AWS Certified AI Practitioner AIF-C01 AIF-C01 Questions 526–600 | Page 8/12

526

Multi-Selecthard

A financial services firm is deploying a generative AI assistant that answers employee questions about internal policy documents. The security team requires that answers be traceable to source text and that the model not invent policy details. Which TWO techniques should be implemented to ground responses and reduce fabricated content? (Choose two.)

Select 2 answers

A.Instruct the model to answer only from the supplied context and to state when the context is insufficient

B.Increase the temperature setting so the model explores a wider range of responses

C.Retrieve relevant passages from the policy corpus and include them in the prompt before generation

D.Remove all system instructions so the model can respond more naturally to each question

E.Expand the model's parameter count by switching to the largest available variant

AnswersA, C

Grounding instructions constrain the model to the retrieved material and give it an explicit escape hatch when evidence is missing. That behavior converts a potential hallucination into an honest abstention, which is exactly what the security team wants, and it pairs naturally with retrieval so the model has context to cite.

Why this answer

Grounding comes from combining retrieval of authoritative passages with instructions that confine the model to that evidence and permit abstention. Together they make answers traceable and suppress invention. Higher temperature, larger parameter counts, and removal of system instructions either increase variability, fail to supply evidence, or discard the constraints that keep output faithful.

Exam trap

The trap here is treating a bigger model as a fix for hallucination, when grounding requires supplying evidence and constraining the model to it rather than scaling parameters.

Full explanation →

527

MCQmedium

A company is using Amazon Bedrock to generate embeddings for a semantic search application. They want to ensure that semantically similar phrases (e.g., "car" and "vehicle") produce similar vector representations. Which type of model should they use?

A.An embedding model like Amazon Titan Embeddings

B.An image generation model like Stable Diffusion

C.A text generation model like Anthropic Claude

D.A multimodal model that supports both text and image input

AnswerA

Amazon Titan Embeddings maps text into a dense vector space where semantic proximity is preserved, so synonyms such as "car" and "vehicle" yield closely aligned vectors. This directly satisfies the stem's requirement for similar vector representations, unlike generative or classification models, which output tokens or labels rather than comparable embeddings.

Why this answer

Amazon Titan Embeddings is a text embedding model specifically designed to convert textual input into dense vector representations that capture semantic meaning. By mapping semantically similar phrases like 'car' and 'vehicle' to nearby points in the embedding space, it enables accurate similarity comparisons for semantic search applications.

Exam trap

The exam often tests the distinction between generative models (text/image generation) and embedding models (vector representation), leading candidates to confuse a general-purpose LLM like Claude with a specialized embedding model like Amazon Titan Embeddings for semantic search.

How to eliminate wrong answers

Option B is wrong because Stable Diffusion is an image generation model that creates images from text prompts, not a model for generating text embeddings for semantic similarity. Option C is wrong because Anthropic Claude is a text generation model designed for conversational AI and content creation, not for producing vector embeddings that encode semantic relationships between phrases. Option D is wrong because a multimodal model that supports both text and image input is optimized for tasks like image captioning or visual question answering, not for generating embeddings that specifically capture textual semantic similarity.

Full explanation →

528

Multi-Selecthard

A machine learning team is building a binary classifier using Amazon SageMaker. The dataset has 10,000 features and 1,000 samples. The model overfits severely. Which TWO approaches are MOST likely to reduce overfitting? (Choose two.)

Select 2 answers

A.Increase the batch size to the full dataset

B.Use a neural network with more layers

C.Perform feature selection to reduce the number of features

D.Add L2 regularization to the loss function

E.Train the model for more epochs

AnswersC, D

With 10,000 features against only 1,000 samples, the model has far more parameters than observations, so it memorises noise. Feature selection removes irrelevant or redundant predictors, directly reducing the dimensionality that drives the severe overfitting described in the stem.

Why this answer

Option C is correct because with 10,000 features but only 1,000 samples, the model has far more parameters than data points, so performing feature selection to reduce the number of features directly lowers model complexity and the variance that causes overfitting. Option D is correct because adding L2 regularization to the loss function penalizes large weights, shrinking them toward zero and constraining the model's effective capacity, which is a standard, effective remedy for overfitting. Options A, B, and E do not belong: increasing batch size to the full dataset mainly affects gradient noise and optimization dynamics rather than reducing model capacity, adding more layers increases model complexity and would worsen overfitting, and training for more epochs lets the model fit the training data even more closely, also worsening overfitting.

Exam trap

AWS often tests the misconception that increasing batch size or training longer always improves generalization, when in fact these techniques can worsen overfitting in high-dimensional, low-sample scenarios.

Full explanation →

529

MCQeasy

A junior data scientist is building a model to classify incoming customer support tickets into one of eight predefined categories such as Billing, Shipping, or Returns. Historical tickets already have correct category labels. Which type of machine learning is being used?

A.Generative AI fine-tuning

B.Reinforcement learning

C.Unsupervised learning

D.Supervised learning

AnswerD

Supervised learning trains on labeled examples, and here every historical ticket already carries the correct category label, so the model learns a mapping from ticket text to one of the eight predefined classes. This is textbook multi-class classification, a supervised task, which is exactly what the scenario describes.

Why this answer

The presence of correct historical labels plus a fixed set of eight output categories makes this multi-class classification, which belongs to supervised learning. The model is taught from known input-output pairs, then predicts the category for new tickets. Unsupervised, reinforcement, and generative approaches all lack the labeled input-output mapping central to this task.

Exam trap

The trap here is assuming any text-related task must involve generative AI or unsupervised clustering, when predefined labels and fixed categories make it supervised classification.

Full explanation →

530

MCQeasy

A company wants to use AI to automatically transcribe customer service calls into text. Which AWS service is most suitable?

A.Amazon Transcribe

B.Amazon Comprehend

C.Amazon Polly

D.Amazon Rekognition

AnswerA

Amazon Transcribe is a fully managed automatic speech recognition service that converts audio to text, supporting batch and streaming transcription with speaker diarisation. It directly satisfies the requirement to transcribe customer service calls without building custom acoustic models.

Why this answer

Amazon Transcribe is the correct choice because it is a fully managed automatic speech recognition (ASR) service designed specifically to convert speech into text. It can handle real-time streaming or batch processing of audio files, making it ideal for transcribing customer service calls into searchable text.

Exam trap

The trap here is that candidates often confuse Amazon Transcribe (speech-to-text) with Amazon Polly (text-to-speech) or assume Amazon Comprehend can process audio directly, when in fact Comprehend only works on text input.

How to eliminate wrong answers

Option B is wrong because Amazon Comprehend is a natural language processing (NLP) service used for extracting insights like sentiment, entities, and key phrases from text, not for transcribing audio. Option C is wrong because Amazon Polly is a text-to-speech (TTS) service that converts text into lifelike speech, the opposite of the required speech-to-text functionality. Option D is wrong because Amazon Rekognition is a computer vision service for analyzing images and videos, such as object detection and facial recognition, and has no capability to process audio or transcribe speech.

Full explanation →

531

Multi-Selecteasy

A retail company is deploying a machine learning model to analyze customer reviews and predict sentiment. The team wants to follow responsible AI guidelines to ensure fairness, transparency, and accountability. Which TWO actions should the team take? (Choose TWO.)

Select 2 answers

A.Use SageMaker Debugger to optimize training performance.

B.Use SageMaker Clarify to evaluate bias in the training data.

C.Use SageMaker Model Monitor to automatically retrain the model when drift is detected.

D.Use Amazon Rekognition to detect personally identifiable information (PII) in the review text.

E.Use SageMaker Model Cards to document the model's intended use, limitations, and evaluation results.

AnswersB, E

SageMaker Clarify detects statistical bias across training data and features, directly satisfying the fairness requirement by quantifying disparate impact before deployment. It also generates explainability reports showing which features drive predictions, supporting transparency and accountability for the sentiment model's outputs.

