AWS Certified AI Practitioner AIF-C01 AIF-C01 Questions 76–150 | Page 2/12

MCQmedium

A financial services firm is deploying a generative AI assistant that answers employee questions about internal policies. Compliance requires that every response cite the exact policy document and section used. The assistant currently relies only on the foundation model's pretrained knowledge and frequently invents policy details. Which technique should the team implement to ground responses in the firm's own documents and produce citations?

A.Implement Retrieval Augmented Generation by embedding the policy documents in a vector store and retrieving relevant passages at query time.

B.Increase the model's temperature setting so it explores more of its pretrained knowledge about corporate policy.

C.Lower the maximum token limit so the assistant gives shorter answers that are less likely to contain errors.

D.Fine-tune the foundation model on a large corpus of public financial regulations.

AnswerA

Retrieval Augmented Generation embeds source documents, retrieves the passages most semantically similar to the user's question, and passes them into the prompt so the model answers from that supplied context. Because the retrieved chunks carry document and section metadata, the assistant can cite the exact source. It also keeps answers current as policies change without retraining.

Why this answer

Grounding a model in proprietary content requires supplying that content at inference time rather than relying on pretrained weights. Retrieval Augmented Generation retrieves the most relevant document chunks and places them in the prompt, so the model's answer is conditioned on real policy text and can reference the source document and section. Sampling or length adjustments do not add knowledge.

Exam trap

The trap here is believing that fine-tuning on domain text guarantees accurate citations, when only retrieval of the actual source documents provides verifiable provenance.

Full explanation →

MCQeasy

According to AWS's responsible AI principles, which principle focuses on the idea that AI systems should produce consistent and reliable results even under unexpected conditions?

A.Safety

B.Veracity

C.Fairness

D.Robustness

AnswerD

Robustness covers an AI system's ability to maintain consistent, reliable performance when inputs or conditions deviate from training expectations. This directly matches the stem's requirement for dependable results under unexpected conditions, distinguishing it from fairness, explainability, or privacy principles.

Why this answer

Robustness in AWS's responsible AI framework refers to an AI system's ability to maintain consistent, reliable performance even when inputs are unexpected, adversarial, or drawn from edge cases outside the training distribution. It encompasses resilience to noise, distributional shift, and adversarial manipulation. This is distinct from safety, which focuses on preventing harm, and veracity, which concerns truthfulness of outputs.

Exam trap

AIF-C01 often tests the overlap between Safety, Robustness, and Veracity; candidates confuse 'reliable under unexpected conditions' (Robustness) with 'truthful outputs' (Veracity) or 'no harm' (Safety).

How to eliminate wrong answers

Option A is wrong because Safety focuses on preventing AI systems from causing physical, psychological, or societal harm, not on consistency under unexpected conditions. Option B is wrong because Veracity addresses whether AI outputs are truthful and accurate, not whether the system remains reliable under stress or distributional shift. Option C is wrong because Fairness concerns equitable treatment across demographic groups and avoidance of bias, not operational reliability under unexpected inputs.

Full explanation →

MCQmedium

A data science team is using SHAP values to explain a complex model. They notice that for a particular prediction, the SHAP value for feature 'age' is +0.3. What does this indicate?

A.The age feature is the most important feature globally

B.The age feature has a negative impact on the prediction

C.The age feature is causing bias in the model

D.The age feature increased the prediction by 0.3 units from the baseline

AnswerD

A SHAP value expresses a feature's marginal contribution to a single prediction relative to the baseline expected output. A value of +0.3 for age therefore means that feature pushed this prediction 0.3 units above the baseline, not that age equals 0.3.

Why this answer

A SHAP value of +0.3 for the 'age' feature on a specific prediction means that, relative to the model's baseline (expected) output, the age feature pushed the prediction upward by 0.3 units. SHAP values are additive and local: they decompose a single prediction into per-feature contributions that sum to the difference between the prediction and the base value. The sign indicates direction, and the magnitude indicates the size of that feature's contribution for that instance.

Exam trap

AIF-C01 often tests the confusion between local SHAP values (per-prediction feature contributions) and global feature importance; candidates incorrectly assume a large SHAP value means the feature is globally most important.

How to eliminate wrong answers

Option A is wrong because a single SHAP value is a local explanation for one prediction; global feature importance requires aggregating absolute SHAP values across many instances. Option B is wrong because the positive sign (+0.3) indicates an upward push on the prediction, not a negative impact. Option C is wrong because SHAP values explain model behavior, not whether the feature causes bias; bias detection requires separate fairness analysis across groups.

Full explanation →

MCQmedium

A company wants to automatically categorize customer support tickets into predefined categories such as 'billing', 'technical', and 'account'. Which AWS service is BEST suited for this task?

A.Amazon Transcribe

B.Amazon Rekognition

C.Amazon Textract

D.Amazon Comprehend

AnswerD

Amazon Comprehend provides pre-trained natural language processing that performs custom multi-class text classification, letting you train a classifier on labelled tickets to assign predefined categories. This directly satisfies the requirement to categorise support tickets automatically, without building or hosting your own model.

Why this answer

Amazon Comprehend is a natural language processing (NLP) service that uses machine learning to extract insights from text, including custom classification. It can be trained to automatically categorize support tickets into predefined categories like 'billing', 'technical', and 'account' by analyzing the text content of each ticket.

Exam trap

The trap is that candidates often confuse Amazon Comprehend (NLP for text analysis) with Amazon Textract (OCR for document text extraction), mistakenly thinking that extracting text from a ticket is the same as understanding or categorizing it.

How to eliminate wrong answers

Option A is wrong because Amazon Transcribe is a speech-to-text service that converts audio to text, not a text classification or categorization service. Option B is wrong because Amazon Rekognition is an image and video analysis service, not designed for processing or categorizing text-based content. Option C is wrong because Amazon Textract is an OCR service that extracts text and data from scanned documents, but it does not perform natural language understanding or classification of the extracted text.

Full explanation →

Multi-Selectmedium

A marketing team wants to build a generative AI application that produces short promotional copy and accompanying images for product launches. They are evaluating Amazon Bedrock and want to understand which statements accurately describe its capabilities for this use case. (Choose two.)

Select 2 answers

A.Amazon Bedrock requires customers to provision and manage the underlying GPU infrastructure for each model.

B.Amazon Bedrock supports both text generation models and image generation models.

C.Amazon Bedrock only supports models hosted in a single AWS Region worldwide.

D.Amazon Bedrock provides access to multiple foundation models from different providers through a single API.

E.Amazon Bedrock automatically guarantees that generated promotional claims are legally compliant.

AnswersB, D

Bedrock includes text models for copywriting and image models such as Amazon Titan Image Generator and Stability AI models for visuals. This means the team can generate promotional copy and accompanying images from the same service, satisfying the multi-modal requirement without stitching together separate platforms.

Why this answer

Amazon Bedrock offers a unified API across many foundation models and includes both text and image generation capabilities, which fits a promotional copy and imagery workload. It is serverless with respect to model hosting, is available in many Regions, and does not guarantee legal compliance, so the two accurate capability statements are the multi-provider single API and support for text and image models.

Exam trap

The trap here is conflating Bedrock's managed model access with infrastructure management or automatic compliance guarantees, neither of which Bedrock provides.

Full explanation →

MCQeasy

A media company wants to build a generative AI assistant that drafts scripts and answers questions about its own style guide. The team has no machine learning engineers and wants to avoid managing GPU infrastructure or training any models. Which approach BEST describes how they should build this solution?

A.Build a rules-based template engine that reassembles stored sentences from previous scripts.

B.Deploy an Amazon SageMaker endpoint hosting an open-source model and manually patch the underlying EC2 instances.

C.Collect a labeled dataset and train a new transformer from scratch on Amazon SageMaker training jobs.

D.Use a foundation model through Amazon Bedrock and supply the style guide as context at inference time.

AnswerD

Foundation models accessed through Amazon Bedrock are pretrained, so no training or GPU fleet management is required, and the style guide can be supplied as retrieved context in the prompt. This matches the goal of a no-ML-team, serverless generative AI solution while still grounding responses in proprietary content.

Why this answer

A pretrained foundation model delivered through a managed service removes the need to train models or operate GPU infrastructure, and the proprietary style guide can be injected as prompt context so outputs follow house rules. This combination satisfies the no-ML-team constraint while still producing grounded, task-specific generative output.

Exam trap

The trap here is assuming that any generative AI use case requires training or fine-tuning a model, when pretrained foundation models plus prompt context are usually sufficient.

Full explanation →

Multi-Selectmedium

Which TWO AWS services can be used to build a chatbot that responds to customer inquiries using a company's documentation as source? (Select two.)

Select 2 answers

A.Amazon Bedrock with RAG

B.Amazon Polly

C.Amazon Q Business

D.Amazon Transcribe

E.Amazon Lex

AnswersA, C

Amazon Bedrock supplies foundation models, while Retrieval Augmented Generation grounds responses in the company's own documentation by retrieving relevant passages and passing them to the model, satisfying the requirement to answer inquiries from that source rather than relying on pretrained knowledge alone.

Why this answer

Amazon Bedrock with RAG (A) is correct because it lets you build generative AI chatbots that retrieve relevant passages from your company documentation (stored in services like Amazon S3 or OpenSearch) and pass them to a foundation model as context, so answers are grounded in the source material. Amazon Q Business (C) is correct because it is a fully managed generative AI assistant that natively connects to enterprise data sources (S3, SharePoint, Confluence, etc.), indexes the documentation, and answers customer inquiries with citations from that content. Amazon Polly (B) is only a text-to-speech service and cannot retrieve or reason over documentation, Amazon Transcribe (D) only converts speech to text, and Amazon Lex (E) builds conversational interfaces/intents but does not itself perform retrieval-augmented generation over a company's documentation.

Exam trap

AWS often tests the distinction between services that provide conversational interfaces (like Lex) versus those that enable retrieval-augmented generation from custom data sources (like Bedrock with RAG or Q Business), leading candidates to mistakenly select Lex because it is a chatbot service, even though it lacks native document retrieval capabilities.

Full explanation →

MCQmedium

A team is deploying a regression model for loan approval. To ensure transparency for regulators, they need to explain individual predictions. Which interpretability method can provide local explanations by approximating the model with a simpler surrogate?

A.SHAP values

B.Partial dependence plots

C.LIME

D.Permutation feature importance

AnswerC

LIME perturbs individual instances and fits a sparse interpretable surrogate, such as a linear model, around each prediction. This yields local feature attributions explaining why a specific loan application was approved or rejected, giving regulators case-level transparency rather than only global feature importance.

Why this answer

LIME (Local Interpretable Model-agnostic Explanations) is the correct choice because it generates local explanations by fitting a simpler, interpretable surrogate model (e.g., linear regression or decision tree) around a single prediction. This allows the team to explain why a specific loan application was approved or rejected, meeting regulatory transparency requirements without needing access to the original model's internals.

Exam trap

The trap here is that candidates confuse SHAP values (which also provide local explanations) with LIME, but SHAP does not use a simpler surrogate model—it directly computes feature attributions from the original model, which is a key distinction the exam tests.

How to eliminate wrong answers

Option A is wrong because SHAP values provide local explanations based on cooperative game theory (Shapley values) but do not approximate the model with a simpler surrogate; instead, they compute additive feature contributions directly from the original model. Option B is wrong because partial dependence plots show the average marginal effect of a feature on the model's predictions across the entire dataset, not local explanations for individual predictions. Option D is wrong because permutation feature importance measures the global drop in model performance when a feature is shuffled, offering no local or surrogate-based interpretability for a single prediction.

Full explanation →

MCQeasy

A company is using Amazon Bedrock to generate code snippets. Developers report that the generated code sometimes contains security vulnerabilities. Which action should the team take to mitigate this risk?

A.Deploy the model in a sandbox environment to limit its access to sensitive systems.

B.Implement a manual code review process after generation.

C.Add a system prompt that instructs the model to follow security best practices and avoid known vulnerabilities.

D.Reduce the temperature parameter to 0 to make the output deterministic.

AnswerC

A system prompt steers the foundation model at inference time, so instructing it to follow security best practices directly reduces vulnerable code output without retraining or infrastructure changes. This satisfies the scenario's constraint of mitigating risk within the existing Amazon Bedrock setup, though prompt-level guidance offers weaker assurance than automated code scanning.

Why this answer

Adding a system prompt that instructs the model to follow security best practices and avoid known vulnerabilities directly influences the model's output at inference time. Amazon Bedrock supports system prompts that act as high-level instructions to guide the foundation model's behavior, making this a proactive, scalable mitigation that does not require manual intervention or architectural changes.

Exam trap

AWS often tests the misconception that reducing temperature or isolating the environment can fix output quality issues, when in fact only prompt-level guidance directly addresses the model's generation behavior.

How to eliminate wrong answers

Option A is wrong because deploying the model in a sandbox environment limits access to sensitive systems but does not prevent the model from generating code with security vulnerabilities; the model's output itself remains unchanged. Option B is wrong because implementing a manual code review process after generation is a reactive measure that does not reduce the risk at the source; it adds human overhead and delays, and is not a mitigation that addresses the model's tendency to produce insecure code. Option D is wrong because reducing the temperature parameter to 0 makes the output deterministic but does not teach or enforce security best practices; it only reduces randomness, not the likelihood of generating vulnerable patterns.

Full explanation →

MCQeasy

A developer needs to evaluate the quality of a text summarization model by comparing its output to reference summaries. Which automated metric measures the overlap of n‑grams between the generated and reference summaries?

A.ROUGE

B.BERTScore

C.METEOR

D.BLEU

AnswerA

ROUGE computes recall-oriented n-gram overlap between generated and reference summaries, directly satisfying the stem's requirement for an automated metric measuring n-gram coincidence. Its variants (ROUGE-1, ROUGE-2, ROUGE-L) quantify unigram, bigram and longest-common-subsequence matching, making it the standard summarisation evaluation metric.

Why this answer

ROUGE measures n‑gram overlap and is commonly used for summarization. BLEU is for translation, BERTScore uses embeddings, and METEOR accounts for synonyms — but the question specifically asks about n‑gram overlap.

Full explanation →

MCQeasy

A data scientist needs to predict whether a transaction is fraudulent (Yes/No). Which type of machine learning problem is this?

A.Classification

B.Clustering

C.Regression

D.Reinforcement learning

AnswerA

Predicting a discrete Yes/No label means the target variable is categorical, which is supervised classification. Regression would output a continuous value, and clustering is unsupervised, so neither fits the fraudulent-or-not constraint. Classification models learn a decision boundary separating the two classes.

