Identity and Access Management (IAM)
Identity and Access Management (IAM)

The framework for managing user identities and their permissions across systems. In AI workloads it governs who can invoke models, access data, and change configurations.

Related: Access Control, Audit Trail
Audit Trail

A chronological record of system activity, such as who accessed data or invoked a model and when. It supports investigations, compliance reporting, and accountability.

Related: Governance, Identity and Access Management (IAM)
Governance

The policies, roles, and oversight processes that direct how AI systems are approved, monitored, and retired. It aligns AI use with legal, ethical, and organizational requirements.

Related: Audit Trail, Responsible AI
Reinforcement Learning from Human Feedback (RLHF)

A training method that uses human rankings of model outputs to shape a reward signal, then tunes the model toward responses people prefer. It improves helpfulness and alignment.

Related: Reinforcement Learning (RL), Fine-Tuning

AWS AI Practitioner study plan

8 weeks · About 6 hours per week

Week 1

Fundamentals of Generative AI

  • Define generative AI, foundation models, large language models, and tokenization in your own words without notes

  • Distinguish pretraining, fine-tuning, and in-context learning by writing one concrete example of each

  • Explain embeddings, vectors, and semantic similarity well enough to sketch how a retrieval system matches a question to a document

  • Compare foundation model capabilities and limitations against traditional supervised ML for three business scenarios

Your checkpoint

A one-page glossary of 20 generative AI terms, each with a self-written definition and one example

Week 2

Fundamentals of Generative AI

  • Describe how transformer attention enables context handling and explain why context windows constrain prompt design

  • Differentiate prompting techniques (zero-shot, few-shot, chain-of-thought) and select the right one for four sample tasks

  • Identify sources of hallucination and non-determinism and propose one mitigation for each

  • Compare Amazon Bedrock, Amazon SageMaker JumpStart, and Amazon Q at a conceptual level for model access and use

Your checkpoint

A comparison table of three prompting techniques with task fit, example prompt, and expected failure mode

Week 3

Applications of Foundation Models

  • Map foundation model application patterns (text generation, summarization, RAG, agents, image generation) to business use cases

  • Explain retrieval-augmented generation end to end, including ingestion, chunking, embedding, vector search, and augmented prompt construction

  • Describe how Amazon Bedrock supports model choice, knowledge bases, guardrails, and agents without managing infrastructure

  • Evaluate a proposed use case for feasibility, cost drivers, and latency trade-offs

Your checkpoint

A written RAG architecture walkthrough for one chosen use case, covering data flow from document ingestion to final answer

Week 4

Applications of Foundation Models & Fundamentals of AI and ML

  • Distinguish supervised, unsupervised, and reinforcement learning and match each to a realistic example

  • Explain the roles of training, validation, and test data splits and why leakage invalidates evaluation

  • Select evaluation metrics (accuracy, precision, recall, F1, RMSE) appropriate to classification and regression problems

  • Describe how foundation model evaluation differs from classical ML evaluation, including human review and benchmark limits

Your checkpoint

A 15-question self-quiz on ML fundamentals and model evaluation, scored with a written list of missed concepts

Week 5

Fundamentals of AI and ML

  • Explain the AWS shared responsibility model as it applies to AI workloads and identify which controls remain the customer's

  • Describe core AWS AI services (Amazon Rekognition, Comprehend, Transcribe, Polly, Textract, Translate) and match each to a task

  • Differentiate Amazon SageMaker capabilities for building, training, and deploying custom models at a conceptual level

  • Identify which AWS AI service or approach fits five given business requirements

Your checkpoint

A service-selection matrix mapping 10 business requirements to a recommended AWS AI service with a one-line justification each

Week 6

Guidelines for Responsible AI

  • Define the core responsible AI dimensions: fairness, explainability, privacy, safety, controllability, verifiability, and robustness

  • Identify bias sources across data collection, labeling, model training, and deployment, and propose a mitigation for each stage

  • Explain how Amazon Bedrock Guardrails filters harmful content, redacts sensitive information, and grounds responses

  • Draft transparency and human-oversight requirements for a customer-facing generative AI assistant

Your checkpoint

A responsible AI review checklist of at least 12 items applied to one previously chosen use case, with pass/fail notes

Week 7

Security, Compliance, and Governance for AI Solutions

  • Apply least-privilege access control to AI workloads using IAM roles, policies, and separation of duties

  • Describe data protection practices for AI systems including encryption at rest and in transit, and handling of sensitive data in prompts

  • Explain logging, monitoring, and auditability requirements for generative AI applications using AWS services

  • Identify governance practices for model inventory, change management, and regulatory and compliance considerations

Your checkpoint

A one-page security and governance control map linking 10 controls to the specific risk each mitigates

Week 8

Fundamentals of AI and ML & Fundamentals of Generative AI & Applications of Foundation Models & Guidelines for Responsible AI & Security, Compliance, and Governance for AI Solutions

  • Consolidate weak areas by re-reviewing every missed item from prior quizzes and deliverables

  • Complete two timed full-length practice exams and record per-domain accuracy

  • Explain each missed practice question aloud or in writing, citing the underlying concept rather than memorizing the answer

  • Build a final one-page summary sheet covering the highest-weight domain concepts for last-minute review

Your checkpoint

A scored practice exam report with per-domain accuracy, a ranked list of remaining weak topics, and a final review sheet

Milestones

Week 2: Score at least 80% on a self-made 20-question generative AI fundamentals quiz covering terminology, prompting, and hallucination mitigation

Week 4: Explain a complete RAG pipeline and the difference between classical ML and foundation model evaluation without notes, and score at least 75% on the week 4 self-quiz

Week 6: Produce a responsible AI checklist applied to a real use case that identifies at least three concrete bias or safety risks with mitigations

Week 8: Achieve at least 85% overall on a timed full-length practice exam with no domain below 70%

AWS AI Practitioner exam day

Your checklist

Government-issued photo ID

Bring the original, physical ID you used when registering. Confirm the exact accepted ID types and name-match rules with the test provider before exam day.