Why this answer

Option B is correct because SageMaker Clarify is the AWS service specifically designed to detect potential bias in training data and models, providing bias metrics (such as class imbalance and disparate impact) that directly support the fairness pillar of responsible AI. Option E is correct because SageMaker Model Cards provide a structured way to document a model's intended use, limitations, evaluation results, and risk information, which directly supports transparency and accountability requirements. Option A is not correct because SageMaker Debugger focuses on training performance and convergence issues (tensor analysis, profiling), not on fairness, transparency, or accountability.

Option C is not correct because SageMaker Model Monitor detects data and model drift and can trigger retraining, but it addresses operational model quality rather than the responsible AI goals of fairness and transparency. Option D is not correct because Amazon Rekognition is a computer vision service for image and video analysis and cannot detect PII in review text; Amazon Comprehend would be the appropriate text-based service.

Exam trap

AWS often tests the distinction between monitoring for operational drift (Model Monitor) and evaluating for ethical bias (Clarify), leading candidates to confuse Model Monitor's drift detection with fairness analysis.

Full explanation →

532

MCQeasy

A content team wants to build an internal tool that drafts blog posts from short outlines. They have no machine learning engineers on staff and want AWS to manage the underlying model infrastructure while they focus only on prompts and output. Which AWS service should they use?

A.Amazon SageMaker AI

B.Amazon Polly

C.Amazon Bedrock

D.Amazon Rekognition

AnswerC

Amazon Bedrock is a fully managed service that exposes foundation models from multiple providers through a single API, so a team with no ML engineers can send prompts and receive generated text without provisioning or tuning any infrastructure. It fits the drafting use case directly, since the team only needs prompt design and output handling rather than model hosting.

Why this answer

The team needs managed access to generative foundation models with minimal operational overhead, and Amazon Bedrock provides exactly that by serving multiple providers' models through one API without requiring the team to train or host anything. SageMaker AI offers more control but demands ML engineering, while Polly and Rekognition serve speech and vision tasks respectively.

Exam trap

The trap here is assuming that any AWS AI service can generate text, when several popular ones such as Polly and Rekognition are specialized for speech or image tasks only.

Full explanation →

533

MCQmedium

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

A.Fine-tune a base LLM on the policy documents monthly

B.Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store

C.Use a larger foundation model with a longer context window and paste all documents into each prompt

D.Train a custom model from scratch on the policy documents each month

AnswerB

Retrieval-Augmented Generation (RAG) avoids retraining by storing policy documents as vector embeddings in a vector store, then retrieving the most relevant chunks at query time to ground the language model’s response. This satisfies the constraint that documents are updated monthly, as only the vector index needs re-indexing, not the underlying model.

Why this answer

Retrieval-Augmented Generation (RAG) is the most appropriate approach because it allows the chatbot to answer questions based on the latest policy documents without retraining the underlying language model. By indexing the documents in a vector store and retrieving relevant chunks at query time, RAG provides up-to-date, grounded responses while keeping the base LLM static, which is both cost-effective and scalable for monthly document updates.

Exam trap

AWS often tests the misconception that fine-tuning is the only way to incorporate new data, but the trap here is that candidates overlook RAG's ability to handle dynamic, frequently updated knowledge bases without retraining, which is a core AIF-C01 concept in the AI and ML Fundamentals domain.

How to eliminate wrong answers

Option A is wrong because fine-tuning a base LLM monthly on updated policy documents is computationally expensive, time-consuming, and risks catastrophic forgetting of previous knowledge, making it impractical for frequent updates. Option C is wrong because pasting all policy documents into each prompt exceeds typical context window limits (e.g., 4K–32K tokens for most models), leading to truncation, high token costs, and degraded performance due to irrelevant context. Option D is wrong because training a custom model from scratch each month is prohibitively expensive and resource-intensive, requiring massive datasets and compute, which is unnecessary when a pre-trained LLM with RAG can achieve the same goal with far less overhead.

Full explanation →

534

MCQhard

A team is fine-tuning a Meta Llama 2 model on Amazon Bedrock for a legal document classification task. After fine-tuning, the model performs well on the training set but poorly on the validation set. Which adjustment is MOST likely to reduce overfitting?

A.Increase the size of the training dataset and apply dropout

B.Increase the learning rate

C.Add more layers to the model

D.Reduce the number of training epochs

AnswerA

More data helps generalization; dropout randomly drops units during training, reducing overfitting.

Why this answer

Increasing the training dataset size provides more diverse examples, helping the model generalize better, while dropout randomly deactivates neurons during training to prevent co-adaptation, both of which directly combat overfitting. In the context of fine-tuning Meta Llama 2 on Amazon Bedrock, these techniques are standard regularization methods to improve validation performance when the model memorizes the training set.

Exam trap

AWS often tests the misconception that reducing epochs alone is a sufficient fix for overfitting, when in reality, regularization techniques like dropout and data augmentation are more targeted and effective for deep learning models.

How to eliminate wrong answers

Option B is wrong because increasing the learning rate can cause the model to overshoot optimal weights, leading to unstable training or divergence, which does not address overfitting and may worsen validation performance. Option C is wrong because adding more layers increases model capacity, making overfitting more likely on a fixed dataset, not reducing it. Option D is wrong because reducing the number of training epochs may help if the model is overtrained, but it is a coarse adjustment that often underfits the model; the primary cause of overfitting here is insufficient regularization or data, not simply too many epochs, and reducing epochs alone is less effective than combining data augmentation and dropout.

Full explanation →

535

MCQmedium

A data science team is using Amazon SageMaker Studio. To meet compliance requirements, they need to ensure that all user activity in the environment is logged and that any unauthorized access attempts are detected. Which approach should they take?

A.Enable SageMaker Model Monitor and configure Amazon S3 server access logs.

B.Enable AWS CloudTrail and Amazon GuardDuty for threat detection.

C.Use AWS Config rules to track changes and Amazon Inspector for vulnerability scanning.

D.Enable SageMaker Studio with VPC only mode and use AWS CloudTrail.

AnswerB

CloudTrail records all API activity across SageMaker Studio, providing the audit trail compliance requires, while GuardDuty continuously analyses logs and behaviour to detect unauthorised access attempts. Together they satisfy both the logging and threat-detection requirements.

Why this answer

AWS CloudTrail logs all API activity in SageMaker Studio, including user actions and access attempts, while Amazon GuardDuty provides intelligent threat detection by analyzing CloudTrail logs, VPC flow logs, and DNS logs for unauthorized access patterns. Together, they meet compliance requirements for logging and detecting unauthorized access without additional configuration overhead.

Exam trap

The trap here is that candidates often confuse logging (CloudTrail) with threat detection (GuardDuty) and assume that enabling CloudTrail alone satisfies both requirements, but GuardDuty is specifically needed to analyze logs for unauthorized access attempts.

How to eliminate wrong answers

Option A is wrong because SageMaker Model Monitor is designed for detecting data drift and model quality issues, not for logging user activity or detecting unauthorized access; Amazon S3 server access logs only capture requests to S3 buckets, not SageMaker Studio user actions. Option C is wrong because AWS Config rules track resource configuration changes and compliance, not user activity logging, and Amazon Inspector focuses on vulnerability scanning of EC2 instances and container images, not threat detection for user access. Option D is wrong because VPC only mode restricts network access but does not provide logging of user activity or threat detection; AWS CloudTrail alone logs API calls but lacks the intelligent threat detection capability that GuardDuty provides for identifying unauthorized access attempts.

Full explanation →

536

Multi-Selectmedium

Which TWO actions can help reduce bias in a foundation model’s outputs? (Choose two.)

Select 2 answers

A.Fine-tune the model on a balanced, representative dataset

B.Use careful prompt engineering with neutral wording

C.Restrict model access to a subset of users

D.Increase temperature to add randomness

E.Use a larger foundation model

AnswersA, B

Fine-tuning on a balanced, representative dataset corrects skewed statistical associations in the training data, so the model's outputs reflect fairer distributions across demographic groups. This directly targets the root cause of bias rather than merely masking its symptoms at inference time.

Why this answer

Option A is correct because fine-tuning on a balanced, representative dataset directly addresses bias by exposing the model to fair, diverse examples during training, which adjusts its learned weights and reduces skewed associations in its outputs. Option B is correct because careful prompt engineering with neutral wording avoids injecting biased framing or leading context into the input, so the model is less likely to amplify stereotypes or produce skewed responses. Option C is incorrect because restricting access to a subset of users is an access-control measure that does nothing to change the model's inherent biases.