Why this answer

This is a classification problem because the output is a discrete category (fraudulent or not). The data scientist is predicting a binary label (Yes/No), which is the defining characteristic of binary classification in supervised learning.

Exam trap

The AWS AI Practitioner exam often tests the distinction between supervised and unsupervised learning, and the trap here is confusing classification (supervised, discrete output) with clustering (unsupervised, no labels) because both involve grouping or categorizing data.

How to eliminate wrong answers

Option B is wrong because clustering is an unsupervised learning technique that groups data into clusters based on similarity, without predefined labels; it cannot predict a specific binary outcome like fraud. Option C is wrong because regression predicts continuous numerical values (e.g., transaction amount), not discrete categories. Option D is wrong because reinforcement learning involves an agent learning optimal actions through rewards and penalties in an environment, not predicting a static label from historical data.

Full explanation →

MCQeasy

A company wants to identify customer segments based on purchasing behavior. They have unlabeled transaction data and do not know the segment definitions beforehand. Which type of machine learning should they use?

A.Semi-supervised learning

B.Unsupervised learning

C.Supervised learning

D.Reinforcement learning

AnswerB

Unsupervised learning finds structure in unlabelled data without predefined outputs, so clustering algorithms can group transactions by purchasing behaviour and reveal segments. This satisfies the stem's constraints: no labels exist and segment definitions are unknown beforehand.

Why this answer

Unsupervised learning is the correct choice because the company has unlabeled transaction data and no predefined segment definitions. This type of machine learning discovers hidden patterns, groupings, or structures in data without requiring labeled outputs, making it ideal for customer segmentation tasks like clustering.

Exam trap

AWS often tests the distinction between supervised and unsupervised learning by presenting unlabeled data scenarios, where candidates mistakenly choose supervised learning because they confuse 'identifying segments' with 'predicting a known label.'

How to eliminate wrong answers

Option A is wrong because semi-supervised learning uses a small amount of labeled data alongside a larger unlabeled dataset, which does not match the scenario of having no labels or segment definitions. Option C is wrong because supervised learning requires labeled training data with known target outputs (e.g., predefined customer segments), which the company lacks. Option D is wrong because reinforcement learning involves an agent learning through trial-and-error interactions with an environment to maximize a reward signal, not for discovering patterns in static transaction data.

Full explanation →

MCQeasy

A company wants to use a foundation model to automatically moderate user-generated content. The model must filter out inappropriate content with high accuracy. Which Amazon service is best suited for this task?

A.Amazon Translate

B.Amazon Rekognition

C.Amazon Polly

D.Amazon Comprehend

AnswerD

Amazon Comprehend provides pre-trained content moderation that detects harmful or inappropriate text categories, returning confidence scores for filtering user-generated content. It delivers the required accuracy without building custom models, unlike general-purpose services such as Bedrock or Rekognition.

Why this answer

Amazon Comprehend is the correct choice because it is a natural language processing (NLP) service that can analyze text for sentiment, key phrases, and — critically — toxicity and inappropriate content using built-in or custom classifiers. This directly matches the requirement to moderate user-generated text with high accuracy, as it can detect hate speech, profanity, and other harmful language.

Exam trap

The trap here is that candidates may confuse Amazon Rekognition's ability to detect unsafe content in images with the need to moderate text, leading them to select Rekognition instead of recognizing that Comprehend is the NLP service for text analysis.

How to eliminate wrong answers

Option A is wrong because Amazon Translate is a machine translation service that converts text between languages; it has no capability to analyze or moderate content for appropriateness. Option B is wrong because Amazon Rekognition is designed for image and video analysis (e.g., object detection, facial recognition, unsafe content detection in visual media), not for moderating text-based user-generated content. Option C is wrong because Amazon Polly is a text-to-speech service that converts written text into lifelike speech; it performs no content moderation or analysis.

Full explanation →

MCQeasy

A startup needs to build a real-time text translation feature for a customer chat application. Latency must be under 200 ms per request. Which AWS approach is BEST suited?

A.Use Amazon Comprehend for language detection and a custom translation model

B.Use Amazon Bedrock with a multilingual foundation model

C.Use Amazon Translate with real-time translation

D.Use Amazon Transcribe and then Amazon Bedrock

AnswerC

Amazon Translate provides synchronous, real-time translation via its TranslateText API, returning results within milliseconds, which satisfies the sub-200 ms latency constraint for live chat. Purpose-built neural machine translation avoids the overhead of provisioning or invoking general-purpose models, making it the appropriate managed service for low-latency text translation.

Why this answer

Amazon Translate's real-time translation API is purpose-built for low-latency text translation, typically achieving sub-200 ms response times for small payloads. It directly translates text without the overhead of running a large foundation model or performing intermediate steps like transcription, making it the best fit for this latency-sensitive chat application.

Exam trap

The trap here is that candidates may assume a large foundation model (e.g., via Bedrock) is always the best choice for multilingual tasks, overlooking that purpose-built services like Amazon Translate are specifically optimized for low-latency, high-throughput translation at a fraction of the cost and complexity.

How to eliminate wrong answers

Option A is wrong because Amazon Comprehend is designed for natural language processing (e.g., sentiment analysis, entity extraction), not real-time translation, and building a custom translation model would introduce significant latency and complexity. Option B is wrong because Amazon Bedrock with a multilingual foundation model introduces inference latency that often exceeds 200 ms for real-time requests, and it is not optimized for the single-purpose, high-throughput translation task required here. Option D is wrong because Amazon Transcribe is for speech-to-text, not text translation, and chaining it with Bedrock adds unnecessary latency and complexity for a text-only translation feature.

Full explanation →

MCQmedium

A developer is using Amazon Bedrock Agents to build a multi-step reasoning workflow. The agent needs to query an external REST API to fetch data. What must the developer define in the agent configuration?

A.A prompt template with the API endpoint URL

B.A knowledge base with the API schema

C.An action group with the API schema and a Lambda function

D.A guardrail to filter API responses

AnswerC

An action group binds the REST API to the agent by supplying its OpenAPI schema, which tells the agent which operations exist and their parameters, plus a Lambda function that executes the call and returns results. This satisfies the stem's requirement to query an external REST API during multi-step reasoning.

Why this answer

Action groups allow agents to invoke external APIs (via AWS Lambda) or knowledge bases. The developer must define an action group that describes the API operations the agent can call, along with the Lambda function that implements the API integration.

Full explanation →

Multi-Selecthard

A financial services company is deploying a machine learning model that must comply with SOC 2 and PCI DSS. They need to ensure that the model artifacts and training data are encrypted, access is audited, and the environment is protected from network threats. Which THREE AWS services should they implement?

Select 3 answers

A.Amazon DynamoDB Accelerator (DAX)

B.Amazon GuardDuty

C.AWS CloudTrail

D.AWS KMS

E.AWS WAF

AnswersB, C, D

GuardDuty continuously monitors for malicious activity and network threats.

Why this answer

Amazon GuardDuty is a threat detection service that continuously monitors for malicious activity and unauthorized behavior, which helps protect the environment from network threats as required by SOC 2 and PCI DSS. By analyzing VPC Flow Logs, DNS logs, and CloudTrail events, it can detect anomalies such as port scanning or data exfiltration, directly addressing the need for network threat protection in a compliant ML deployment.

Exam trap

AWS often tests the distinction between network threat detection (GuardDuty) and web application layer protection (WAF), leading candidates to incorrectly choose WAF for general network threat protection when it only addresses HTTP/S-based attacks.

Full explanation →

MCQhard

A company has trained a model using Amazon SageMaker and stored the model artifacts in S3 with SSE-KMS encryption. The development team wants to grant cross-account access to the model artifacts so a partner can deploy the model in their own account. Which steps are required?

A.Create a new IAM role in the partner account with S3 read permissions

B.Create an S3 bucket policy allowing the partner account to read the objects, and ensure the KMS key policy allows the partner account to use the key

C.Use AWS Lake Formation to share the data

D.Copy the model artifacts to a public S3 bucket

AnswerB

Cross-account access requires permissions at both layers: the S3 bucket policy grants the partner account read access to the objects, while the SSE-KMS key policy must separately allow that account to decrypt using the customer managed key. Missing either blocks access.

Why this answer

To share SSE-KMS encrypted objects cross-account, you must grant the partner account access to both the S3 object (via bucket policy) and the KMS key (via key policy). The partner must also have the correct IAM permissions.

Full explanation →

MCQhard

A data science team has trained a gradient boosting model using Amazon SageMaker to predict equipment failures. The model's confusion matrix shows 100 true negatives, 5 false positives, 20 false negatives, and 75 true positives. The cost of a false negative (missed failure) is $10,000, and the cost of a false positive (false alarm) is $500. What is the total cost of the model's predictions on this evaluation set?

A.$2,500

B.$20,500

C.$1,020,500

D.$202,500

AnswerD

Multiplying the 20 false negatives by the $10,000 missed-failure penalty gives $200,000, and the 5 false positives by the $500 false-alarm cost gives $2,500. Summing these weighted error costs yields $202,500, matching the stem's asymmetric cost constraint rather than treating all misclassifications equally.

Why this answer

The total cost is calculated by multiplying the number of false negatives (20) by their unit cost ($10,000) and the number of false positives (5) by their unit cost ($500), then summing these values: (20 × $10,000) + (5 × $500) = $200,000 + $2,500 = $202,500. This matches option D. The true negatives and true positives have no associated cost in this evaluation.

Exam trap

AWS certification exams often test the ability to correctly apply a cost matrix to a confusion matrix, where the trap is that candidates either forget to include all misclassification costs or mistakenly assign costs to correct predictions (true positives/negatives).

How to eliminate wrong answers

Option A is wrong because $2,500 only accounts for the cost of false positives (5 × $500) and completely ignores the much larger cost of false negatives (20 × $10,000 = $200,000). Option B is wrong because $20,500 incorrectly multiplies the number of false negatives by $1,000 instead of $10,000 (20 × $1,000 = $20,000) and adds the false positive cost ($500), revealing a unit cost error. Option C is wrong because $1,020,500 mistakenly includes the cost of true positives (75 × $10,000 = $750,000) and true negatives (100 × $500 = $50,000) in addition to the false negative and false positive costs, which is not part of the cost model.

Full explanation →

MCQhard

A media company uses Amazon Bedrock to generate short product descriptions. They notice that outputs vary in tone and sometimes include unwanted promotional claims. They want consistent, brand-aligned results while keeping the same foundation model. Which action best addresses this requirement?

A.Switch to a larger foundation model with more parameters to improve output quality and consistency.

B.Create a guardrail in Amazon Bedrock with a denied topics policy and a word filter for prohibited promotional terms.

C.Use a prompt template that specifies brand voice, required structure, and forbidden claims, and reuse it for every generation request.

D.Lower the top-p value to 0.1 so the model only considers the most probable tokens.

AnswerC

A well-crafted prompt template embeds the desired tone, format, and explicit exclusions directly into every request, giving the model clear instructions for consistent output. Reusing the template standardizes results across many generations without changing the model. This is the most direct way to enforce brand alignment for text style and claims.

Why this answer

Consistent brand-aligned generation is achieved by controlling the instructions the model receives. A reusable prompt template that defines voice, structure, and forbidden claims gives the model explicit guidance on every request, producing more uniform output. Guardrails, sampling parameters, and larger models can support quality but do not by themselves encode brand style and claim restrictions.

Exam trap

The trap here is treating guardrails or sampling settings as a substitute for clear prompt instructions, when style and claim control require explicit guidance in the prompt.

Full explanation →

MCQhard

A company is building a real-time document analysis tool using Amazon Bedrock. Their documents average 15,000 tokens each. Users submit a document and ask a single question about it. The team wants to minimize latency while maintaining answer quality. Which approach is MOST suitable?

A.Use Amazon Bedrock's Knowledge Base (RAG) to retrieve relevant chunks from the document and then answer the question.

B.Split the document into chunks of 1,000 tokens each, process each chunk with the model separately, and aggregate results.

C.Use Amazon Bedrock's Converse API with a prompt that includes the full document text and the user's question in a single invocation.

D.Use a fine-tuned model that has been trained on similar documents to avoid context processing.

AnswerC

A single Converse invocation passes the full 15,000-token document and question together, avoiding the extra retrieval round trip and embedding lookup that a RAG pipeline would add. With one question per document, prompt-stuffing satisfies the latency constraint directly while preserving answer quality.

Why this answer

Using Amazon Bedrock's Converse API with a prompt that includes the full document text and the user's question in a single invocation minimizes latency by avoiding the overhead of multiple API calls or retrieval steps. This approach maintains answer quality because the model has access to the entire document context (up to 15,000 tokens) in one pass, which is well within the context window of models like Claude 3 or Llama 2, ensuring accurate responses without the need for chunking or external retrieval.

Exam trap

The AIF-C01 exam often tests the misconception that RAG or chunking is always necessary for large documents, but the trap here is that 15,000 tokens is well within the context window of modern foundation models, making direct prompting the most latency-efficient and quality-preserving approach for single-document Q&A.

How to eliminate wrong answers

Option A is wrong because it suggests using a retrieval-augmented generation (RAG) approach with vector embeddings and a knowledge base, which introduces additional latency from embedding generation, vector search, and retrieval steps, making it unsuitable for real-time single-document analysis where the entire document fits in the model's context. Option B is wrong because it proposes splitting the document into chunks and processing each chunk separately with multiple Bedrock invocations, then aggregating results; this increases latency due to multiple API calls and risks losing cross-chunk context, degrading answer quality. Option D is wrong because it recommends using a smaller, faster model like Amazon Titan Text Lite, which may reduce latency but sacrifices answer quality due to its limited capacity to handle 15,000-token documents and complex question-answering tasks, leading to incomplete or inaccurate responses.

Full explanation →

MCQeasy

A developer runs this AWS CLI command to invoke a model in us-west-2 but receives an error: 'An error occurred (ModelNotFoundException) when calling the InvokeModel operation: Model not found'. What is the most likely cause?

A.The request body is not properly formatted

B.The --region parameter is missing from the command

C.The model is not available in the us-west-2 region

D.The user's IAM role lacks permissions to invoke the model

AnswerC

Model availability is region-specific in Amazon Bedrock; model IDs valid in one region return ModelNotFoundException elsewhere. Since the command targets us-west-2, the requested model simply is not offered there, so invoking it fails regardless of credentials or permissions.