Appointment confirmation and test-provider login

Have your appointment confirmation number and the username/password for your testing account ready. For online proctoring, run the provider's system test in advance.

Know your exam code and format

Confirm the exact exam code, question count, and time limit in the official AWS exam guide, since these can change. Do not rely on secondhand summaries.

Review the official exam guide and task domains

Use the published domain weightings to decide where to spend your last study hours. Note any domains you have not practiced.

Practice with timed question sets

Complete at least one full-length timed practice run so you know your pacing. Do not memorize question text; focus on why each answer is right or wrong.

Prepare a one-page summary sheet

Write concise notes on core concepts, service comparisons, and common traps. Review this sheet, not new material, in the final hours.

Test-day logistics plan

For a test center, confirm the address, parking, and arrival time. For online proctoring, confirm your room, desk, and internet meet the provider's requirements. Verify details with the provider.

Quiet, cleared testing environment

For online exams, clear your desk and room of notes, phones, and smart devices. Confirm the provider's rules on permitted items and room scans.

Backup internet and power

For online proctoring, have a backup connection (for example, a mobile hotspot) and a charged device. Confirm the provider's policy on disconnections.

Rest and hydration plan

Plan a normal night's sleep and a light meal before the exam. Avoid cramming new topics in the final hours.

Confirm break and reschedule policies

Check whether your exam allows scheduled or unscheduled breaks and what happens if you need to reschedule. Policies vary; confirm with the provider.

Confirm score-report and retake rules

Know how and when you will receive your result and what the retake waiting period is. These details can change; confirm with the provider.

Plan the day

72 hours before

Take one full-length timed practice exam under realistic conditions. Review every missed question and tag weak domains. Spend the remaining time on those domains only, using official documentation and your summary sheet.

72 hours before

Re-read the official exam guide and confirm logistics: test center address or online-proctoring system check, ID requirements, and permitted items. Resolve any uncertainty with the provider now, not on exam day.

24 hours before

Stop taking new practice tests. Review your one-page summary sheet and a short list of commonly confused concepts. Prepare your ID, confirmation details, and environment. Set alarms and get a full night's sleep.

Morning of

Eat a light meal and avoid heavy caffeine if you are sensitive. Do a final 15-20 minute review of your summary sheet only. Arrive early or log in early as the provider instructs; do not start new topics.

During the test

Read each question fully and identify what is being asked before looking at options. Answer easy questions first, flag uncertain ones, and return to them. Eliminate clearly wrong options, then choose the best fit. Watch the clock and keep a steady pace. If a break is allowed, confirm the rules before taking it.

Final minutes

Use remaining time to review flagged questions. Change an answer only when you have a clear reason. Submit according to the provider's instructions and note how you will receive your score.

Rules to check

  • Bring the original government-issued photo ID you registered with. The name must match your registration; confirm accepted ID types and name-change procedures with the provider.

  • Follow all proctor instructions. Do not communicate with others, read questions aloud, or use unauthorized materials, devices, or notes during the exam.

  • For online proctoring, keep your desk clear and remain in the camera's view. Confirm the provider's rules on room scans, permitted items, and movement before you begin.

  • Breaks are governed by the provider's policy for your exam and delivery method. Confirm in advance whether breaks are allowed, whether the clock stops, and what check-in is required on return.

  • Scoring is based on the number of correct answers against a passing standard set by the certifying body. There is no penalty for guessing, so answer every question.

  • If you experience a technical issue, stop and contact the proctor or provider support immediately. Do not attempt to fix it yourself in ways that violate testing rules; confirm the incident-reporting process in advance.

  • Rescheduling, cancellation, and retake rules have deadlines and fees that can change. Confirm the current policy with the provider before your appointment.

  • Do not share, reproduce, or discuss actual exam content. Doing so violates the certification agreement and can invalidate your result.

AWS AI Practitioner FAQ

How many AWS AI Practitioner practice questions are available?

The practice set has 200 AWS AI Practitioner questions. Each one includes the choices, the correct answer, and an explanation, and the set is free to read with no signup.

How many AWS AI Practitioner flashcards are available?

The flashcard set has 32 AWS AI Practitioner cards. Each card shows the term and its definition.

Which topics does the AWS AI Practitioner study material cover?

The AWS AI Practitioner study material has 5 sections, including Fundamentals of AI and ML, Fundamentals of Generative AI, and Applications of Foundation Models.

What format facts does the catalog list for the AWS AI Practitioner exam?

The catalog lists these item types: multiple choice and multiple response for the AWS AI Practitioner exam.