Option D is incorrect because increasing temperature only adds randomness to token sampling, which can make outputs more erratic rather than less biased. Option E is incorrect because a larger foundation model does not inherently reduce bias and may even reproduce or amplify biases present in its larger training corpus.

Exam trap

AWS AI Practitioner exam often tests the misconception that increasing randomness (temperature) or model size can inherently fix bias, when in fact these changes do not address the underlying data or prompt-level causes of biased outputs.

Full explanation →

537

Multi-Selecthard

A company is using Amazon Fraud Detector to detect fraudulent transactions. Which TWO actions can be taken to improve model accuracy? (Select TWO.)

Select 2 answers

A.Increase the volume of event data

B.Deploy the model to multiple endpoints

C.Use a different detector type

D.Use a different model version

E.Select event variables that are more predictive

AnswersA, E

Fraud Detector's models learn from historical event data, so a larger volume of labelled fraud and legitimate events gives the algorithm more signal to distinguish patterns. This directly addresses the accuracy constraint by reducing overfitting to sparse samples.

Why this answer

Option A is correct because Amazon Fraud Detector's model accuracy improves with more historical event data — a larger volume of labeled fraud/legitimate events gives the training algorithm more examples to learn patterns from, reducing overfitting and improving predictions. Option E is correct because model accuracy depends heavily on feature quality; choosing event variables with stronger predictive power (e.g., customer age, order price, IP address, email domain) directly improves the model's ability to distinguish fraudulent from legitimate transactions. Option B is incorrect because deploying a model to multiple endpoints only affects availability/scalability of predictions, not the underlying model's accuracy.

Option C is incorrect because changing the detector type (e.g., online fraud, transaction fraud) alters the use case rather than inherently improving accuracy for the existing scenario. Option D is incorrect because switching to a different model version simply selects an already-trained model; it does not by itself improve accuracy unless retraining with better data or variables occurs.

Exam trap

The AIF-C01 exam often tests the misconception that changing model versions or detector types alone improves accuracy, when in reality accuracy improvements require data or feature enhancements.

Full explanation →

538

Multi-Selectmedium

Which TWO actions are recommended for improving the factual accuracy of a foundation model's responses when using RAG?

Select 2 answers

A.Include relevant context from the knowledge base in the prompt

B.Increase the max_tokens parameter

C.Provide clear instructions in the system prompt

D.Use the largest foundation model available

E.Increase the temperature parameter

AnswersA, C

Injecting retrieved passages into the prompt gives the model authoritative source text to condition on, so answers cite supplied evidence rather than relying on parametric memory. This grounds generation in the knowledge base, directly improving factual accuracy.

Why this answer

Option A is correct because RAG's core mechanism is retrieving relevant passages from the knowledge base and injecting them into the prompt as grounding context, which directly supplies the model with the facts it needs and reduces hallucination. Option C is correct because clear system-prompt instructions (for example, telling the model to answer only from the provided context and to say it doesn't know when the context is insufficient) constrain the model's behavior and improve factual fidelity. Option B is not recommended because max_tokens only caps response length and does not affect factual grounding.

Option D is not recommended because a larger model does not guarantee factual accuracy and does not address retrieval quality. Option E is not recommended because raising temperature increases randomness and creativity, which typically worsens factual accuracy.

Exam trap

AWS often tests the misconception that larger models or higher randomness (temperature) inherently improve response quality, but in RAG, factual accuracy depends on retrieval quality and prompt engineering, not model size or creativity parameters.

Full explanation →

539

MCQmedium

A company wants to build a customer support chatbot that answers questions based on a large internal knowledge base. Which AWS service is most suitable for implementing RAG to retrieve relevant documents?

A.Amazon Lex

B.Amazon Polly

C.Amazon Connect

D.Amazon Kendra

AnswerD

Amazon Kendra is an ML-powered enterprise search service that indexes internal document repositories and returns semantically relevant passages, which the chatbot feeds to the model as RAG context. It satisfies the requirement for retrieving relevant documents from a large knowledge base.

Why this answer

Amazon Kendra is an intelligent enterprise search service powered by machine learning that is specifically designed for retrieving relevant documents from large knowledge bases, making it ideal for RAG implementations. It supports natural language queries and returns precise answers with citations, which is exactly what a customer support chatbot needs to retrieve relevant documents.

Exam trap

AIF-C01 often tests the confusion between services that build chatbots (Lex) and services that retrieve documents (Kendra), and candidates may pick Lex because it sounds like a chatbot service.

How to eliminate wrong answers

Option A is wrong because Amazon Lex is a service for building conversational interfaces (chatbots) using voice and text, but it does not provide document retrieval capabilities for RAG. Option B is wrong because Amazon Polly is a text-to-speech service, not a document retrieval service. Option C is wrong because Amazon Connect is a cloud contact center service, not a document retrieval or search service.

Full explanation →

540

MCQeasy

A marketing team uses Amazon Bedrock to generate promotional copy for a global campaign. A reviewer discovers that some outputs include biased stereotypes about certain nationalities. The team wants a configurable control that detects and blocks this category of harmful content before the copy reaches reviewers. Which Amazon Bedrock feature should they configure?

A.Amazon Bedrock model evaluation jobs that compare foundation models on benchmark datasets

B.Amazon CloudWatch alarms on the model's invocation latency

C.Amazon Bedrock Guardrails content filters for hate and insult categories

D.Amazon Bedrock provisioned throughput to reserve dedicated model capacity

AnswerC

Content filters in Amazon Bedrock Guardrails detect and block harmful categories such as hate, insults, sexual content, and violence, with configurable strength per category. Biased stereotypical statements about nationalities fall under hate and insult detection, so enabling and tuning these filters directly prevents such copy from reaching reviewers and aligns with responsible AI content controls.

Why this answer

The requirement is an inline, configurable control that detects and blocks harmful stereotyped content at generation time. Amazon Bedrock Guardrails content filters provide exactly that, with tunable strength for hate and insult categories. Throughput reservation, latency alarms, and offline model evaluation jobs address capacity, operations, and model selection rather than runtime content safety.

Exam trap

The trap here is confusing model evaluation, which scores models before deployment, with guardrail content filters, which block harmful content on every live response.

Full explanation →

541

Multi-Selectmedium

A data scientist is using a foundation model to summarize long documents. Which TWO of the following steps are most likely to improve the quality of the summaries?

Select 2 answers

A.Break the input document into chunks and summarize each chunk separately.

B.Use a high temperature parameter to increase creativity.

C.Provide few-shot examples of desired summaries in the prompt.

D.Use a low frequency penalty to reduce repetition.

E.Use a longer context length by increasing the max tokens parameter.

AnswersA, C

Chunking splits the long document into segments that fit the model's context window, letting each chunk be summarised without truncation. This satisfies the long-document constraint, since whole-document input would exceed context limits and lose detail, degrading summary quality.

Why this answer

Option A is correct because chunking a long document and summarizing each chunk separately keeps each request within the model's effective context window, avoiding truncation and the degraded recall that occurs when a foundation model must attend to very long inputs, and the chunk summaries can then be combined into a final summary. Option C is correct because few-shot prompting supplies concrete examples of the desired summary style, length, and level of detail, which steers the model's output distribution toward the target format more reliably than a bare instruction. Option B is not appropriate because a high temperature increases randomness and creativity, which harms factual fidelity and consistency in summarization.

Option D is not the priority here because frequency penalty only discourages repeated tokens and does not address the core problems of long-input context limits or output style alignment. Option E is not correct because increasing max tokens only raises the output length cap; it does not extend the model's usable context window for the input document and can even encourage overly long, unfocused summaries.

Exam trap

AWS often tests the misconception that increasing max tokens extends the model's input capacity, when in reality it only controls the output length, while the input is constrained by the model's inherent context window.

Full explanation →

542

MCQeasy

Which metric is commonly used to evaluate the quality of a text summarization model by comparing the generated summary with a reference summary, measuring the overlap of n-grams?