Why this answer

The error 'ModelNotFoundException' specifically indicates that the model ID or ARN is not recognized in the specified region. AWS Bedrock models are region-specific; not all foundation models are available in every AWS region. The developer invoked the model in us-west-2, but the requested model may only be available in regions like us-east-1 or us-west-1, causing the service to return a ModelNotFoundException rather than a permissions or formatting error.

Exam trap

Candidates often confuse the ModelNotFoundException with permission errors (AccessDeniedException). In AWS Bedrock, this error indicates the model is not available in the specified region, not a lack of IAM permissions. Always check regional model availability before assuming a permissions issue.

How to eliminate wrong answers

Option A is wrong because a malformed request body would result in a ValidationException or MalformedRequestBody error, not a ModelNotFoundException. Option B is wrong because if the --region parameter were missing, the CLI would use the default region from the AWS config or environment variables; if no default region were set, the CLI would return a 'You must specify a region' error, not a ModelNotFoundException. Option D is wrong because insufficient IAM permissions would produce an AccessDeniedException or UnauthorizedOperation error, not a ModelNotFoundException.

Full explanation →

MCQmedium

A financial services company wants to deploy a chatbot using Amazon Bedrock that must never discuss investment advice. Which Bedrock feature should they configure to enforce this policy?

A.Grounding check

B.PII detection

C.Topic denial

D.Content filtering

AnswerC

Topic denial prevents the model from discussing defined topics like investment advice.

Why this answer

Amazon Bedrock Guardrails lets administrators define denied topics (for example, 'investment advice') that the model must not discuss. When a user prompt or model response is detected as related to a denied topic, Guardrails blocks the response and can return a predefined message. This is enforced through Bedrock Guardrails denied topics, not a standalone 'Topic denial' feature.

Exam trap

AWS AI Practitioner exams often test the distinction between built-in moderation capabilities (content filtering, PII detection) and custom business policy enforcement using Bedrock Guardrails denied topics. Candidates may mistakenly choose Content filtering because it sounds like it blocks unwanted content, but custom prohibited-topic rules are handled through Guardrails denied topics (the intended meaning of the 'Topic denial' option).

How to eliminate wrong answers

Option A is wrong because Grounding check verifies that the model's response is based on a provided source document (e.g., a company policy PDF) but does not actively block specific topics; it only checks for factual consistency. Option B is wrong because PII detection is designed to identify and redact personally identifiable information (e.g., SSNs, credit card numbers) from inputs or outputs, not to enforce content policies about prohibited topics. Option D is wrong because Content filtering applies to harmful or offensive content (e.g., hate speech, violence) based on configurable thresholds, not to domain-specific business policies like 'no investment advice'.

Full explanation →

Multi-Selecteasy

A data scientist wants to deploy a custom model built with TensorFlow to Amazon SageMaker for real-time inference. Which TWO steps are required? (Choose two.)

Select 2 answers

A.Create an Amazon ECR repository for the inference container

B.Upload the model artifacts to an S3 bucket

C.Submit a training job to SageMaker

D.Create a SageMaker endpoint configuration

E.Convert the model to ONNX format

AnswersB, D

SageMaker real-time endpoints pull model artefacts from Amazon S3, so the trained TensorFlow files must reside in a bucket the execution role can read. This satisfies the stem's deployment requirement: without artefacts in S3, CreateModel cannot locate the model data, and endpoint creation fails.

Why this answer

Option B is correct because SageMaker requires model artifacts (the trained TensorFlow SavedModel or model.tar.gz) to be stored in an Amazon S3 bucket, which is then referenced by the SageMaker model when deploying. Option D is correct because deploying to a real-time endpoint requires creating an endpoint configuration that specifies the production variant, instance type, and initial instance count, which is then used to create the endpoint. Option A is not required because the TensorFlow inference container is already provided and maintained by SageMaker as a prebuilt Docker image, so no custom ECR repository is needed unless using a custom container.

Option C is not required because the model is already built with TensorFlow; a SageMaker training job is only needed if training within SageMaker. Option E is not required because SageMaker's TensorFlow container supports native TensorFlow SavedModel format, so ONNX conversion is unnecessary.

Exam trap

The trap here is that candidates often think they must build a custom container (Option A) or convert the model (Option E), but SageMaker's pre-built TensorFlow containers eliminate those steps, and the key requirements are simply uploading artifacts to S3 and creating the endpoint configuration.

Full explanation →

MCQmedium

A developer is using the Amazon Bedrock InvokeModel API with the above request to summarize meeting notes. The response is a single word repeated many times. Which parameter is MOST likely causing this issue?

A.topP set to 0.9

B.stopSequences is empty

C.maxTokenCount set to 100

D.temperature set to 0

AnswerD

Temperature at 0 makes the model greedy, always picking the highest-probability token, so a repetitive loop can dominate the output. It satisfies the stem's constraint by explaining the degenerate single-word repetition, though top-p or top-k sampling would restore diversity.

Why this answer

A temperature of 0 forces the model to always select the highest-probability token at each step, which can lead to repetitive loops if the most likely token repeatedly points back to itself (e.g., the same word). This deterministic behavior eliminates randomness, causing the model to get stuck in a single-word cycle rather than generating diverse or coherent text.

Exam trap

AWS often tests the misconception that temperature only affects 'creativity' or 'randomness,' when in fact a temperature of 0 causes deterministic argmax selection, which can paradoxically produce repetitive or stuck outputs rather than simply 'less creative' text.

How to eliminate wrong answers

Option A is wrong because topP set to 0.9 (nucleus sampling) actually increases diversity by considering tokens whose cumulative probability reaches 0.9, which would reduce repetition, not cause it. Option B is wrong because an empty stopSequences list means no custom stopping conditions are applied, but this does not force repetition; the model would still generate until a natural stop (e.g., EOS token) or maxTokenCount is reached. Option C is wrong because maxTokenCount set to 100 only limits the total number of tokens generated; it does not influence token selection probability or cause a single word to repeat—it would simply stop after 100 tokens regardless of content.

Full explanation →

100

MCQeasy

Which of the following is a key advantage of using a diffusion model for image generation compared to a GAN?

A.Diffusion models produce more diverse and higher-quality images with stable training

B.Diffusion models generate images faster than GANs during inference

C.Diffusion models are inherently conditional and do not require labels

D.Diffusion models require less training data than GANs

AnswerA

Diffusion models iteratively denoise random noise, avoiding the adversarial min-max game that destabilises GAN training. This yields greater sample diversity and higher fidelity, directly satisfying the stem's demand for a key advantage over GANs. Mode collapse, a common GAN failure, is largely absent, giving the stable training the option claims.

Why this answer

Diffusion models offer a key advantage over GANs because they are trained with a stable, non-adversarial objective—denoising score matching—which avoids the mode collapse and training instability common in GANs. This leads to more diverse outputs and, with sufficient steps, higher-quality images that can rival or exceed GANs, especially in large-scale text-to-image tasks.

Exam trap

AWS often tests the misconception that diffusion models are faster than GANs because they are newer or more advanced, but the trap is that their iterative sampling process makes them significantly slower at inference time.

How to eliminate wrong answers

Option B is wrong because diffusion models are inherently slower during inference; they require many iterative denoising steps (e.g., 50–1000 steps) to generate an image, whereas GANs produce an image in a single forward pass. Option C is wrong because diffusion models are not inherently conditional; they require explicit conditioning mechanisms (e.g., cross-attention to text embeddings) and labels or prompts to guide generation, whereas GANs can also be conditioned but are not inherently so. Option D is wrong because diffusion models typically require large amounts of training data to learn the underlying data distribution effectively, often more than GANs, and do not have a data efficiency advantage.

Full explanation →

101

MCQhard

A data scientist runs the SageMaker Clarify job shown in the exhibit for a credit risk model. After reviewing the results, they find a high bias metric for the gender facet. Which action is most consistent with responsible AI?

A.Proceed with deployment because the model is already in production

B.Remove the gender attribute from the training data and retrain

C.Investigate the root cause and retrain with balanced data

D.Increase the acceptance threshold for the model

AnswerC

SageMaker Clarify's bias metric indicates disparate impact across gender, so investigating the root cause and retraining with balanced data addresses the underlying skew. This satisfies responsible AI by remediating the model rather than merely documenting or ignoring the measured bias.

Why this answer

Responsible AI requires understanding and mitigating bias at its source, not just masking it. Investigating the root cause (e.g., data collection bias, labeling bias, or proxy features) and retraining with balanced data directly addresses the high bias metric detected by SageMaker Clarify, aligning with AWS's principle of fairness. Simply removing the gender attribute may not eliminate bias if other features act as proxies, and increasing the threshold does not fix the underlying model bias.

Exam trap

The AIF-C01 exam often tests the misconception that simply removing a sensitive attribute (like gender) is sufficient to eliminate bias, but the trap here is that proxy features can still encode the same bias, making root-cause investigation and balanced retraining the only responsible action.

How to eliminate wrong answers

Option A is wrong because deploying a model with a known high bias metric violates responsible AI principles and could lead to unfair outcomes, even if the model is already in production; SageMaker Clarify is designed to detect such issues before or during deployment. Option B is wrong because removing the gender attribute alone does not guarantee bias removal—other features like zip code or income can act as proxies for gender, and the model may still learn biased correlations. Option D is wrong because increasing the acceptance threshold (e.g., for a binary classifier) only changes the decision boundary, not the underlying biased patterns learned by the model; it does not reduce the bias metric reported by Clarify.

Full explanation →

102

Multi-Selecthard

A data science team is using Amazon Bedrock to generate synthetic data for training a new model. They need to ensure the generated data is diverse and covers edge cases. Which THREE parameters should they adjust to maximize diversity? (Select THREE.)

Select 3 answers

A.Increase the top-p value to 0.9

B.Increase the top-k value to 50

C.Decrease the top-p value to 0.1

D.Set the seed to a fixed value

E.Increase the temperature

AnswersA, B, E

Raising top-p to 0.9 widens nucleus sampling, so the model draws from a larger cumulative probability mass rather than the most likely tokens alone. This directly satisfies the diversity and edge-case coverage requirement by admitting lower-probability token choices during synthetic data generation.

Why this answer

Option E (increase the temperature) is correct because temperature scales the logits before softmax, flattening the probability distribution so lower-probability tokens are sampled more often, which directly increases output diversity and helps surface edge cases. Option A (increase top-p to 0.9) is correct because top-p (nucleus) sampling retains the smallest set of tokens whose cumulative probability reaches 0.9, keeping a broad candidate pool instead of truncating to only the most likely tokens. Option B (increase top-k to 50) is correct because top-k sampling restricts sampling to the 50 highest-probability tokens, and a larger k widens that pool, allowing more varied token choices than a small k.

Option C (decrease top-p to 0.1) is not correct because a narrow nucleus of only the top ~10% cumulative probability makes generation more deterministic and less diverse. Option D (set a fixed seed) is not correct because a fixed seed makes sampling reproducible, not more diverse, and can actually reduce variation across runs.

Exam trap

AWS often tests the misconception that decreasing top-p or fixing the seed increases diversity, when in fact both actions reduce randomness and limit the model's ability to generate varied outputs.

Full explanation →

103

MCQeasy

A social media company uses Amazon Comprehend to moderate user comments. They want to avoid censoring legitimate speech while catching hate speech. Which approach aligns with responsible AI governance?

A.Implement a human-in-the-loop review for borderline cases

B.Use multiple models and average their scores

C.Use a single model with high confidence threshold

D.Rely solely on automated filtering

AnswerA

Human-in-the-loop review routes borderline confidence scores to a person, so ambiguous comments are judged contextually rather than auto-removed. This satisfies the stem's constraint of avoiding censorship of legitimate speech while still catching hate speech, and it keeps accountability with a human decision-maker.

Why this answer

A human-in-the-loop (HITL) review for borderline cases aligns with responsible AI governance by balancing automated detection with human judgment. Amazon Comprehend can flag comments with moderate confidence scores (e.g., 0.5–0.9) for manual review, ensuring that ambiguous or context-dependent hate speech is not censored while still catching clear violations. This approach mitigates false positives and respects free expression, which is a core tenet of responsible AI.

Exam trap

The AWS AI Practitioner exam often tests the misconception that higher confidence thresholds or multiple models alone are sufficient for responsible AI, when in fact human oversight is required to handle edge cases and ensure ethical outcomes.

How to eliminate wrong answers

Option B is wrong because averaging scores from multiple models does not inherently address the nuance of borderline cases; it may still produce a false positive or negative if all models share similar biases or training data, and it lacks the contextual understanding that human review provides. Option C is wrong because using a single model with a high confidence threshold (e.g., >0.95) will reduce false positives but will also miss many true instances of hate speech that fall below the threshold, leading to under-censorship and failing to catch subtle or coded hate speech. Option D is wrong because relying solely on automated filtering ignores the need for human oversight in ambiguous cases, which can lead to over-censorship of legitimate speech or failure to detect nuanced hate speech, violating responsible AI principles of fairness and accountability.

Full explanation →

104

Multi-Selecthard

A media company is evaluating foundation models for a generative AI application that produces image captions and short video summaries. The team must balance output quality, latency, and operational cost. Which TWO considerations are most important when selecting a foundation model for this multimodal task? (Choose two.)

Select 2 answers

A.Whether the model can generate outputs in multiple human languages simultaneously.

B.Whether the model was trained using a specific programming language's libraries.

C.Whether the model supports the required input modalities, such as images and video frames.

D.Whether the model's inference latency and cost align with the application's throughput requirements.

E.Whether the model's license permits the company's intended commercial distribution of outputs.

AnswersC, D

The task requires processing images and video frames, so the model must accept those modalities as input. A text-only model cannot caption or summarize visual content regardless of its quality. Verifying modality support is therefore a fundamental selection criterion, ensuring the chosen model can actually ingest the required data types before any other trade-off is evaluated.

Why this answer

For a multimodal captioning and summarization workload, the model must accept images and video frames as input, so modality support is a prerequisite. Because the team must also balance latency and cost, evaluating inference speed and pricing against throughput needs is equally critical. Together these ensure the model can handle the data and remain practical to operate at scale.

Exam trap

The trap here is gravitating toward tangential factors like licensing or language coverage while overlooking that modality support and performance economics are the decisive selection criteria.