A.Perplexity

B.BLEU

C.ROUGE

D.BERTScore

AnswerC

ROUGE scores summarisation by counting overlapping n-grams between the generated summary and the reference summary, reporting recall-oriented precision. That direct n-gram comparison is exactly the overlap metric the question describes, unlike embedding-based or classification metrics.

Why this answer

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is the standard metric for evaluating text summarization by measuring the overlap of n-grams (e.g., ROUGE-1, ROUGE-2, ROUGE-L) between a generated summary and a reference summary. It focuses on recall, capturing how much of the reference content is preserved in the generated summary, making it ideal for summarization tasks.

Exam trap

The AWS AI Practitioner exam often tests the distinction between BLEU (precision-focused, for translation) and ROUGE (recall-focused, for summarization), and candidates mistakenly choose BLEU because both involve n-gram overlap, but BLEU is not the primary metric for summarization quality.

How to eliminate wrong answers

Option A is wrong because Perplexity measures how well a language model predicts a sequence of tokens, typically used for evaluating language models (e.g., GPT) rather than comparing generated summaries to reference summaries via n-gram overlap. Option B is wrong because BLEU (Bilingual Evaluation Understudy) is designed for machine translation, emphasizing precision of n-gram matches, which is less suitable for summarization where recall of key content is critical. Option D is wrong because BERTScore uses contextual embeddings from BERT to compute semantic similarity between tokens, not n-gram overlap, making it a semantic metric rather than a lexical n-gram-based one.

Full explanation →

543

MCQmedium

A team deployed a text generation model on Amazon Bedrock. They want to monitor for toxic content in model outputs. Which evaluation approach is MOST effective?

A.Enable CloudWatch Logs and set a metric filter for toxic words

B.Use Amazon SageMaker Ground Truth for human annotation

C.Manually review a sample of outputs each week

D.Use Amazon Bedrock Model Evaluation with toxicity metrics

AnswerD

Bedrock Model Evaluation runs automatic toxicity metrics against model outputs, directly satisfying the requirement to monitor generated text for toxic content. It provides scored, repeatable assessment rather than ad-hoc inspection, making it the most effective ongoing evaluation approach for this deployment.

Why this answer

Amazon Bedrock Model Evaluation with toxicity metrics is the most effective approach because it provides automated, built-in evaluation of model outputs for toxic content using predefined metrics, directly integrated with the Bedrock service. This eliminates the need for manual effort or custom filtering, ensuring consistent and scalable monitoring of harmful content.

Exam trap

The trap here is that candidates may choose CloudWatch metric filters (Option A) because they associate monitoring with logs, but fail to recognize that toxicity detection requires semantic understanding beyond simple keyword matching.

How to eliminate wrong answers

Option A is wrong because CloudWatch Logs with a metric filter for toxic words is a simplistic, keyword-based approach that cannot detect nuanced or context-dependent toxicity, such as sarcasm or implicit hate speech, and requires manual setup of word lists. Option B is wrong because Amazon SageMaker Ground Truth for human annotation is designed for creating labeled datasets, not for real-time or automated monitoring of model outputs, and introduces latency and cost overhead. Option C is wrong because manually reviewing a sample of outputs each week is not scalable, introduces human bias, and fails to provide continuous or real-time monitoring, making it ineffective for production systems.

Full explanation →

544

MCQmedium

A company is developing an AI system that generates news articles. To comply with transparency regulations, they must clearly indicate when content is AI-generated. Which action should they take?

A.Register the model with a government agency

B.Remove any identifiable information about the model from the output

C.Use a watermark that is invisible to users

D.Include a disclaimer in the article metadata and a visible label stating 'This content was generated by AI'

AnswerD

Labelling both the article metadata and the visible page satisfies transparency rules requiring disclosure that content is machine-generated. Metadata supports automated detection and audit, while the visible label informs human readers, covering both machine-readable and human-readable disclosure obligations.

Why this answer

Transparency regulations (such as the EU AI Act's Article 50 and similar disclosure requirements) mandate that AI-generated content be clearly identifiable to the audience. Option D satisfies this by combining a machine-readable disclaimer in the article metadata with a human-readable visible label, covering both automated detection and end-user awareness. This dual approach is the standard compliance pattern for generative AI disclosure.

Exam trap

AIF-C01 often tests the misconception that invisible watermarking alone satisfies transparency requirements, when regulations actually demand perceivable disclosure to end users.

How to eliminate wrong answers

Option A is wrong because registering a model with a government agency is not a transparency mechanism for content consumers and is not required by content-labeling regulations. Option B is wrong because removing identifiable model information actually reduces transparency and obscures provenance rather than disclosing AI generation. Option C is wrong because an invisible watermark alone does not satisfy transparency requirements that call for clear, perceivable disclosure to users — it only aids forensic detection.

Full explanation →

545

MCQeasy

A company wants to automatically detect anomalies in their AWS CloudTrail logs to identify potential security threats. Which AWS service is specifically designed for this purpose?

A.Amazon Macie

B.AWS Config

C.Amazon GuardDuty

D.Amazon Inspector

AnswerC

GuardDuty is a managed threat-detection service that continuously analyses CloudTrail management events, VPC Flow Logs and DNS logs using machine learning and threat intelligence, surfacing anomalous or malicious activity without agents — precisely the automated anomaly detection the company needs.

Why this answer

Amazon GuardDuty is a threat detection service that continuously monitors AWS accounts and workloads using machine learning, anomaly detection, and integrated threat intelligence. It specifically analyzes CloudTrail management and data events, VPC Flow Logs, and DNS logs to identify unauthorized behavior or potential security threats, making it the correct choice for automatically detecting anomalies in CloudTrail logs.

Exam trap

The AIF-C01 exam often tests the distinction between services that detect threats (GuardDuty) versus services that protect data (Macie), assess vulnerabilities (Inspector), or track configuration compliance (Config), leading candidates to confuse their primary use cases.

How to eliminate wrong answers

Option A is wrong because Amazon Macie is a data security and data privacy service that uses machine learning to discover, classify, and protect sensitive data stored in Amazon S3, not to analyze CloudTrail logs for security threats. Option B is wrong because AWS Config is a service that evaluates and records resource configurations and compliance against desired policies, not designed for real-time anomaly detection in log data. Option D is wrong because Amazon Inspector is a vulnerability management service that scans EC2 instances and container images for software vulnerabilities and unintended network exposure, not for analyzing CloudTrail logs.

Full explanation →

546

Multi-Selecteasy

Which TWO of the following are types of feature scaling?

Select 2 answers

A.One-hot encoding

B.Principal Component Analysis (PCA)

C.Standardization

D.Binning

E.Normalization (Min-Max)

AnswersC, E

Standardization rescales each feature to zero mean and unit variance by subtracting the mean and dividing by the standard deviation, a distinct scaling technique alongside normalization, which instead maps values into a fixed range such as 0 to 1.

Why this answer

Standardization (C) is a feature-scaling technique that transforms each feature to have a mean of 0 and a standard deviation of 1 (z-score, (x − μ)/σ), which is exactly what the question asks for. Normalization (Min-Max) (E) is also a feature-scaling method that rescales values into a fixed range, typically [0, 1], via (x − min)/(max − min). One-hot encoding (A) is a categorical-encoding technique that creates binary columns, not a scaling method.

Principal Component Analysis (B) is a dimensionality-reduction technique that projects data onto principal components, not a scaling method. Binning (D) is a discretization technique that groups continuous values into bins, not a scaling method.

Exam trap

AWS often tests the distinction between feature scaling (changing the numeric range of features) and data transformation techniques like encoding or dimensionality reduction, leading candidates to confuse one-hot encoding or PCA with scaling methods.

Full explanation →

547

MCQeasy

A company is using Amazon Comprehend to analyze customer feedback. They need to ensure that the documents are encrypted at rest. What should they do?

A.No action is needed; Amazon Comprehend automatically encrypts data at rest using AES-256

B.Enable encryption using AWS KMS in the Comprehend console

C.Store documents in an encrypted S3 bucket and use a VPC endpoint

D.Use SSL/TLS for all API calls to Comprehend

AnswerA

Amazon Comprehend encrypts data at rest by default using AES-256, with AWS-managed keys, so no customer configuration is required to meet the encryption requirement. This default behaviour satisfies the constraint without additional key management or bucket policies.

Why this answer

Amazon Comprehend automatically encrypts all data at rest using AES-256 encryption by default, with no additional configuration required. This encryption covers both the documents processed by the service and any models or artifacts stored internally. Therefore, no action is needed from the customer to enable encryption at rest.

Exam trap

The trap here is that candidates often assume they need to manually enable encryption or use KMS, but Amazon Comprehend enforces encryption at rest automatically with no user action required, making 'No action needed' the correct answer.

How to eliminate wrong answers

Option B is wrong because Amazon Comprehend does not expose a console option to enable or disable encryption via AWS KMS; encryption is always-on and managed by the service. Option C is wrong because while storing documents in an encrypted S3 bucket is a best practice for data in transit to Comprehend, it does not affect how Comprehend encrypts data at rest within its own storage; the service already encrypts at rest regardless of the source bucket's encryption. Option D is wrong because SSL/TLS protects data in transit, not data at rest, and is already enforced by Comprehend for API calls.

Full explanation →

548

MCQhard

A team is using Amazon SageMaker to train a deep learning model. The training job is taking too long. Which action is MOST likely to reduce training time while maintaining model quality?

A.Use incremental training

B.Enable managed spot training

C.Switch to a smaller instance type

D.Increase the number of training instances for distributed training

AnswerD

Distributed training spreads gradient computation across multiple instances, so each epoch processes more data in parallel and wall-clock training time falls. Adding instances scales throughput while preserving model quality, since the same algorithm and hyperparameters are used across the cluster.

Why this answer

Increasing the number of training instances for distributed training (Option D) directly reduces training time by parallelizing the workload across multiple machines, which is the most effective approach for deep learning models that are computationally intensive. SageMaker's distributed training libraries (e.g., SageMaker Distributed Data Parallel) split the data and model across instances, enabling linear or near-linear speedups while preserving model quality through synchronized gradient updates.

Exam trap

A common misconception is that using Spot Instances reduces training time, but it actually only reduces cost and can increase time due to interruptions. Distributed training with SageMaker's data parallelism is the correct approach for reducing wall-clock time.

How to eliminate wrong answers

Option A is wrong because incremental training (e.g., using a pre-trained model and fine-tuning on new data) does not reduce the time for a single training job from scratch; it is designed for iterative updates on new data, not for speeding up a long-running initial training job. Option B is wrong because enabling managed spot training reduces cost by using spare EC2 capacity, but it does not inherently reduce training time; in fact, spot instances can be interrupted, potentially increasing total time due to checkpointing and restarts. Option C is wrong because switching to a smaller instance type reduces computational resources, which typically increases training time rather than reducing it, and may degrade model quality if the instance lacks sufficient memory or compute for the model.

Full explanation →

549

MCQhard

A team is building a real-time anomaly detection system for IoT sensor data. The data is unlabeled, and the team expects the anomalies to be rare but of high importance. Which combination of approach and AWS service should the team use?

A.Use Amazon Rekognition Custom Labels to detect anomalies in sensor images

B.Train a supervised classifier using Amazon SageMaker; use Amazon Kinesis Data Analytics for real-time inference

C.Apply k-means clustering with Amazon SageMaker; deploy the model as a real-time endpoint

D.Use a semi-supervised one-class classifier with Amazon Lookout for Equipment

AnswerD

Amazon Lookout for Equipment trains one-class models on normal sensor telemetry alone, flagging deviations without labelled anomaly examples — matching the unlabeled, rare-anomaly constraint. It ingests multivariate time-series data directly from IoT sensors, so the semi-supervised one-class approach fits the scenario's requirement for detecting high-importance, infrequent faults.

Why this answer

Amazon Lookout for Equipment is purpose-built for anomaly detection on unlabeled industrial sensor data. It uses a semi-supervised one-class classifier that learns the normal operating envelope from historical data and flags rare, high-importance anomalies in real time, matching the exact requirements of the scenario.

Exam trap

AWS often tests the distinction between supervised, unsupervised, and semi-supervised learning in the context of unlabeled data, and the trap here is that candidates may choose k-means clustering (option C) thinking any unsupervised method works, but k-means is not optimized for rare anomaly detection and lacks the one-class modeling that Lookout for Equipment provides.

How to eliminate wrong answers

Option A is wrong because Amazon Rekognition Custom Labels is designed for image classification and object detection, not for analyzing time-series sensor data. Option B is wrong because supervised classifiers require labeled training data, which the team does not have, and Amazon Kinesis Data Analytics is a streaming SQL engine, not a real-time inference service for custom ML models. Option C is wrong because k-means clustering is an unsupervised method that groups data into clusters but does not inherently detect rare anomalies; it requires post-processing to identify outliers, and deploying it as a real-time endpoint adds unnecessary complexity compared to a purpose-built service.

Full explanation →

550

MCQeasy

A retail company is deploying a chatbot to handle customer inquiries. During testing, they notice the chatbot occasionally uses offensive language when responding to certain user inputs. Which responsible AI principle is being violated?

A.Privacy

B.Transparency

C.Fairness

D.Accountability

AnswerC

Fairness ensures AI systems treat all users equitably; offensive language is a fairness issue.

Why this answer

The chatbot's use of offensive language in responses indicates a bias in the model's training data or behavior, which violates the Fairness principle. Fairness in responsible AI requires that systems do not discriminate against or harm individuals or groups based on protected characteristics, and generating offensive outputs directly undermines this. This is distinct from privacy (data protection), transparency (explainability), or accountability (responsibility for outcomes).

Exam trap

AWS often tests the distinction between Fairness and other principles by presenting a scenario where the harm is not about data leaks (Privacy), lack of explanation (Transparency), or who is responsible (Accountability), but rather about the system producing discriminatory or harmful content.

How to eliminate wrong answers

Option A is wrong because Privacy concerns the protection of personal data and user consent, not the generation of offensive language. Option B is wrong because Transparency focuses on making AI decisions understandable and explainable, not on preventing biased or harmful outputs. Option D is wrong because Accountability refers to assigning responsibility for AI system outcomes and governance, not the specific ethical violation of producing offensive content.

Full explanation →

551

MCQeasy

A team trained a deep learning model that achieves 99% accuracy on training data but only 70% on validation data. What is the most likely issue?

A.Underfitting

B.Overfitting

C.Data leakage

D.Feature scaling

AnswerB

The large gap between 99% training accuracy and 70% validation accuracy shows the model memorised training data rather than learning generalisable patterns. Overfitting is the specific condition producing this divergence, matching the stem's reported metrics exactly.

Why this answer

The model performs exceptionally well on training data (99% accuracy) but significantly worse on validation data (70% accuracy). This large gap indicates the model has memorized the training data, including noise and irrelevant patterns, rather than learning generalizable features — a classic symptom of overfitting.

Exam trap

The AIF-C01 exam often tests the distinction between overfitting and underfitting by presenting a scenario where training accuracy is high but validation accuracy is low, tempting candidates to incorrectly choose underfitting if they focus only on the low validation score.

How to eliminate wrong answers

Option A is wrong because underfitting would show poor performance on both training and validation data, not high training accuracy with low validation accuracy. Option C is wrong because data leakage typically causes both training and validation accuracy to be artificially high, not a large gap between them. Option D is wrong because feature scaling issues would generally affect model convergence or performance uniformly across datasets, not create a specific training-validation accuracy disparity.

Full explanation →

552

Multi-Selectmedium

A healthcare company is building a medical diagnosis assistant using Amazon Bedrock. They need to ensure the model’s responses are based on the latest medical research and do not include outdated information. The company also wants to minimize costs. Which TWO actions should they take? (Select TWO)

Select 2 answers

A.Use the Converse API to maintain conversation history

B.Use a smaller model like Amazon Titan Text Lite to reduce inference costs

C.Implement RAG by indexing the latest medical journals in a vector store

D.Fine-tune a large model on the latest medical data monthly

E.Choose a model with the largest context window to include all research in the prompt

AnswersB, C

Amazon Titan Text Lite cuts inference cost per token, satisfying the stated cost-minimisation constraint. However, a smaller model alone cannot ground responses in current medical research, so it must be paired with Retrieval Augmented Generation against an updated knowledge base to prevent outdated output.