Full explanation →

105

MCQhard

A media company generates AI-written summaries of news articles using Amazon Bedrock and publishes them automatically. Legal counsel is concerned that the model might reproduce long verbatim passages from copyrighted source articles. The team wants a configurable safeguard that detects and filters responses containing text closely matching the source documents before publication. Which approach should they implement?

A.Configure a sensitive information filter in Amazon Bedrock Guardrails to block personally identifiable information

B.Implement a plagiarism or similarity detection step that compares generated text against the source corpus before publishing

C.Apply automated reasoning checks in Amazon Bedrock Guardrails to validate the summary against policy rules

D.Use a word filter in Amazon Bedrock Guardrails configured with the publisher's restricted terms

AnswerB

Detecting verbatim copying requires comparing generated output against the original documents using similarity measures such as n-gram overlap or embedding distance. Building this comparison step into the publishing pipeline directly targets the legal concern and allows a configurable threshold for blocking or rewriting flagged summaries. None of the guardrail content filters evaluate text-to-source similarity in this way.

Why this answer

Copyright overlap is a text-similarity problem, so the safeguard must compare generated summaries against the source article corpus and flag passages exceeding a similarity threshold. PII filters, word filters, and policy-based automated reasoning checks each address different risk categories and cannot measure verbatim reproduction of ordinary prose from a reference document.

Exam trap

The trap here is reaching for a Bedrock Guardrails filter by habit, when the actual risk is textual overlap with source documents, which requires similarity comparison rather than pattern or policy matching.

Full explanation →

106

MCQmedium

A company wants to personalize its generative AI model for its specific domain without sharing data with third-party model providers. Which method should they use?

A.Fine-tuning the foundation model on their proprietary data

B.Prompt engineering with domain-specific examples

C.Retrieval-augmented generation (RAG) with a domain-specific knowledge base

D.Model distillation using a larger foundation model

AnswerA

Fine-tuning adjusts a foundation model's weights using the company's proprietary domain data, and the resulting custom model remains within the organisation's Azure AI resource. No training data is shared with the third-party provider, satisfying the data-confidentiality constraint in the stem.

Why this answer

Fine-tuning the foundation model on proprietary data allows the company to adapt the model's weights to its specific domain without sharing data with third parties. This method trains the model on private datasets, enabling it to learn domain-specific patterns and terminology while keeping data in-house, which is critical for data privacy and compliance.

Exam trap

AWS often tests the distinction between methods that modify model parameters (fine-tuning) versus those that only augment input or retrieval (prompt engineering, RAG), leading candidates to mistakenly choose RAG for personalization when fine-tuning is required for deep domain adaptation.

How to eliminate wrong answers

Option B is wrong because prompt engineering with domain-specific examples does not modify the model's weights; it only influences output through input prompts, which cannot achieve the same depth of domain adaptation as fine-tuning and still relies on the base model's knowledge. Option C is wrong because retrieval-augmented generation (RAG) with a domain-specific knowledge base retrieves external information at inference time but does not train the model on proprietary data, so the model itself remains unchanged and may not fully internalize domain nuances. Option D is wrong because model distillation compresses a larger model into a smaller one for efficiency, but it does not involve training on proprietary domain data and does not address the requirement of personalization without sharing data.

Full explanation →

107

MCQhard

A financial services company uses Amazon Bedrock to power an internal assistant that answers employee questions about HR policies. The company must ensure that the assistant never reveals sensitive employee data. The HR policy documents are stored in an Amazon S3 bucket and are updated frequently. The company wants the assistant to cite the exact policy document and section for each answer. Which solution meets these requirements with the LEAST operational overhead?

A.Build a custom retrieval system using Amazon OpenSearch Service and write a Lambda function to call the model with retrieved passages.

B.Use the model's built-in knowledge by prompting it to answer from its training data, and add a guardrail to filter sensitive information.

C.Fine-tune a foundation model on the HR policy documents and deploy it with a guardrail that blocks sensitive data.

D.Create an Amazon Bedrock knowledge base backed by the S3 bucket, and use the RetrieveAndGenerate API with citations enabled.

AnswerD

Amazon Bedrock knowledge bases can be connected to an S3 bucket, automatically chunk and index documents, and support frequent updates. The RetrieveAndGenerate API can return generated answers with citations that point to the exact source documents and sections. This approach minimizes operational overhead because Bedrock manages the vector store and retrieval pipeline, and it meets both the citation and data freshness requirements.

Why this answer

The company needs accurate, cited answers from frequently updated HR documents with minimal operational overhead. Amazon Bedrock knowledge bases integrate with S3, automatically manage indexing and retrieval, and support citations via the RetrieveAndGenerate API. This managed solution avoids custom infrastructure and ensures answers are grounded in the latest policies.

Fine-tuning, custom retrieval, or relying on model training data all fail to meet the citation and maintenance requirements efficiently.

Exam trap

The trap here is assuming that fine-tuning a model on internal documents will provide citations, when in fact fine-tuning does not produce source references and requires retraining for updates.

Full explanation →

108

Multi-Selecteasy

A developer wants to quickly prototype a generative AI application using Amazon Bedrock. They need to test different foundation models and prompts interactively. Which two services or features are designed for this purpose? (Choose TWO.)

Select 2 answers

A.Amazon Bedrock Playground

B.Amazon CloudWatch

C.AWS Lambda

D.Amazon Bedrock Studio

E.Amazon SageMaker Studio

AnswersA, D

Amazon Bedrock Playground provides a console interface for interactively selecting foundation models, adjusting parameters and testing prompts without writing code. This directly satisfies the stem's requirement to prototype quickly and compare different models and prompts interactively.

Why this answer

Amazon Bedrock Playground (A) is a console-based interactive environment specifically designed for experimenting with different foundation models and prompts without writing code, making it ideal for quick prototyping and model comparison. Amazon Bedrock Studio (D) is a web-based workspace that provides a collaborative, interactive interface for building and testing generative AI applications with Bedrock models, prompts, and knowledge bases, which directly supports rapid prototyping. Amazon CloudWatch (B) is a monitoring and observability service for logs, metrics, and alarms, not an interactive model/prompt testing tool.

AWS Lambda (C) is a serverless compute service for running code in response to events, not a purpose-built Bedrock prototyping interface. Amazon SageMaker Studio (E) is an integrated development environment for the broader machine learning lifecycle (data prep, training, tuning, deployment), not the interactive Bedrock model/prompt testing feature described here.

Exam trap

AWS often tests the distinction between general-purpose ML tools (SageMaker Studio) and Bedrock-specific prototyping features (Playground and Studio), leading candidates to mistakenly choose SageMaker Studio because it sounds similar to 'studio' in Bedrock Studio.

Full explanation →

109

Multi-Selecthard

A company is fine-tuning an Amazon Titan Text model on custom data using Amazon Bedrock. They want to ensure the fine-tuned model retains general language capabilities while learning domain-specific knowledge. Which THREE best practices should they follow? (Select THREE.)

Select 3 answers

A.Freeze the first few layers of the model to prevent overfitting

B.Train only on domain-specific data to maximize accuracy

C.Monitor validation loss to detect overfitting and stop training early if needed

D.Use a low learning rate to avoid catastrophic forgetting

E.Use a diverse dataset that includes both general and domain-specific examples

AnswersC, D, E

Monitoring validation loss detects divergence between training and held-out data, triggering early stopping before the model memorises domain samples and degrades general language capabilities. This directly satisfies the stem's constraint of retaining general ability while acquiring domain knowledge during Amazon Bedrock fine-tuning.

Why this answer

Option C is correct because monitoring validation loss during fine-tuning lets you detect when the model begins to overfit the domain data and apply early stopping, preserving generalization to broader language tasks. Option D is correct because using a low learning rate makes smaller weight updates, which reduces the risk of catastrophic forgetting of the model's pre-trained general language capabilities while it absorbs domain knowledge. Option E is correct because mixing general and domain-specific examples in the training dataset helps the fine-tuned model retain broad language understanding alongside the new specialized knowledge.

Option A is not appropriate here because freezing the first few layers is a parameter-efficient technique aimed at reducing overfitting and compute, not a stated best practice for preserving general capabilities in Amazon Bedrock Titan fine-tuning. Option B is incorrect because training only on domain-specific data maximizes specialization at the expense of general language ability, which is exactly the outcome the company wants to avoid.

Exam trap

A common misconception in fine-tuning LLMs is that freezing layers or using only domain-specific data is effective, when in practice, low learning rates and diverse datasets are required to prevent catastrophic forgetting and overfitting.

Full explanation →

110

MCQeasy

A startup wants to build a product recommendation engine for their e-commerce platform. They have user purchase history and item metadata. They want a fully managed solution that can automatically train and deploy a recommendation model without needing to manage the underlying ML lifecycle. The solution should provide personalized recommendations based on collaborative filtering. Which AWS service should they use?

A.Use Amazon Kendra

B.Use Amazon Lex

C.Use Amazon Personalize

D.Use Amazon SageMaker built-in Factorization Machines algorithm

AnswerC

Amazon Personalize is fully managed, handling training, tuning and deployment of collaborative-filtering recommenders without ML lifecycle management. It ingests purchase history and item metadata to produce personalised recommendations, meeting the startup's requirement for a managed solution requiring no underlying ML operations.

Why this answer

Amazon Personalize is a fully managed service that enables you to build and deploy recommendation models without managing the underlying ML lifecycle. It supports collaborative filtering out of the box, using user purchase history and item metadata to generate personalized recommendations, which directly matches the startup's requirements.

Exam trap

The trap here is that candidates may confuse Amazon SageMaker's built-in algorithms (like Factorization Machines) with a fully managed recommendation service, overlooking the requirement for automatic lifecycle management and instead focusing only on the algorithm capability.

How to eliminate wrong answers

Option A is wrong because Amazon Kendra is an intelligent search service that uses natural language processing to answer questions and retrieve documents, not a recommendation engine for collaborative filtering. Option B is wrong because Amazon Lex is a service for building conversational interfaces (chatbots) using speech and text, not for generating product recommendations. Option D is wrong because while Amazon SageMaker's built-in Factorization Machines algorithm can perform collaborative filtering, it requires you to manage the ML lifecycle (data preparation, training, deployment, scaling), which contradicts the requirement for a fully managed solution that automatically handles these tasks.

Full explanation →

111

Multi-Selectmedium

A financial institution is developing a model to detect fraudulent transactions. They want to ensure the model is robust and does not exhibit bias. Which TWO actions should they take?

Select 2 answers

A.Use Amazon SageMaker Clarify to compute fairness metrics like demographic parity

B.Use a single model for all customers without segmenting by region

C.Deploy the model with a high confidence threshold to reduce false positives

D.Encrypt all transaction data at rest and in transit

E.Collect a balanced training dataset representing all transaction types and customer demographics

AnswersA, E

SageMaker Clarify quantifies bias by computing fairness metrics such as demographic parity across groups, exposing disparities the institution could otherwise miss. Measuring these metrics lets the team detect and remediate bias before deployment, satisfying the robustness requirement.

Why this answer

Option A is correct because Amazon SageMaker Clarify is the AWS service specifically designed to detect bias in data and models, and it computes fairness metrics such as demographic parity, disparate impact, and equal opportunity across sensitive groups, directly addressing the requirement to ensure the model does not exhibit bias. Option E is correct because a balanced training dataset that represents all transaction types and customer demographics reduces sampling bias and prevents the model from overfitting to majority groups, which is a foundational step for building a robust and fair fraud-detection model. Option B is incorrect because using a single unsegmented model for all customers can mask region-specific fraud patterns and amplify bias, rather than mitigate it.

Option C is incorrect because raising the confidence threshold only tunes precision/recall trade-offs for false positives; it does nothing to detect or reduce bias. Option D is incorrect because encryption at rest and in transit is a security control for data protection and compliance, not a fairness or bias-mitigation measure.

Exam trap

The trap is selecting security or threshold-tuning options (C, D) that sound responsible but do not address bias — candidates must distinguish fairness actions from general best practices.

Full explanation →

112

Multi-Selectmedium

Which THREE practices are recommended for promoting robustness and security in AI systems?

Select 3 answers

A.Deploy the model immediately after training without validation

B.Implement strong access controls and encryption for model artifacts

C.Regularly test the model against adversarial examples

D.Monitor model performance for data drift and concept drift

E.Remove logging and monitoring to improve performance

AnswersB, C, D

Security controls protect models from unauthorized access and tampering.

Why this answer

Robustness and security in AI systems require multiple complementary practices. (B) Implementing strong access controls (e.g., IAM policies, role-based access control) and encryption (e.g., AES-256 for data at rest, TLS 1.2+ for data in transit) protects model artifacts from unauthorized access, tampering, and exfiltration, ensuring confidentiality and integrity throughout the lifecycle. (C) Regularly testing the model against adversarial examples surfaces vulnerabilities to evasion, poisoning, and prompt-injection style attacks, allowing defenses to be hardened before real-world exploitation. (D) Monitoring model performance for data drift and concept drift detects when the model's inputs or the underlying relationships change, so degradation, unexpected behavior, or emerging security-relevant anomalies can be caught and remediated early. Together these practices cover artifact protection, adversarial resilience, and ongoing operational vigilance.

Exam trap

Candidates often mistakenly think that immediate deployment or removing monitoring can improve performance, but these actions severely compromise robustness and security by skipping validation and eliminating visibility into model degradation.

Full explanation →

113

MCQmedium

A city government uses an Amazon SageMaker model to score affordable-housing applications. To meet its responsible AI commitments, the IT team must ensure that every automated decision can be traced back to the exact model version and training dataset used, and that changes are reviewed before deployment. Which combination of AWS practices best provides this accountability?

A.Register model versions in the SageMaker Model Registry with approval status and lineage, and require manual approval before deployment

B.Configure a SageMaker endpoint with a larger instance type to reduce inference latency for applicants

C.Store application data in an Amazon S3 bucket encrypted with AWS KMS customer managed keys

D.Enable automatic scaling on the SageMaker endpoint so it can handle fluctuating application volumes

AnswerA

The SageMaker Model Registry stores versioned model packages with metadata, approval status, and lineage to the training job and dataset. Gating deployment on an approved status creates a documented review step and a traceable record linking each decision to a specific model version and its training data, which is exactly the accountability the city requires.

Why this answer

Accountability for automated decisions requires a durable record of what produced each outcome and a controlled release process. The SageMaker Model Registry captures versioned model packages with approval status and lineage to training data, and its approval workflow enforces review before deployment. Scaling, instance sizing, and encryption improve performance or confidentiality but do not deliver traceability or approval gating.