Why this answer

Option B is correct because Amazon Titan Text Lite is a smaller, lower-cost model, so using it for inference directly reduces per-token costs, which aligns with the company's goal to minimize costs. Option C is correct because Retrieval Augmented Generation (RAG) retrieves relevant passages from an up-to-date vector store of the latest medical journals and injects them into the prompt, grounding responses in current research and avoiding outdated information without retraining the model. Option A is not correct because the Converse API only manages conversation history and does not by itself ensure access to the latest medical research or reduce costs.

Option D is not correct because monthly fine-tuning of a large model is expensive, slow, and risks stale knowledge between training cycles, conflicting with the cost-minimization goal. Option E is not correct because stuffing all research into a large context window increases token usage and cost and does not guarantee retrieval of the most current, relevant information.

Exam trap

The trap here is that candidates sometimes assume they need a large model like Claude or Titan Text Express for clinical accuracy, but with RAG, a smaller model like Amazon Titan Text Lite can generate reliable responses using retrieved data, reducing costs. Also, fine-tuning is expensive and not needed if the model uses up-to-date external sources.

Full explanation →

553

MCQmedium

A machine learning practitioner is training a model to forecast product demand and observes that the model performs well on training data but poorly on unseen data. Which of the following is the MOST likely cause?

A.High bias-variance tradeoff favoring bias

B.Underfitting due to model being too simple

C.Data leakage causing artificially high training scores

D.Overfitting due to excessive model complexity

AnswerD

Excessive model complexity lets the network memorise training examples, including noise, so training performance stays high while generalisation to unseen demand data degrades. The gap between strong training results and poor unseen-data results directly indicates overfitting rather than underfitting or data leakage.

Why this answer

The model performs well on training data but poorly on unseen data, which is the classic symptom of overfitting. Overfitting occurs when the model is excessively complex (e.g., too many parameters, deep neural networks with high capacity) and learns noise and random fluctuations in the training data rather than the underlying pattern. This leads to high variance, causing excellent training performance but poor generalization to new data.

Exam trap

The AWS AI Practitioner exam often tests the distinction between overfitting and underfitting by describing performance on training vs. unseen data; the trap here is that candidates may confuse 'high bias' (underfitting) with 'high variance' (overfitting) when the model performs well on training data but poorly on test data.

How to eliminate wrong answers

Option A is wrong because a high bias-variance tradeoff favoring bias would cause underfitting, not overfitting; the model would perform poorly on both training and unseen data. Option B is wrong because underfitting due to a model being too simple would result in poor performance on training data as well, not good training performance. Option C is wrong because data leakage artificially inflates both training and test scores, not just training scores; it would cause the model to appear good on unseen data during evaluation, contradicting the scenario of poor performance on unseen data.

Full explanation →

554

Multi-Selecthard

A company uses Amazon SageMaker to build and deploy models. They want to enforce compliance that all model endpoints are encrypted in transit and use least privilege access. Which THREE steps should they take? (Choose THREE.)

Select 3 answers

A.Configure the SageMaker endpoint to use a custom SSL certificate via AWS Certificate Manager

B.Use an interface VPC endpoint (AWS PrivateLink) for SageMaker

C.Attach an IAM policy to the execution role that only allows specific actions on the endpoint

D.Enable AWS CloudTrail to log all endpoint invocations

E.Disable root access on the SageMaker notebook instances

AnswersA, B, C

This ensures HTTPS for encryption in transit.

Why this answer

Configuring a SageMaker endpoint to use a custom SSL certificate from AWS Certificate Manager (ACM) ensures that all data transmitted between clients and the endpoint is encrypted in transit using TLS. This enforces the compliance requirement for encryption in transit by replacing the default SageMaker certificate with a customer-managed certificate, which can be validated and rotated as needed.

Exam trap

The trap here is that candidates often confuse logging (CloudTrail) with enforcement of encryption or access control, or they mistakenly think disabling root access on notebooks affects endpoint security, when in fact it only secures the development environment.

Full explanation →

555

MCQmedium

A company deploys a large language model to automatically generate product descriptions. They want to ensure customers are aware that the content is AI-generated, as part of transparency requirements. What should they implement?

A.Include a disclosure statement such as 'This content was generated by AI' in the output

B.Use Amazon Rekognition to add a visible watermark to images only

C.Store metadata in the content database indicating AI generation

D.Embed an invisible watermark in the generated text

AnswerA

A disclosure statement directly satisfies the transparency requirement by informing customers the text is AI-generated. This is the standard mechanism for meeting AI transparency obligations, as it makes the artificial origin of the content explicit to the end user at the point of consumption.

Why this answer

Transparency requirements for AI-generated content are best met by clearly disclosing to the end user that the content was generated by AI. Including a visible disclosure statement in the output directly informs customers at the point of consumption. This satisfies emerging AI transparency regulations and ethical guidelines.

Exam trap

AIF-C01 often tests the difference between visible disclosure (user awareness) and invisible watermarking (provenance) — candidates pick the technical-sounding watermark option when the requirement is customer awareness.

How to eliminate wrong answers

Option B is wrong because Amazon Rekognition is an image/video analysis service and watermarks images only; it does not address text-based product descriptions or provide the required disclosure. Option C is wrong because storing metadata in a database is not visible to customers and does not fulfill transparency obligations to end users. Option D is wrong because an invisible watermark is not perceivable by customers and therefore does not satisfy the requirement that customers be aware the content is AI-generated.

Full explanation →

556

MCQeasy

A developer is building a customer-facing chatbot using Amazon Bedrock. To ensure the chatbot does not generate offensive or inappropriate content, which AWS feature should they implement?

A.AWS Identity and Access Management (IAM) policies

B.Amazon Bedrock Guardrails

C.Prompt engineering with system prompts

D.Increasing the model temperature parameter

AnswerB

Amazon Bedrock Guardrails applies configurable content filters and denied-topic policies that intercept harmful prompts and responses at inference time. This directly satisfies the requirement to stop offensive or inappropriate output in a customer-facing chatbot, independent of the underlying foundation model.

Why this answer

Amazon Bedrock Guardrails is the correct choice because it provides configurable safeguards that allow developers to define denied topics, content filters (e.g., hate, insults, sexual content), and sensitive information filters to prevent the model from generating offensive or inappropriate responses. Unlike prompt engineering or parameter tuning, Guardrails enforce policy-based constraints at inference time, independent of the underlying model's behavior.

Exam trap

A common pitfall in this exam is assuming that prompt engineering with system prompts or adjusting model parameters like temperature can reliably block offensive content. While these techniques can influence behavior, they do not provide enforceable, policy-based safeguards. Amazon Bedrock Guardrails must be used to define and enforce content filters, denied topics, and sensitive information filters at inference time.

How to eliminate wrong answers

Option A is wrong because IAM policies control access to AWS resources and API actions, not the content generated by a model; they cannot filter or block specific output text. Option C is wrong because prompt engineering with system prompts can guide model behavior but is not a guaranteed or enforceable mechanism—models can still bypass or ignore system instructions, especially under adversarial inputs. Option D is wrong because increasing the temperature parameter increases randomness in token selection, which may actually increase the likelihood of generating inappropriate content, not reduce it.

Full explanation →

557

MCQhard

During a security review, it is found that an Amazon SageMaker notebook instance has outbound internet access, which could lead to data exfiltration. The notebook must only access resources within the VPC. Which step should be taken to restrict internet access?

A.Modify the notebook instance's IAM role to deny s3:GetObject

B.Attach a security group that denies all outbound traffic to 0.0.0.0/0

C.Configure the notebook instance in a VPC with no internet gateway or NAT device, and set the notebook's 'Direct Internet Access' option to 'Disabled'

D.Disable the SageMaker notebook instance's root volume encryption

AnswerC

Placing the notebook in a VPC subnet with no internet gateway or NAT device removes any route to the internet, and disabling Direct Internet Access prevents SageMaker from provisioning its own connectivity, confining traffic to VPC resources.

Why this answer

Disabling 'Direct Internet Access' on a SageMaker notebook instance and placing it in a VPC without an internet gateway or NAT device ensures the notebook cannot reach the public internet. This configuration forces all traffic to stay within the VPC, preventing data exfiltration via outbound internet connections while still allowing access to VPC resources.

Exam trap

The trap here is that candidates may confuse network-level controls (security groups, VPC routing) with IAM permissions, thinking that denying S3 access prevents all exfiltration, or they may incorrectly assume that disabling encryption or blocking all outbound traffic is the correct approach.

How to eliminate wrong answers

Option A is wrong because modifying the IAM role to deny s3:GetObject only restricts access to S3 objects, not outbound internet traffic; data exfiltration could still occur via other protocols (e.g., HTTP, DNS tunneling). Option B is wrong because attaching a security group that denies all outbound traffic to 0.0.0.0/0 would block all outbound traffic, including legitimate VPC resources (e.g., other services within the same VPC), which is overly restrictive and not the intended solution. Option D is wrong because disabling root volume encryption does not affect internet access; it only removes encryption at rest, which is a security risk but unrelated to network egress control.

Full explanation →

558

MCQeasy

A developer is building an application using Amazon Bedrock and needs to ensure that the model's responses do not include any toxic or harmful language. Which Bedrock feature should they configure?

A.Bedrock Knowledge Bases

B.Bedrock Playground

C.Bedrock Agents

D.Bedrock Guardrails

AnswerD

Bedrock Guardrails apply configurable content filters that evaluate prompts and responses, blocking toxic or harmful language before it reaches the application. This directly satisfies the requirement that model responses exclude harmful content, without retraining or prompt engineering.

Why this answer

Bedrock Guardrails provide content filters, topic denial, PII detection, and other safety controls. Knowledge Bases, Agents, and Playground do not directly enforce content safety rules.

Full explanation →

559

MCQeasy

What is the main purpose of a system prompt in a large language model?

A.To increase the temperature for more creative responses

B.To provide an example of the desired output format

C.To list all possible tokens to be used in the response

D.To define the high-level instructions that set the model's behavior and persona

AnswerD

A system prompt supplies the high-level instructions that establish the model's behaviour, persona, tone and boundaries before any user turn. It satisfies the stem's requirement for the main purpose by shaping every subsequent response, unlike user prompts which only govern a single exchange.

Why this answer

The system prompt in a large language model (LLM) defines the high-level instructions that set the model's behavior, persona, and constraints for the entire conversation. Unlike user prompts, which are task-specific, the system prompt establishes the model's role (e.g., 'You are a helpful assistant') and governs how it interprets all subsequent interactions, ensuring consistent and aligned outputs.

Exam trap

AWS often tests the distinction between system prompts (high-level behavior) and user prompts (task-specific instructions), trapping candidates who confuse providing an output format example (few-shot prompting) with the system prompt's role of defining persona and constraints.

How to eliminate wrong answers

Option A is wrong because increasing the temperature parameter controls randomness in token sampling, not the purpose of a system prompt; temperature is a generation hyperparameter, not a prompt function. Option B is wrong because providing an example of the desired output format is the role of a few-shot prompt or user prompt, not the system prompt, which sets overarching behavior rather than specific formatting examples. Option C is wrong because listing all possible tokens is not a prompt function; the model's vocabulary is fixed and defined by its tokenizer, and a system prompt cannot enumerate tokens.

Full explanation →

560

MCQeasy

Which Amazon Titan model is specifically designed to convert text into numerical vectors for use in semantic search and Retrieval-Augmented Generation (RAG)?

A.Amazon Titan Text Express

B.Amazon Titan Multimodal Embeddings

C.Amazon Titan Image Generator

D.Amazon Titan Embeddings

AnswerD

Amazon Titan Embeddings converts text into dense numerical vectors, which semantic search and Retrieval-Augmented Generation rely on for similarity comparison and retrieval. Other Titan variants generate text or images, so they cannot produce the vector representations the scenario requires.

Why this answer

Amazon Titan Embeddings is the correct model because it is specifically designed to convert text into numerical vectors (embeddings) that capture semantic meaning. These vectors are essential for semantic search and Retrieval-Augmented Generation (RAG), where they enable similarity comparisons and efficient retrieval of relevant documents from a vector database.

Exam trap

The trap here is that candidates may confuse 'embedding models' with 'generative models' (like Text Express) or assume that any multimodal model (like Multimodal Embeddings) can handle text-only tasks, but the question explicitly asks for a model designed specifically for text-to-vector conversion.

How to eliminate wrong answers

Option A is wrong because Amazon Titan Text Express is a generative text model designed for tasks like summarization, translation, and conversation, not for producing numerical embeddings. Option B is wrong because Amazon Titan Multimodal Embeddings generates embeddings for both text and images, not exclusively for text as required by the question. Option C is wrong because Amazon Titan Image Generator is a generative model for creating and editing images, not for converting text into vectors.

Full explanation →

561

MCQhard

A security team needs to detect anomalies in AWS CloudTrail logs to identify potential unauthorized access. They want to use machine learning without manually labeling data or training custom models. Which AWS service should they use?

A.Amazon Macie

B.Amazon SageMaker

C.Amazon GuardDuty

D.Amazon Detective

AnswerC

Amazon GuardDuty continuously analyses CloudTrail management and data events, VPC flow logs and DNS logs using AWS-managed threat intelligence and anomaly detection, raising findings without any labelling or model training by the security team, satisfying the no-custom-model constraint.

Why this answer

Amazon GuardDuty is a threat detection service that uses machine learning and anomaly detection to continuously monitor AWS CloudTrail logs, VPC Flow Logs, and DNS logs for unauthorized access and malicious activity. It requires no manual labeling of data or custom model training, as it leverages pre-built ML models and integrated threat intelligence to identify suspicious behavior.

Exam trap

The trap here is that candidates often confuse Amazon GuardDuty with Amazon Detective, assuming Detective performs the initial anomaly detection, when in fact Detective is a post-incident investigation tool that consumes findings from GuardDuty and other sources.

How to eliminate wrong answers

Option A is wrong because Amazon Macie is a data security service that uses ML to discover, classify, and protect sensitive data (e.g., PII) in S3 buckets, not to detect anomalies in CloudTrail logs. Option B is wrong because Amazon SageMaker is a fully managed ML platform for building, training, and deploying custom models, which requires manual data labeling and model training, contrary to the requirement of no manual labeling or custom training. Option D is wrong because Amazon Detective is a security investigation service that analyzes and visualizes security data (including GuardDuty findings) to help root-cause analysis, but it does not perform the initial anomaly detection or use ML to identify threats in real time.

Full explanation →

562

MCQmedium

A developer is building a chatbot using Amazon Bedrock and Claude. They notice that the model sometimes generates harmful or biased responses. Which AWS service can they use to implement guardrails?

A.AWS WAF

B.Amazon GuardDuty

C.AWS Shield

D.Amazon Bedrock Guardrails

AnswerD

Amazon Bedrock Guardrails applies configurable content filters, denied topics and word filters directly to model inference, blocking harmful or biased outputs before they reach users. It satisfies the stem's requirement for guardrails on a Bedrock-hosted Claude chatbot, unlike standalone moderation services that would need custom integration outside the Bedrock invocation flow.

Why this answer

Amazon Bedrock Guardrails is the correct choice because it is a native feature of Amazon Bedrock designed specifically to implement safety controls, content filters, and topic policies for foundation models like Claude. It allows developers to define denied topics, filter harmful content (e.g., hate speech, violence), and redact sensitive information, directly addressing the need to prevent harmful or biased responses in a chatbot built on Bedrock.

Exam trap

The trap here is that candidates may confuse AWS security services (WAF, GuardDuty, Shield) with AI-specific safety mechanisms, assuming any 'guard' or 'shield' service can filter model outputs, when only Amazon Bedrock Guardrails is purpose-built for content safety in generative AI.