Exam trap

The trap here is equating any governance-adjacent control, such as encryption or scaling, with accountability, when only versioned lineage plus an approval gate provides traceability and review.

Full explanation →

114

MCQmedium

A social media company deploys a content moderation model. They want to minimize the risk of over-censoring legitimate posts (false positives) while still catching harmful content. Which metric should they prioritize?

A.F1 score

B.Recall

C.Accuracy

D.Precision

AnswerD

Precision measures the proportion of predicted harmful posts that are genuinely harmful, so maximising it directly reduces false positives, satisfying the stem's goal of avoiding over-censoring legitimate content. Recall would instead prioritise catching all harmful posts.

Why this answer

Precision measures the proportion of positive identifications that are actually correct. High precision means fewer false positives, which aligns with the goal of not over-censoring legitimate posts.

Full explanation →

115

Multi-Selectmedium

A company is deploying a large language model (LLM) for customer support. They want to reduce the risk of hallucinations. Which TWO approaches should they implement? (Choose two.)

Select 2 answers

A.Implement Amazon Bedrock Guardrails to define topics the model should avoid

B.Use a larger model without any retrieval mechanism

C.Fine-tune the model on a dataset of hallucinated examples

D.Increase the maximum token length to give the model more room to elaborate

E.Use Retrieval-Augmented Generation (RAG) to provide factual context

AnswersA, E

Guardrails apply configurable content filters and topic-deny policies that block the model from generating responses outside approved subject areas. This constrains outputs at inference time, reducing the chance of fabricated or off-topic answers, which addresses the stated hallucination-reduction requirement.

Why this answer

Option A is correct because Amazon Bedrock Guardrails lets you define denied topics, content filters, and contextual grounding checks that constrain what the model can say, directly reducing the chance it produces off-topic or fabricated responses in a customer support setting. Option E is correct because Retrieval-Augmented Generation (RAG) retrieves authoritative documents from a knowledge base and injects that factual context into the prompt, grounding the model's answers in real data rather than relying solely on parametric memory, which is the standard mitigation for hallucinations. Option B is wrong because a larger model without retrieval still generates from learned parameters and can hallucinate confidently; scale alone does not guarantee factual accuracy.

Option C is wrong because fine-tuning on hallucinated examples would teach the model to reproduce fabrications, worsening the problem. Option D is wrong because increasing max token length only allows longer outputs and does not improve factual grounding; it can even give the model more room to elaborate on incorrect claims.

Exam trap

The trap is assuming that simply increasing model size or token limit will reduce hallucinations, when in fact these can exacerbate the problem. Candidates may also overlook that fine-tuning on bad examples worsens the issue.

Full explanation →

116

MCQmedium

A data scientist is training a binary classification model to predict customer churn. The dataset has 10,000 records with 9,500 non-churners and 500 churners. After training a logistic regression model, the model achieves 95% accuracy on the test set. However, the business team reports that the model is not useful because it predicts almost all customers as non-churners. Which metric should the data scientist use to evaluate the model's performance in this scenario?

A.Accuracy

B.R-squared

C.Precision

D.Recall

AnswerD

Recall measures the proportion of actual churners correctly identified, directly exposing the model's failure on the 500-positive minority class. With 95% accuracy achievable by predicting all non-churners, recall reveals that sensitivity to churn is near zero, satisfying the need for a metric robust to this class imbalance.

Why this answer

(Recall) is correct because in this highly imbalanced dataset (95% non-churners vs 5% churners), the model's 95% accuracy is misleading—it can achieve this by simply predicting the majority class (non-churner) for all samples. Recall measures the proportion of actual churners correctly identified (True Positives / (True Positives + False Negatives)), directly addressing the business need to detect churn. A high recall ensures the model captures most churners, even at the cost of some false positives.

Exam trap

The AIF-C01 exam often tests the misconception that high accuracy always indicates a good model, especially in imbalanced datasets, leading candidates to overlook metrics like recall or precision that better reflect model utility for the specific business problem.

How to eliminate wrong answers

Option A is wrong because accuracy is a poor metric for imbalanced datasets; a model that predicts all samples as the majority class can achieve high accuracy (95% here) while failing to identify any churners, making it useless for the business goal. Option B is wrong because R-squared is a metric for regression models, measuring the proportion of variance explained by the independent variables, and is not applicable to binary classification tasks like churn prediction. Option C is wrong because precision (True Positives / (True Positives + False Positives)) focuses on the correctness of positive predictions; while important, it does not capture the model's ability to find all churners—a model with high precision but low recall might still miss most churners, which is the core issue reported by the business team.

Full explanation →

117

MCQmedium

A developer is using Amazon Bedrock to generate product descriptions. The developer notices that the model sometimes outputs descriptions that contradict the provided product specifications. Which parameter adjustment would MOST directly reduce factual inconsistencies?

A.Increase maxTokens to allow longer descriptions

B.Decrease temperature to a value close to 0

C.Increase topP to 1.0

D.Set topK to a higher value

AnswerB

Lowering temperature sharpens the token probability distribution, so the model samples near-deterministic, highest-likelihood continuations rather than exploring alternatives. This directly constrains the randomness that produces specification-contradicting output, satisfying the stem's requirement to reduce factual inconsistency most directly.

Why this answer

Decreasing temperature to a value close to 0 makes the model more deterministic and less creative, which reduces the likelihood of generating random or contradictory content. In Amazon Bedrock, temperature controls the randomness of token selection; lower values cause the model to choose the most probable tokens, aligning outputs more closely with the provided product specifications and minimizing factual inconsistencies.

Exam trap

AWS often tests the misconception that increasing output length (maxTokens) or expanding token selection (topP, topK) improves accuracy, when in fact these parameters increase variability and the risk of factual errors, whereas lowering temperature is the direct control for reducing randomness.

How to eliminate wrong answers

Option A is wrong because increasing maxTokens only extends the maximum length of the generated text, which does not address the randomness or creativity that leads to contradictions; it may even allow more room for errors. Option C is wrong because increasing topP to 1.0 includes all possible tokens up to the cumulative probability threshold, which actually increases diversity and can worsen factual inconsistencies by allowing less probable tokens. Option D is wrong because setting topK to a higher value expands the pool of candidate tokens considered, increasing randomness and the chance of generating contradictory content, rather than reducing it.

Full explanation →

118

Multi-Selectmedium

A data scientist is building a regression model to predict house prices. During feature engineering, they have categorical variables (e.g., neighborhood) and numerical variables (e.g., square footage) with missing values. Which TWO actions should the data scientist take? (Choose two.)

Select 2 answers

A.Drop all rows with any missing values

B.Normalize all numerical features to [0,1] range

C.Impute missing square footage values with the median

D.Apply label encoding to the neighborhood variable

E.Apply one-hot encoding to the neighborhood variable

AnswersC, E

Square footage is numerical and skewed by outliers, so median imputation preserves the central tendency without being distorted by extreme values, unlike the mean. This keeps the regression model's feature distribution stable while retaining rows that would otherwise be dropped.

Why this answer

Option C is correct because imputing missing square footage values with the median is a robust strategy for numerical features, especially when outliers may be present, and it preserves the row count for model training. Option E is correct because one-hot encoding is the appropriate technique for nominal categorical variables like neighborhood, as it creates binary columns without imposing an artificial ordinal relationship. Option A is not ideal because dropping all rows with any missing values can discard substantial useful data and introduce bias.

Option B is not required for regression models and may not be necessary unless the algorithm is sensitive to feature scale. Option D is incorrect because label encoding would impose an arbitrary order on neighborhood categories, which is inappropriate for a nominal variable.

Exam trap

In the AWS AI Practitioner exam, a common trap is that candidates may incorrectly choose label encoding (Option D) because it seems simpler, without recognizing that it introduces an arbitrary ordinal relationship that degrades model performance.

Full explanation →

119

MCQhard

A generative AI application occasionally produces factually incorrect responses. The team has already tried prompt engineering and increasing the temperature parameter. Which next step is MOST effective to improve factual accuracy?

A.Use a larger foundation model

B.Fine-tune the model on company data

C.Reduce the temperature to 0

D.Implement a Retrieval Augmented Generation (RAG) pipeline

AnswerD

RAG retrieves relevant documents from a knowledge base and injects them into the prompt, grounding generation in verifiable source content. Prompt engineering and temperature tuning cannot supply missing facts, so retrieval directly targets factual accuracy.

Why this answer

Retrieval Augmented Generation (RAG) is the most effective next step because it grounds the model's responses in a verified external knowledge base, directly addressing factual inaccuracies without requiring retraining. Unlike prompt engineering or temperature adjustments, RAG provides real-time access to authoritative documents, reducing hallucinations by constraining the model's output to retrieved evidence.

Exam trap

AWS often tests the misconception that reducing temperature or using a larger model directly fixes factual accuracy, when in fact these methods address output randomness and capacity, not the root cause of hallucination, which is lack of grounded knowledge.

How to eliminate wrong answers

Option A is wrong because using a larger foundation model does not inherently improve factual accuracy; larger models can still hallucinate and may even produce more confident but incorrect answers. Option B is wrong because fine-tuning on company data improves domain-specific performance but does not guarantee factual correctness for dynamic or external facts, and it requires substantial computational resources and labeled data. Option C is wrong because reducing temperature to 0 makes the model deterministic but does not prevent it from generating plausible-sounding but factually incorrect responses; it only reduces randomness, not hallucinations.

Full explanation →

120

MCQhard

A healthcare company needs to use Amazon SageMaker Ground Truth for data labeling. The data includes protected health information (PHI) that must remain in the US. Which configuration meets the compliance requirements?

A.Use a vendor-managed workforce and set up data encryption

B.Use a private workforce consisting of the company's employees and launch the labeling job in the us-east-1 region

C.Use a public workforce (Mechanical Turk) and select the US East region

D.Use a private workforce but launch the labeling job in the eu-west-1 region

AnswerB

A private workforce of the company's own employees keeps PHI within the organisation, and launching in us-east-1 ensures the data remains in the US. This satisfies the stem's requirement that PHI stays on US soil.

Why this answer

A private workforce consisting of the company's own employees ensures that PHI is never exposed to external workers, and launching the labeling job in us-east-1 keeps all data within the US, satisfying the data residency requirement. Amazon SageMaker Ground Truth allows you to restrict data access to a private workforce that you manage, and by selecting a US region, you ensure that data processing and storage remain within US borders.

Exam trap

The trap here is that candidates often assume that selecting a US region with a public workforce is sufficient for compliance, overlooking that PHI must not be exposed to external workers regardless of geographic location.

How to eliminate wrong answers

Option A is wrong because a vendor-managed workforce involves third-party vendors who may not have the same compliance controls for PHI, and data encryption alone does not guarantee that data remains within the US. Option C is wrong because a public workforce (Mechanical Turk) exposes PHI to anonymous external workers, which violates HIPAA and data privacy requirements, even if the region is set to US East. Option D is wrong because launching the labeling job in eu-west-1 (Ireland) violates the requirement that PHI must remain in the US, as data would be processed and stored in the European Union.

Full explanation →

121

MCQmedium

A media company is using Amazon Bedrock to generate captions for images. They have a batch processing pipeline that sends thousands of images daily to the Bedrock API using the Titan Image Generator G1 model. Recently, they started receiving ThrottlingException errors during peak hours. The team needs to process all images within 24 hours without changing the model or the application code. The current account has a default quota of 10 requests per second (RPS) for the Titan model in us-east-1. The team estimates they need 50 RPS during peak hours. They have already implemented exponential backoff in the client, but the errors persist. What is the MOST effective solution to resolve the throttling issue?

A.Request a service quota increase for the InvokeModel API for the Titan model in us-east-1

B.Use Amazon SageMaker batch transform to process images offline

C.Distribute the requests across multiple AWS Regions

D.Switch to a different foundation model that has a higher default quota

AnswerA

Raising the InvokeModel quota for Titan in us-east-1 lifts the 10 RPS ceiling to the required 50 RPS, clearing ThrottlingException at source. Exponential backoff cannot exceed a hard per-account, per-region, per-model limit, and the stem forbids changing model or code.

Why this answer

The team has already implemented exponential backoff, but the errors persist because their current quota of 10 RPS is insufficient for the required 50 RPS. Requesting a service quota increase for the InvokeModel API for the Titan Image Generator G1 model in us-east-1 directly addresses the root cause by raising the throughput limit, allowing the existing application code and model to handle the peak load without any architectural changes.

Exam trap

The trap here is that candidates may think exponential backoff or distributing across Regions solves all throttling, but the core issue is a hard service quota that must be increased, not a transient rate limit.

How to eliminate wrong answers

Option B is wrong because Amazon SageMaker batch transform is designed for offline inference on SageMaker endpoints, not for invoking Bedrock APIs; it would require changing the application code and infrastructure, which the question explicitly prohibits. Option C is wrong because distributing requests across multiple AWS Regions would require modifying the application code to route traffic to different endpoints, and it does not address the underlying quota issue in the primary region; it also introduces latency and complexity. Option D is wrong because switching to a different foundation model would require changing the application code and potentially the image generation logic, which is not allowed; moreover, other models may have different default quotas or capabilities, and the goal is to process images with the Titan model.

Full explanation →

122

MCQmedium

A startup is using Amazon Bedrock to power a virtual assistant. They need to ensure that personally identifiable information (PII) is not included in the model's responses. Which feature should they enable?

A.Enable PII redaction in the Bedrock guardrails.

B.Enable model invocation logging.

C.Configure a VPC endpoint.

D.Enable data encryption at rest.

AnswerA

Bedrock guardrails apply PII redaction as a configurable filter that detects and masks sensitive data such as names, addresses and account numbers in both prompts and model responses. This directly satisfies the startup's requirement that personally identifiable information never appears in the virtual assistant's output, without retraining or altering the underlying foundation model.

Why this answer

Amazon Bedrock Guardrails provide a configurable content filtering and PII redaction feature that can automatically detect and mask personally identifiable information (PII) in model inputs and outputs. By enabling PII redaction within guardrails, the startup can ensure that sensitive data like names, addresses, or credit card numbers are removed or obfuscated before the virtual assistant's responses reach the user. This is the direct and intended mechanism for preventing PII leakage in model responses.

Exam trap

The trap here is that candidates often confuse data protection features like encryption or logging with content filtering, not realizing that PII redaction is a specific guardrail policy that actively modifies model outputs in real time.

How to eliminate wrong answers

Option B is wrong because model invocation logging captures metadata and request/response payloads for auditing and debugging, but it does not actively redact or filter PII from responses — it only records what was sent and received. Option C is wrong because configuring a VPC endpoint provides private network connectivity to Bedrock without traversing the public internet, but it has no capability to inspect or modify the content of model responses for PII. Option D is wrong because enabling data encryption at rest protects stored data (e.g., logs, model artifacts) from unauthorized access, but it does not perform real-time redaction of PII in model outputs during inference.

Full explanation →

123

MCQmedium

A data scientist wants to compare the text embeddings generated by Amazon Titan Embeddings for a set of product descriptions. Which metric is MOST appropriate to measure the semantic similarity between two embedding vectors?

A.Hamming distance

B.Manhattan distance

C.Euclidean distance

D.Cosine similarity

AnswerD

Cosine similarity measures the angle between two vectors, ignoring magnitude, so it reflects semantic orientation rather than length. This suits comparing Titan Embeddings, where semantic closeness is encoded in direction, and it remains stable across vectors of differing dimensionality-normalised scales.

Why this answer

Cosine similarity measures the cosine of the angle between two vectors, focusing on their orientation rather than magnitude. For text embeddings from models like Amazon Titan Embeddings, which are normalized to unit length, cosine similarity is the standard metric because it captures semantic similarity even when descriptions differ in length or word count. Euclidean distance would be affected by vector magnitude, making it less suitable for comparing semantic content.

Exam trap

A common mistake is to assume Euclidean distance works for text embeddings, but AWS Titan Embeddings produces unit vectors, so cosine similarity is the standard metric.

How to eliminate wrong answers

Option A is wrong because Hamming distance measures the number of positions at which two binary strings differ, which is not applicable to continuous-valued embedding vectors. Option B is wrong because Manhattan distance sums absolute differences along each dimension, which is sensitive to the magnitude of the vectors and does not isolate directional similarity. Option C is wrong because Euclidean distance computes the straight-line distance between points, which is influenced by vector length; in high-dimensional embedding spaces, it can be dominated by magnitude differences rather than semantic alignment.

Full explanation →

124

MCQeasy

A developer is building an application that generates product descriptions from images using a multimodal model. Which AWS service provides access to multimodal foundation models?

A.Amazon Rekognition

B.Amazon Textract

C.Amazon Comprehend

D.Amazon Bedrock

AnswerD

Amazon Bedrock supplies managed access to multimodal foundation models, including Amazon Titan and third-party options, through a single API. It satisfies the stem's requirement for generating text from images without infrastructure management, unlike services limited to text-only models or custom training workflows.

Why this answer

Amazon Bedrock is a managed service that provides access to a wide range of foundation models (FMs) from leading AI providers, including multimodal models that can process both images and text to generate product descriptions. This makes it the correct choice for building an application that requires multimodal capabilities.

Exam trap

The trap here is that candidates may confuse AWS AI services that handle specific modalities (Rekognition for images, Comprehend for text) with Bedrock, which is the only service that provides access to generative multimodal foundation models capable of combining both modalities in a single inference.

How to eliminate wrong answers

Option A is wrong because Amazon Rekognition is a computer vision service for image and video analysis (e.g., object detection, facial recognition), but it does not provide access to generative multimodal foundation models. Option B is wrong because Amazon Textract is an OCR service that extracts text from documents, not a platform for accessing or running multimodal generative models. Option C is wrong because Amazon Comprehend is a natural language processing (NLP) service for text analysis (e.g., sentiment, entities), and it lacks support for multimodal input or generative model access.

Full explanation →

125

Multi-Selecthard

A fintech startup is preparing its first machine learning project to detect fraudulent card transactions. The team must decide which characteristics make a problem well suited to supervised learning. Which TWO characteristics indicate that supervised learning is appropriate? (Choose two.)

Select 2 answers

A.The team wants the algorithm to discover hidden structure without any predefined output.

B.Historical transactions exist that were confirmed as fraudulent or legitimate by investigators.

C.The team prefers a model that explains each prediction with feature importance values.

D.The goal is to predict a discrete outcome for each new transaction.

E.The dataset contains millions of unlabeled transactions with no investigator feedback.

AnswersB, D

Confirmed historical outcomes are exactly the labeled data supervised learning requires. Each transaction carries a target the model can learn from, allowing it to map transaction features to a fraud or legitimate decision. Without such labels, a supervised classifier could not be trained, so this characteristic is a defining indicator that the approach fits.

Why this answer

Supervised learning needs labeled examples and a defined target to predict. Confirmed fraudulent or legitimate transactions supply those labels, and a discrete fraud-or-not decision supplies a suitable classification target. Unlabeled data, label-free exploration goals, and explainability preferences do not establish the labeled input-output pairing that supervised learning fundamentally depends on.

Exam trap

The trap here is assuming that a large transaction volume or a desire for explainability alone makes a problem supervised, when labeled outcomes and a defined target are what actually matter.

Full explanation →

126

Multi-Selectmedium

Which TWO actions should a data scientist take to evaluate fairness of a binary classification model using Amazon SageMaker Clarify? (Choose two.)

Select 2 answers

A.Use post-training bias metrics like Difference in Positive Proportions

B.Ensure the training dataset is balanced by resampling

C.Generate SHAP values for feature importance

D.Use pre-training bias metrics such as Class Imbalance

E.Run a data quality monitoring job on unlabeled data

AnswersA, D

Difference in Positive Proportions in Predicted Labels is a post-training bias metric, comparing predicted positive rates across facets. It quantifies disparate impact in the model's actual decisions, satisfying the requirement to evaluate fairness of the deployed binary classifier.

Why this answer

Option A is correct because SageMaker Clarify's post-training bias metrics, such as Difference in Positive Proportions (DPP), compare predicted outcomes across facets (e.g., gender or age groups) to quantify disparate impact in the model's actual predictions, which is essential for evaluating fairness of a binary classifier. Option D is correct because pre-training bias metrics like Class Imbalance (CI) measure skew in the underlying training data's label distribution across facets before any model is trained, helping detect whether the data itself could lead to biased outcomes. Together, these two metric types cover both data-level and model-level fairness assessment as Clarify is designed to do.

Option B is not a Clarify fairness evaluation action but a data preprocessing technique, and balancing data does not by itself measure bias. Option C, SHAP values, explains feature importance/attribution for interpretability, not fairness metrics. Option E, data quality monitoring on unlabeled data, addresses data drift/quality issues and does not evaluate model fairness.

Exam trap

The trap here is that candidates may confuse bias detection with data preprocessing or model explainability, leading them to select resampling (B) or SHAP values (C) instead of recognizing that SageMaker Clarify specifically provides pre-training and post-training bias metrics as separate evaluation steps.

Full explanation →

127

MCQhard

A fintech company wants its Amazon Bedrock assistant to answer customer questions only from its approved policy documents and to avoid fabricating answers when the documents do not cover a topic. The team plans to use a knowledge base with retrieval augmented generation. Which Bedrock Guardrails feature should they configure to detect and block responses that are not supported by the retrieved source passages?

A.A contextual grounding check with a grounding threshold and relevance threshold

B.A sensitive-information filter that blocks personally identifiable information

C.A word filter that blocks a custom list of profane terms

D.Content filters for hate, violence, and insult

AnswerA

Contextual grounding evaluates whether a response is supported by the retrieved source and whether it is relevant to the user query, using configurable grounding and relevance thresholds. Setting these thresholds makes the assistant block or flag answers not entailed by the approved policy passages, directly preventing fabrication when the documents do not cover a topic.

Why this answer

Preventing fabricated answers requires checking whether a response is entailed by the retrieved context. Bedrock Guardrails contextual grounding performs exactly that comparison, and its grounding and relevance thresholds determine how strictly unsupported or off-topic responses are blocked. Harmful-content, PII, and word filters address toxicity or confidentiality but never verify factual support from source passages.

Exam trap

The trap here is assuming any Guardrails filter prevents hallucination, when only the contextual grounding check compares responses against retrieved source passages.

Full explanation →

128

MCQmedium

An application uses this configuration to enable RAG. What is required for the knowledge base to function?

A.The agent must have internet access to retrieve documents

B.The embedding model ARN must include the account ID

C.The embedding model must be fine-tuned on the domain data

D.The knowledge base must have a vector index configured in Amazon OpenSearch Serverless

AnswerD

RAG retrieval depends on semantic similarity search, which requires vector embeddings stored in an index. Amazon OpenSearch Serverless provides that vector index, so the knowledge base can match queries to relevant document chunks before passing them to the model.

Why this answer

For a knowledge base to function in a RAG (Retrieval-Augmented Generation) setup on AWS, the knowledge base must have a vector index configured in Amazon OpenSearch Serverless. This vector index stores the embeddings generated from the source documents, enabling efficient similarity search to retrieve relevant context for the agent. Without a vector index, the knowledge base cannot perform the vector search required to fetch relevant document chunks.

Exam trap

The trap here is that candidates may think the embedding model must be fine-tuned or that internet access is needed, but the core requirement is the vector index in a vector store like Amazon OpenSearch Serverless on AWS, which is essential for the retrieval step in RAG.

How to eliminate wrong answers

Option A is wrong because the agent does not need internet access to retrieve documents; the knowledge base is hosted within AWS and accessed via internal API calls, not over the public internet. Option B is wrong because the embedding model ARN does not need to include the account ID; ARNs already contain the account ID by default, and the requirement is that the model must be accessible (e.g., via Amazon Bedrock) and the ARN must be correctly specified, but the account ID is not an additional requirement. Option C is wrong because the embedding model does not need to be fine-tuned on domain data; a pre-trained embedding model (e.g., from Amazon Bedrock) can generate embeddings for any domain, and fine-tuning is not a prerequisite for RAG functionality.

Full explanation →

129

MCQmedium

A company deployed a chatbot using Amazon Lex integrated with a Lambda function that invokes Claude on Amazon Bedrock. The Lambda function retrieves relevant documents from an Amazon Kendra index to use as context. Users report that the chatbot's responses are often irrelevant or incorrect despite the Kendra index containing accurate information. The logs show that the Lambda function is correctly passing retrieved documents to the model. What is the most likely cause and solution?

A.Switch to a larger foundation model like Claude 3 Opus

B.The model's temperature is set too high; reduce it to 0.1

C.The maximum tokens limit is too low; increase it to 4096

D.The chunking strategy for documents is too coarse or inappropriate; refine chunking and use semantic search in Kendra

AnswerD

Coarse chunking embeds large blocks, so Kendra returns passages whose vectors dilute the specific answer, and the model receives loosely relevant context despite accurate documents. Refining chunk size and enabling semantic search sharpens retrieval precision, satisfying the requirement that passed context actually match the user's question.

Why this answer

When retrieved documents are correctly passed to the model but responses are still irrelevant, the problem is almost always upstream in retrieval quality — specifically how documents were chunked and indexed in Kendra. Coarse or poorly aligned chunks dilute the semantic signal, so the model receives context that does not actually answer the query. Refining chunking strategy and enabling semantic search in Kendra directly addresses the root cause.

Exam trap

AIF-C01 often tests the misconception that a bigger or more expensive model fixes RAG quality — candidates must recognize that retrieval (chunking, indexing, semantic search) is the usual bottleneck when context is already being passed.

How to eliminate wrong answers

Option A is wrong because switching to a larger model like Claude 3 Opus does not fix bad retrieval — a bigger model given irrelevant context will still produce irrelevant answers, and it increases cost and latency without addressing the root cause. Option B is wrong because temperature controls randomness/creativity, not factual grounding; a high temperature might cause variability but the logs show documents are being passed correctly, so the issue is retrieval relevance, not sampling. Option C is wrong because max tokens limits response length, not relevance — increasing it just allows longer (still wrong) answers and does not improve the quality of retrieved context.

Full explanation →

130

MCQhard

A media company wants to build a system that generates short promotional descriptions for its articles. The team has no labeled dataset of article-summary pairs but has a large corpus of published articles and descriptions. They want to leverage a pretrained foundation model and adapt it to their domain with minimal labeling effort. Which approach best fits this scenario?

A.Use reinforcement learning with reader click rewards to train a description generator from scratch.

B.Fine-tune a pretrained foundation model on the company's article and description corpus for text generation.

C.Train a supervised classification model that assigns each article to a predefined topic category.

D.Apply unsupervised clustering to group articles by topic and use cluster IDs as descriptions.

AnswerB

Foundation models are pretrained on broad text and can be adapted to a specific domain through fine-tuning on in-domain examples. The company's corpus of articles paired with descriptions provides exactly the material to steer the model toward generating promotional text in its own style. This leverages pretrained capability while requiring far less labeling than training a model from scratch.

Why this answer

The company needs generated text and already holds a domain corpus of articles paired with descriptions. Fine-tuning a pretrained foundation model adapts broad language ability to the company's style with modest additional data, which matches the stated preference for minimal labeling. Classification, clustering, and from-scratch reinforcement learning either produce the wrong output type or discard the pretrained model advantage.

Exam trap

The trap here is assuming any use of a foundation model must be prompt-only, overlooking that fine-tuning on in-domain pairs is a valid adaptation path.

Full explanation →

131

Multi-Selectmedium

A company is building a generative AI application using Amazon Bedrock and needs to ensure that the model does not generate outputs containing personally identifiable information (PII). Which TWO actions should the company take? (Choose 2)

Select 2 answers

A.Implement a custom AWS Lambda function to scan and redact PII from inputs and outputs.

B.Use AWS Identity and Access Management (IAM) policies to restrict model access.

C.Enable Amazon CloudWatch Logs to capture and audit model outputs.

D.Configure Amazon Bedrock Guardrails to block or mask PII.

E.Place the Bedrock model endpoint within a private VPC.

AnswersA, D

A Lambda function provides deterministic, customisable inspection and redaction of PII across both request and response payloads, catching patterns Guardrails filters may not cover. It satisfies the requirement to prevent PII appearing in generated outputs.

Why this answer

Option A is correct because a custom AWS Lambda function can programmatically scan both the prompt inputs and the model responses for PII patterns (for example, using regex or Amazon Comprehend's PII detection APIs) and redact or block them before they reach the user, giving the company full control over what data enters and leaves the generative AI application. Option D is correct because Amazon Bedrock Guardrails provides a native, purpose-built sensitive information filter that can detect and either block or mask PII entities (such as names, emails, and SSNs) in both user inputs and model outputs, directly satisfying the requirement to prevent PII from appearing in generated content. Option B is not correct because IAM policies only control who can invoke which Bedrock models and actions; they do not inspect or filter the content of prompts and responses for PII.

Option C is not correct because CloudWatch Logs only captures and stores model outputs for auditing and monitoring after the fact; it does not prevent PII from being generated or returned. Option E is not correct because placing the Bedrock endpoint in a private VPC only affects network isolation and traffic routing, not the content-level detection or redaction of PII in model outputs.

Exam trap

The AIF-C01 exam often tests the distinction between network-level security controls (like VPCs) and content-level data protection mechanisms, leading candidates to mistakenly choose VPC isolation as a solution for PII redaction.

Full explanation →

132

MCQeasy

Which component of the Transformer architecture allows the model to weigh the importance of different tokens in the input sequence when generating output?

A.Layer normalization

B.Self-attention mechanism

C.Positional encoding

D.Feed-forward neural network

AnswerB

Self-attention computes pairwise compatibility scores between every token and all others, producing weighted representations that emphasise contextually relevant tokens. This directly satisfies the stem's requirement to weigh token importance during output generation, unlike feed-forward layers or positional encodings, which handle transformation and ordering respectively.

Why this answer

The self-attention mechanism (B) is the core component that enables the Transformer to dynamically assign importance weights to every token in the input sequence relative to every other token. This allows the model to capture long-range dependencies and contextual relationships, which is essential for generating coherent output in tasks like translation or summarization.

Exam trap

AWS often tests the distinction between components that add information (positional encoding) versus those that compute relationships (self-attention), leading candidates to confuse positional encoding as the mechanism for weighting token importance.

How to eliminate wrong answers

Option A is wrong because layer normalization stabilizes training by normalizing activations across features, but it does not weigh token importance or model relationships between tokens. Option C is wrong because positional encoding adds information about token order to the input embeddings, but it does not perform any weighting or attention computation. Option D is wrong because the feed-forward neural network processes each token independently after attention has combined information; it does not weigh token importance across the sequence.

Full explanation →

133

MCQhard

A company wants to use a large language model to generate code based on natural language descriptions. They need to minimize latency and control costs by running inference on their own infrastructure. Which approach is most suitable?

A.Use Amazon Bedrock API

B.Use Amazon SageMaker to deploy a custom LLM

C.Use Amazon Comprehend

D.Use Amazon Lex

AnswerB

SageMaker deploys the model on infrastructure the company controls, so inference runs locally rather than through a third-party API. This satisfies both constraints: reduced network latency and predictable, controllable cost, unlike fully managed model endpoints.

Why this answer

Amazon SageMaker allows you to deploy a custom large language model (LLM) on your own infrastructure, giving you full control over inference latency and cost. By using SageMaker endpoints with auto-scaling and instance selection, you can optimize for low-latency responses while avoiding per-token API charges from managed services.

Exam trap

AWS often tests the distinction between managed API services (like Bedrock) and self-managed deployment options (like SageMaker), where candidates mistakenly choose Bedrock for 'control' over costs and latency, not realizing that Bedrock is a pay-per-token managed service with no infrastructure control.

How to eliminate wrong answers

Option A is wrong because Amazon Bedrock is a managed API service that charges per-token and does not allow you to run inference on your own infrastructure, so you cannot control latency or costs at the infrastructure level. Option C is wrong because Amazon Comprehend is a natural language processing (NLP) service for tasks like sentiment analysis and entity extraction, not a generative AI service capable of code generation from natural language. Option D is wrong because Amazon Lex is designed for building conversational chatbots using intent-based models, not for deploying large language models for code generation.

Full explanation →

134

MCQmedium

Refer to the exhibit. A data scientist runs an Amazon SageMaker Clarify bias analysis on a binary classifier. The pre-training ClassImbalance is 1.5 and the post-training DPPL is 0.15. What should the data scientist conclude?

A.The data is highly imbalanced and the model is unbiased.

B.The data has a mild class imbalance, but the model shows a noticeable bias in predictions.

C.The pre-training metric indicates a fairness issue, but the post-training metric is acceptable.

D.The data is perfectly balanced and the model is fair.

AnswerB

ClassImbalance of 1.5 sits just above the balanced threshold, indicating only mild skew in the training data. DPPL of 0.15 exceeds the typical 0.1 tolerance, meaning predicted positive rates differ noticeably between groups, so the model exhibits real predictive bias.

Why this answer

The pre-training ClassImbalance metric of 1.5 indicates a mild class imbalance (values close to 1.0 indicate balance, while values significantly above 1.0 indicate imbalance). The post-training DPPL (Difference in Positive Proportions in Labels) metric of 0.15 exceeds the commonly accepted fairness threshold of 0.10, indicating a noticeable bias in the model's predictions. Therefore, the data has a mild imbalance, but the model exhibits a bias that warrants further investigation.

Exam trap

In AWS AI Practitioner exams, a common misconception is that a low pre-training imbalance automatically means the model is fair, but the post-training DPPL metric directly measures prediction bias and can reveal unfairness even when the data appears balanced.

How to eliminate wrong answers

Option A is wrong because a ClassImbalance of 1.5 indicates a mild imbalance, not a highly imbalanced dataset, and the DPPL of 0.15 suggests the model is biased, not unbiased. Option C is wrong because the pre-training metric of 1.5 does not indicate a fairness issue—it only measures class distribution, not fairness—and the post-training DPPL of 0.15 is above the 0.10 threshold, making it unacceptable. Option D is wrong because a ClassImbalance of 1.5 is not perfectly balanced (perfect balance is 1.0), and a DPPL of 0.15 indicates the model is not fair.

Full explanation →

135

MCQmedium

A company runs a Retrieval Augmented Generation (RAG) application on Amazon Bedrock. The application uses an Amazon OpenSearch Serverless vector index to store internal HR documents. The security team must ensure that the vector index is encrypted at rest with a customer-managed AWS KMS key so that they can control key rotation and revoke access independently. Which configuration should they implement?

A.Configure the OpenSearch Serverless collection with an AWS owned key and enable automatic key rotation in AWS KMS.

B.Enable Amazon S3 default encryption with SSE-KMS on the bucket that stores the source documents, then rebuild the vector index.

C.Use AWS PrivateLink to connect to the OpenSearch Serverless collection and rely on TLS in transit for data protection.

D.Create a customer managed KMS key and associate it with the OpenSearch Serverless collection at creation time using the encryption policy.

AnswerD

OpenSearch Serverless supports specifying a customer managed AWS KMS key in the collection's encryption policy at creation. This key encrypts the vector index at rest and allows the security team to manage rotation and revoke access by disabling or deleting the key. This directly satisfies the requirement for customer-controlled encryption of the vector store used by the RAG application.

Why this answer

A customer managed KMS key must be specified in the OpenSearch Serverless encryption policy when the collection is created, because the encryption key cannot be changed after creation. This gives the security team control over rotation and the ability to revoke access by disabling the key. Encrypting source documents or using private connectivity does not encrypt the vector index at rest with a customer managed key.

Exam trap

The trap here is assuming that encrypting the source documents in Amazon S3 also encrypts the derived vector index or that the encryption key can be swapped after the collection exists.

Full explanation →

136

MCQeasy

A startup wants to build a mobile app that generates personalized workout plans using a foundation model. They need to minimize infrastructure management and pay only for what they use. Which AWS service should they use to access foundation models via a single API?

A.Amazon Polly

B.Amazon Comprehend

C.Amazon SageMaker

D.Amazon Bedrock

AnswerD

Amazon Bedrock is a fully managed service that provides access to multiple foundation models from different providers through a single API. It eliminates infrastructure management and offers pay-as-you-go pricing, making it ideal for a startup that wants to minimize operational overhead and control costs.

Why this answer

The startup needs a managed service that offers foundation models via a single API with pay-as-you-go pricing. Amazon Bedrock provides exactly that, allowing developers to experiment with and deploy models from multiple providers without managing infrastructure. Other services like SageMaker require more management, while Comprehend and Polly serve different purposes.

Exam trap

The trap here is confusing Amazon SageMaker with Amazon Bedrock; SageMaker is a broader ML platform that requires more infrastructure management and does not offer a unified API for multiple foundation models.

Full explanation →

137

MCQmedium

Refer to the exhibit. You receive this response from Amazon Bedrock. What is the most likely cause of the incomplete information?

A.The max_tokens limit was reached

B.The prompt was too short

C.The temperature was too high

D.The model lacks knowledge about capitals

AnswerA

Generation halts once the max_tokens ceiling is hit, so the response ends mid-sentence with content omitted. The truncated, incomplete output in the exhibit matches that hard cutoff, since the model cannot emit further tokens beyond the configured limit.

Why this answer

The response from Amazon Bedrock shows an incomplete sentence that cuts off mid-thought, which is a classic symptom of hitting the max_tokens limit. When the generated output reaches the specified maximum number of tokens, the model stops generating immediately, resulting in truncated text. This is the most likely cause because the output is syntactically incomplete but otherwise coherent up to the cutoff point.

Exam trap

AWS often tests the distinction between output truncation (max_tokens) and output quality issues (temperature, prompt engineering), so the trap here is that candidates may incorrectly attribute a truncated response to model ignorance or randomness rather than the explicit token limit.

How to eliminate wrong answers

Option B is wrong because the prompt length does not directly cause incomplete output; a short prompt can still produce a complete response if the max_tokens limit is high enough. Option C is wrong because temperature controls randomness and creativity, not the length or truncation of the output; high temperature might produce less coherent text but would not cut off mid-sentence. Option D is wrong because the model's lack of knowledge about capitals would result in incorrect or hallucinated information, not a truncated or incomplete sentence.

Full explanation →

138

MCQeasy

An organization wants to control which topics their AI chatbot can discuss. For example, they want to block all conversations about investment advice. Which Amazon Bedrock Guardrails feature should they configure?

A.Contextual grounding check

B.Content filtering with category-based harmful content filters

C.Sensitive information filters

D.Topic restrictions

AnswerD

Topic restrictions define denied subjects, so the guardrail evaluates prompts and responses against them and blocks investment-advice conversations. Content filters target harmful categories such as hate or violence, whereas denied topics are the mechanism for excluding specific subject matter.

Why this answer

Amazon Bedrock Guardrails topic restrictions allow you to define a set of topics that the model should avoid discussing. By configuring topic restrictions, you can block conversations about investment advice by providing a natural language description of the topic and example phrases. This is the specific feature designed to deny-list topics.

Exam trap

AIF-C01 often tests the distinction between content filters (harmful categories) and topic restrictions (custom denied topics), causing candidates to choose content filtering when the requirement is to block a specific subject like investment advice.

How to eliminate wrong answers

Option A is wrong because contextual grounding check is used to detect and filter hallucinations by comparing responses to source information, not to block topics. Option B is wrong because content filtering with category-based harmful content filters targets harmful categories like hate, violence, and sexual content, not specific topics like investment advice. Option C is wrong because sensitive information filters are for detecting and redacting PII or custom regex patterns, not for blocking discussion topics.

Full explanation →

139

MCQmedium

A company is using Amazon Bedrock to build a multilingual support chatbot. They need a model that can understand and generate text in multiple languages without requiring separate fine-tuning per language. Which model capability is MOST important?

A.Large context window

B.Multilingual training data

C.Support for streaming responses

D.High temperature setting

AnswerB

Multilingual training data lets a single model learn shared representations across languages, so it understands and generates text in many languages without per-language fine-tuning. This directly satisfies the constraint of avoiding separate fine-tuning for each supported language.

Why this answer

A model's ability to understand and generate text in multiple languages without separate fine-tuning depends on being trained on a diverse, multilingual dataset. This allows the model to learn cross-lingual representations and perform zero-shot transfer between languages, which is essential for a multilingual chatbot.

Exam trap

AWS often tests the misconception that a large context window or streaming responses are prerequisites for multilingual support, when in fact these features address throughput and latency, not language diversity.

How to eliminate wrong answers

Option A is wrong because a large context window determines how much text the model can process at once (e.g., 8K or 32K tokens), but it does not inherently enable multilingual understanding or generation. Option C is wrong because streaming responses affect how output is delivered (incrementally vs. all at once), not the model's language capabilities. Option D is wrong because a high temperature setting increases randomness in output generation, which can degrade coherence and accuracy, and does not contribute to multilingual proficiency.

Full explanation →

140

Multi-Selectmedium

A company is deploying a foundation model on Amazon Bedrock to generate product descriptions. They want to ensure the model's output is factually consistent with the provided product specifications and avoids hallucinated features. Which TWO techniques should they use? (Choose two.)

Select 2 answers

A.Set the temperature to a high value to encourage the model to explore more creative descriptions.

B.Fine-tune the model on a dataset of existing product descriptions to improve its writing style.

C.Use Amazon Bedrock Guardrails to define a denied topic for features not present in the specifications.

D.Use retrieval-augmented generation (RAG) with Amazon Bedrock Knowledge Bases to ground the model in the product specifications.

E.Provide clear instructions and product specifications in the prompt, and ask the model to only use the given information.

AnswersD, E

RAG with Amazon Bedrock Knowledge Bases retrieves relevant product specifications and provides them to the model as context. This grounds the generation in factual data, reducing hallucinations and ensuring the output aligns with the actual features. It is a direct method to improve factual consistency.

Why this answer

To ensure factual consistency, the company should ground the model in the actual product specifications. Retrieval-augmented generation with Amazon Bedrock Knowledge Bases provides relevant context, while clear prompt instructions that restrict the model to given information further reduce hallucinations. Both techniques work together to keep output aligned with facts.

Exam trap

The trap here is thinking that Guardrails or fine-tuning can enforce factual consistency, when they address different concerns like content filtering or style adaptation, not grounding in specific data.

Full explanation →

141

Multi-Selecteasy

A developer wants to use Amazon Bedrock to build a text summarization application. Which TWO of the following are required steps?

Select 2 answers

A.Create an Amazon SageMaker endpoint for hosting the model

B.Train a custom model from scratch on summarization data

C.Set up an Amazon EMR cluster to preprocess the text data

D.Request access to a foundation model in Amazon Bedrock

E.Invoke the model using the Bedrock API or SDK with a prompt containing the text to summarize

AnswersD, E

Requesting model access in Amazon Bedrock is mandatory before invocation, since Anthropic, Meta and Amazon models are gated by default. This satisfies the stem's requirement that the developer must obtain entitlement to a specific foundation model in the target AWS Region before the summarisation application can call it.

Why this answer

Option D is correct because Amazon Bedrock requires you to request and be granted access to a specific foundation model (for example, Amazon Titan Text or Anthropic Claude) in the target AWS Region before you can invoke it, since models are not enabled by default. Option E is correct because once access is granted, the application must call the model through the Bedrock runtime API (InvokeModel or Converse) using the AWS SDK or CLI, passing a prompt that contains the text to be summarized. Option A is incorrect because Bedrock is a fully managed, serverless service that exposes foundation models via API, so no SageMaker endpoint needs to be created or managed.

Option B is incorrect because Bedrock provides pre-trained foundation models that can be used as-is or customized with techniques like fine-tuning or RAG, but training a model from scratch is not required. Option C is incorrect because Amazon EMR is not a prerequisite for Bedrock; any text preprocessing can be done in the application itself or with simpler services, and EMR is unnecessary for this scenario.

Exam trap

AWS often tests the misconception that you must train or host your own model to use generative AI, when in fact managed services like Bedrock provide pre-trained models accessible via API, eliminating the need for custom infrastructure.

Full explanation →

142

MCQmedium

A developer is using Amazon Bedrock to build a generative AI application. The application must deny any user request that involves asking about the topic of 'mergers and acquisitions'. Which Bedrock feature should the developer use?

A.Bedrock PII Detection

B.Bedrock Content Filtering

C.Bedrock Grounding Check

D.Bedrock Topic Denial (part of Guardrails)

AnswerD

Topic denial policies in Bedrock Guardrails are specifically designed to block or deny conversations on defined topics like 'mergers and acquisitions'.

Why this answer

Bedrock Guardrails includes a Topic Denial feature that allows developers to define specific topics (e.g., 'mergers and acquisitions') that the model must refuse to discuss. When a user input or model output matches a denied topic, Guardrails blocks the response and returns a predefined denial message, ensuring compliance with organizational policies.

Exam trap

The trap here is that candidates often confuse content filtering (which blocks harmful content) with topic denial (which blocks specific business topics), leading them to select Option B instead of D.

How to eliminate wrong answers

Option A is wrong because Bedrock PII Detection is designed to identify and redact personally identifiable information (e.g., SSNs, credit card numbers) from inputs or outputs, not to block entire topics like 'mergers and acquisitions'. Option B is wrong because Bedrock Content Filtering focuses on filtering harmful or offensive content (e.g., hate speech, violence) based on severity thresholds, not on denying specific business topics. Option C is wrong because Bedrock Grounding Check verifies that model responses are grounded in a source document or knowledge base to reduce hallucinations, but it does not provide topic-level denial capabilities.

Full explanation →

143

MCQeasy

A company is using Amazon SageMaker to train machine learning models on sensitive customer data. Which AWS service can be used to encrypt the data at rest in the S3 bucket used by SageMaker?

A.AWS Key Management Service (KMS)

B.AWS CloudHSM

C.AWS Secrets Manager

D.AWS Certificate Manager (ACM)

AnswerA

AWS Key Management Service supplies the customer-managed keys that encrypt SageMaker training data and the associated S3 bucket at rest, with access governed by key policies. This satisfies the stem's requirement for encryption of sensitive data at rest in S3.

Why this answer

AWS Key Management Service (KMS) is the correct service because it provides managed encryption keys that can be used to enable server-side encryption (SSE-KMS) for Amazon S3 buckets. When SageMaker accesses training data from S3, it can use a customer-managed KMS key to encrypt data at rest, ensuring sensitive customer data remains protected. KMS integrates directly with S3 and SageMaker, allowing you to specify a KMS key in the SageMaker training job configuration.

Exam trap

The trap here is that candidates often confuse AWS CloudHSM with KMS, thinking that a dedicated HSM is required for encryption, but KMS is the simpler, fully managed service that directly integrates with S3 and SageMaker for at-rest encryption.

How to eliminate wrong answers

Option B (AWS CloudHSM) is wrong because CloudHSM provides dedicated hardware security modules for key generation and storage, but it does not directly integrate with S3 for server-side encryption; you would need to manage the encryption process yourself, making it more complex and less suitable for simple at-rest encryption. Option C (AWS Secrets Manager) is wrong because Secrets Manager is designed to securely store and rotate secrets like database credentials and API keys, not to manage encryption keys for S3 data at rest. Option D (AWS Certificate Manager (ACM)) is wrong because ACM is used to provision, manage, and deploy SSL/TLS certificates for securing network traffic (in transit), not for encrypting data at rest in S3.

Full explanation →

144

MCQeasy

A startup wants to quickly prototype a generative AI application for summarizing news articles. They have limited ML expertise and want minimal infrastructure management. Which AWS service should they use?

A.Amazon Bedrock with a foundation model accessed via API.

B.Amazon SageMaker to build and train a custom summarization model.

C.AWS Lambda with a custom Python script using the Hugging Face Transformers library.

D.Amazon EC2 instance running a pre-trained model from AWS Marketplace.

AnswerA

Amazon Bedrock provides serverless API access to foundation models, eliminating infrastructure management and ML expertise requirements. This satisfies the startup's constraints of minimal infrastructure and limited ML skills, enabling rapid prototyping of news summarisation without training or hosting models.

Why this answer

Amazon Bedrock is the correct choice because it provides pre-trained foundation models from leading AI providers via a simple API, requiring no ML expertise or infrastructure management. The startup can quickly prototype a summarization application by sending news articles to the API and receiving summaries without training or deploying models.

Exam trap

AWS often tests the distinction between managed AI services (Bedrock) and infrastructure-heavy services (SageMaker, EC2, Lambda), where candidates mistakenly choose SageMaker for its flexibility or Lambda for its serverless nature, overlooking the specific requirement for minimal ML expertise and infrastructure management.

How to eliminate wrong answers

Option B is wrong because Amazon SageMaker is designed for building, training, and deploying custom ML models, which requires significant ML expertise and infrastructure management, contradicting the startup's need for minimal ML expertise and quick prototyping. Option C is wrong because AWS Lambda with a custom Python script using Hugging Face Transformers requires managing dependencies, cold start latency, and model loading within Lambda's memory and time limits, adding complexity and infrastructure management. Option D is wrong because Amazon EC2 requires provisioning, configuring, and managing the instance, including installing the model and handling scaling, which is not minimal infrastructure management and deviates from the goal of quick prototyping.

Full explanation →

145

Multi-Selectmedium

A hospital is deploying an AI system to assist in diagnosing diseases from medical images. According to the EU AI Act, this system may be classified as high-risk. Which THREE requirements should the hospital address to comply with the EU AI Act for high-risk AI systems?

Select 3 answers

A.Ensure transparency and provision of information to users

B.Establish a risk management system throughout the lifecycle

C.Enable human oversight to prevent or minimize risks

D.Use only open-source models

E.Deploy the system without any testing in a sandbox environment

AnswersA, B, C

The EU AI Act requires high-risk systems to be sufficiently transparent, with clear instructions and information supplied to deployers. For a diagnostic imaging tool, informing clinicians about capabilities, limitations and intended purpose satisfies this requirement, enabling informed, safe use.

Why this answer

The EU AI Act requires high-risk systems to have risk management, transparency, and human oversight among other requirements.

Full explanation →

146

MCQeasy

A social media platform uses an AI system to moderate content. They want to ensure that human reviewers can review decisions when the AI is uncertain. Which AWS service can be used to set up a human review workflow for AI predictions?

A.Amazon Mechanical Turk

B.AWS Lambda

C.Amazon SageMaker Ground Truth

D.Amazon Augmented AI (A2I)

AnswerD

Amazon Augmented AI builds human review workflows into ML applications, routing low-confidence predictions to reviewers. It directly satisfies the requirement to let humans review decisions when the AI is uncertain, integrating with Amazon Rekognition and custom models.

Why this answer

Amazon Augmented AI (A2I) is purpose-built to integrate human review into ML workflows, allowing you to define review workflows that trigger when model confidence is low. It integrates natively with Amazon Rekognition, Textract, and custom models via the A2I API, and routes uncertain predictions to human reviewers. This directly matches the requirement for human review of AI moderation decisions.

Exam trap

AIF-C01 often tests the confusion between SageMaker Ground Truth (dataset labeling) and Amazon A2I (production human-in-the-loop review), so candidates must match the use case to the correct service.

How to eliminate wrong answers

Option A is wrong because Amazon Mechanical Turk is a crowdsourcing marketplace for task completion, not a managed human review workflow service integrated with ML predictions. Option B is wrong because AWS Lambda is a compute service for running code, not a human review orchestration tool. Option C is wrong because SageMaker Ground Truth is designed for labeling training datasets, not for reviewing live model predictions in production; A2I is the service for production human-in-the-loop review.

Full explanation →

147

MCQeasy

A developer wants to compare the output quality of several foundation models available in Amazon Bedrock for a text summarization task. They need to evaluate responses side by side using the same prompt. Which Amazon Bedrock feature should they use?

A.Amazon Bedrock model evaluation

B.Amazon Bedrock provisioned throughput

C.Amazon Bedrock Agents

D.Amazon Bedrock Guardrails

AnswerA

Amazon Bedrock model evaluation allows you to compare foundation models using either automatic metrics or human evaluation. You can submit the same prompt to multiple models and assess summarization quality side by side. This directly supports the developer's goal of comparing output quality across models for a specific task, making it the appropriate feature.

Why this answer

Amazon Bedrock model evaluation provides both automatic and human-based evaluation workflows to compare foundation models on tasks like summarization. It lets you run the same prompt across models and review results side by side. Guardrails, Agents, and provisioned throughput serve different purposes such as content filtering, orchestration, and capacity reservation, so they cannot fulfill the comparison requirement.

Exam trap

The trap here is mixing up model evaluation with guardrails, since both relate to model outputs, but only evaluation is designed to compare and score quality across models.

Full explanation →

148

Multi-Selectmedium

A company is deploying an AI-based diagnostic system in healthcare. Which THREE practices align with AWS responsible AI guidelines? (Choose THREE.)

Select 3 answers

A.Deploy the model in production immediately after training without manual review.

B.Continuously monitor model performance for drift using SageMaker Model Monitor.

C.Use only automated decision-making without any human oversight.

D.Document the model's intended use and limitations with model cards.

E.Implement a human-in-the-loop process for high-risk predictions using Amazon A2I.

AnswersB, D, E

Monitoring ensures ongoing reliability and safety.

Why this answer

AWS Responsible AI guidelines emphasize several practices for high-risk systems like healthcare diagnostics. Continuous monitoring with SageMaker Model Monitor helps detect data quality, bias, and feature attribution drift, supporting reliability and safety (B). Documenting the model's intended use, limitations, and performance with model cards promotes transparency and accountability (D).

Implementing a human-in-the-loop review process using Amazon Augmented AI (A2I) ensures meaningful human oversight for high-risk predictions (E). In contrast, deploying without manual review (A) and relying solely on automated decisions (C) violate responsible AI principles requiring human oversight and validation.

Exam trap

A common misconception is that automated decision-making alone satisfies responsible AI, but AWS guidelines require human oversight for high-risk predictions, as emphasized in the AWS Well-Architected Framework and AIF-C01 guidelines.

Full explanation →

149

Multi-Selecteasy

Which TWO of the following are benefits of using Amazon Bedrock for building applications with foundation models?

Select 2 answers

A.No infrastructure management

B.Automatic model fine-tuning

C.Access to multiple foundation models

D.Free tier for all models

E.Built-in image generation capability

AnswersA, C

Amazon Bedrock is fully managed and serverless, so teams consume foundation models through a single API without provisioning or scaling any compute, satisfying the stem's benefit of eliminating infrastructure management. This removes capacity planning and patching overhead, letting developers focus on application logic rather than hosting model endpoints.

Why this answer

Option A (No infrastructure management) is correct because Amazon Bedrock is a fully managed, serverless service: AWS handles provisioning, scaling, patching, and hosting of the underlying compute for the foundation models, so developers just call the API without managing any servers or clusters. Option C (Access to multiple foundation models) is correct because Bedrock provides a single unified API to choose from a range of foundation models from providers such as Anthropic, AI21 Labs, Cohere, Meta, Stability AI, and Amazon, letting you switch or compare models without integrating separate SDKs or endpoints. Option B is not a Bedrock benefit as stated, since fine-tuning is an optional capability you must explicitly configure (and not all models support it), not an automatic feature.

Option D is incorrect because Bedrock is not free for all models; usage is billed per input/output token or per image, and pricing varies by model. Option E is incorrect because image generation is only available through specific models (e.g., Stability AI or Amazon Titan Image Generator), not as a universal built-in capability of the service.

Exam trap

AWS often tests the misconception that Amazon Bedrock includes built-in capabilities like automatic fine-tuning or image generation, when in reality these are model-specific features that you must explicitly select and configure, not inherent service features.

Full explanation →

150

MCQmedium

A data scientist is training a model using Amazon SageMaker and notices the training loss is decreasing but validation loss starts increasing after a few epochs. Which technique should they apply to address this?

A.Increase batch size

B.Increase the learning rate

C.Add more training data

D.Add regularization (e.g., L1 or L2)

AnswerD

Rising validation loss while training loss falls signals overfitting. L1 or L2 regularization penalises large weights, constraining model complexity so it generalises better to unseen data, directly addressing the divergence between the two loss curves.

Why this answer

The scenario describes overfitting, where the model memorizes training data but fails to generalize to validation data. Adding regularization (L1 or L2) penalizes large weights, reducing model complexity and improving generalization. This is a standard technique in SageMaker training jobs, often configured via the `regularizer` hyperparameter in frameworks like TensorFlow or MXNet.

Exam trap

The trap here is that candidates confuse overfitting with underfitting or optimization issues, and incorrectly choose to increase learning rate or batch size, not recognizing that rising validation loss with falling training loss is the classic signature of overfitting.

How to eliminate wrong answers

Option A is wrong because increasing batch size typically stabilizes gradient estimates but does not directly address overfitting; it may even reduce generalization by sharpening minima. Option B is wrong because increasing the learning rate can cause divergence or overshooting of the loss minimum, worsening both training and validation loss. Option C is wrong because adding more training data can help generalization but is not a direct fix for overfitting when validation loss increases; it may not be feasible or sufficient, and regularization is the immediate corrective action.

Full explanation →

AWS Certified AI Practitioner AIF-C01 (AIF-C01) — Questions 76–150