How to eliminate wrong answers

Option A is wrong because AWS WAF is a web application firewall that protects HTTP/HTTPS APIs from common web exploits like SQL injection and cross-site scripting, not a service for implementing content guardrails on generative AI model outputs. Option B is wrong because Amazon GuardDuty is a threat detection service that monitors for malicious activity and unauthorized behavior in AWS accounts and workloads, not a tool for filtering or controlling the responses of a large language model. Option C is wrong because AWS Shield is a managed Distributed Denial of Service (DDoS) protection service that safeguards applications against DDoS attacks, and it has no capability to enforce safety policies or bias filters on AI-generated content.

Full explanation →

563

Multi-Selecthard

A healthcare analytics team is evaluating whether a generative AI solution is appropriate for summarizing patient intake notes into structured clinical fields. They are concerned about the model producing confident but incorrect medical details. Which TWO practices best reduce the risk of fabricated content in this scenario? (Choose two.)

Select 2 answers

A.Ask the model to state its confidence as a percentage after each extracted field and trust fields scored above ninety percent.

B.Instruct the model to respond with a specific phrase such as 'not stated' when the source note does not contain a requested field.

C.Ground each summarization request by supplying the relevant intake note text in the prompt instead of relying on the model's memory.

D.Fine-tune the model on the entire historical patient record database to expand its clinical knowledge.

E.Raise the temperature to its maximum so the model produces diverse clinical phrasing across repeated runs.

AnswersB, C

Explicitly defining an abstention response gives the model a valid way to avoid filling gaps with plausible guesses. Since hallucination often arises when a field is missing and the model is expected to answer anyway, allowing an 'unknown' output removes the pressure to fabricate and makes downstream review of missing data straightforward.

Why this answer

Fabrication drops when the model is anchored to the actual source text and is given permission to report that information is absent. Supplying the intake note in the prompt keeps generation evidence-based, and defining an explicit abstention phrase prevents the model from inventing values for missing fields. Randomness and self-reported confidence do not reliably reduce hallucination.

Exam trap

The trap here is treating a model's self-reported confidence score as a trustworthy signal, when that number is generated text and is not a calibrated probability.

Full explanation →

564

MCQeasy

A developer wants to invoke a foundation model in Amazon Bedrock with a large number of similar requests for cost savings. Which feature can help reduce inference cost?

A.Model caching

B.Fine-tuning

C.Prompt management

D.Batch inference

AnswerA

Correct. Caching avoids recomputing responses for identical prompts, saving cost.

Why this answer

Model caching in Amazon Bedrock stores the results of frequently used or repeated inference requests, allowing the system to serve identical prompts from a cache rather than re-running the foundation model. This directly reduces inference cost by avoiding redundant compute for similar requests, making it the ideal choice for a developer with a large number of similar requests.

Exam trap

AWS often tests the misconception that batch inference or fine-tuning inherently reduces per-request cost, but the key distinction is that model caching eliminates redundant compute for repeated identical prompts, whereas batch inference still incurs compute for every request and fine-tuning adds upfront training cost.

How to eliminate wrong answers

Option B (Fine-tuning) is wrong because fine-tuning customizes a model on a specific dataset, which incurs additional training costs and does not inherently reduce per-request inference cost for repeated similar prompts. Option C (Prompt management) is wrong because it organizes and versions prompts but does not cache or reuse inference results, so it has no direct impact on reducing inference cost. Option D (Batch inference) is wrong because while it processes multiple requests together, it still runs the model for each request in the batch and does not eliminate redundant computation for identical or similar prompts.

Full explanation →

565

MCQhard

A company is using Amazon SageMaker to train a large language model with hundreds of billions of parameters. The model does not fit into the memory of a single GPU. Which approach should they use to train the model efficiently?

A.Use a larger instance with more GPU memory, such as p4d.24xlarge

B.Use SageMaker's data parallelism strategy

C.Use SageMaker's model parallelism strategy with the SageMaker distributed training library

D.Reduce the model size by pruning layers until it fits into memory

AnswerC

Model parallelism shards the model's layers and parameters across multiple GPUs, so no single device must hold the full hundreds-of-billions-parameter model. This directly resolves the stem's constraint that the model does not fit into one GPU's memory.

Why this answer

SageMaker's model parallelism strategy with the SageMaker distributed training library is specifically designed for training large models that do not fit into the memory of a single GPU. It partitions the model layers across multiple GPUs, enabling efficient training of models with hundreds of billions of parameters by overlapping computation and communication.

Exam trap

The AIF-C01 exam often tests the distinction between data parallelism and model parallelism, and the trap here is that candidates may confuse data parallelism (which splits data, not the model) as a solution for models that don't fit in memory, when in fact model parallelism is required for such cases.

How to eliminate wrong answers

Option A is wrong because even the largest GPU instances like p4d.24xlarge have limited GPU memory (40 GB per A100 GPU), which is insufficient for a model with hundreds of billions of parameters; scaling vertically is not feasible for such large models. Option B is wrong because SageMaker's data parallelism strategy replicates the entire model on each GPU and splits the data across GPUs, which requires the model to fit into a single GPU's memory; it does not solve the memory constraint issue. Option D is wrong because pruning layers to reduce model size would degrade model quality and is not a practical or efficient approach for training large language models; the goal is to train the full model, not a smaller version.

Full explanation →

566

MCQhard

A developer is explaining how a large language model generates text so that a business stakeholder understands why the same prompt can yield different answers. Which description accurately captures the generation process?

A.The model queries an external search index and paraphrases the top-ranked web page for the prompt.

B.The model searches a stored database of previously written sentences and returns the closest exact match to the prompt.

C.The model predicts a probability distribution over the next token, samples from it, and appends the chosen token to repeat the process.

D.The model compiles the prompt into a set of if-then rules that deterministically map keywords to canned responses.

AnswerC

Generation is autoregressive: given the prompt and tokens produced so far, the model outputs a probability distribution over the vocabulary, a token is selected according to the sampling strategy, and that token is appended before the next prediction. Because sampling is stochastic, the same prompt can produce different completions, which explains the variability the stakeholder observed.

Why this answer

Text generation is autoregressive and probabilistic. At each step the model computes a distribution over the next token conditioned on everything before it, a token is sampled according to settings such as temperature and top-p, and the sequence grows one token at a time. Stochastic sampling is why identical prompts can yield different outputs, and why temperature changes affect variability.

Exam trap

The trap here is assuming the model stores and retrieves whole sentences, when it actually predicts one token at a time from probability distributions over its vocabulary.

Full explanation →

567

Multi-Selectmedium

Which TWO actions help ensure fairness in an AI system deployed on AWS? (Select two.)

Select 2 answers

A.Train the model on a representative dataset

B.Enable AWS CloudTrail for audit

C.Use SageMaker Clarify to detect bias

D.Use a single validation set

E.Encrypt data at rest using AWS KMS

AnswersA, C

Training on a representative dataset directly reduces sampling bias, ensuring the model's learned patterns reflect the true population distribution rather than over-representing dominant groups. This satisfies the fairness constraint by preventing skewed predictions that disadvantage under-represented cohorts, forming the foundational data-level mitigation before any post-processing correction is applied.

Why this answer

Option A (Train the model on a representative dataset) is correct because fairness in AI depends on the training data reflecting the demographics and conditions of the real-world population the model will serve; unrepresentative or skewed data leads to biased predictions. Option C (Use SageMaker Clarify to detect bias) is correct because SageMaker Clarify provides bias detection metrics (e.g., pre-training and post-training bias metrics such as Disparate Impact and Equal Opportunity Difference) that quantify and help mitigate bias in data and models. Option B (Enable AWS CloudTrail for audit) is not a fairness action; CloudTrail records API activity for security, compliance, and operational auditing, not for measuring or correcting model bias.

Option D (Use a single validation set) is not a fairness measure and can actually increase evaluation variance; fairness requires representative and possibly multiple or cross-validated evaluation sets. Option E (Encrypt data at rest using AWS KMS) addresses data confidentiality and compliance, not fairness or bias in model outcomes.

Exam trap

AWS AI Practitioner candidates often confuse security/audit mechanisms (like CloudTrail and KMS) with fairness-specific tools (like SageMaker Clarify), leading them to select encryption or logging as bias mitigation actions.

Full explanation →

568

MCQmedium

A data scientist uses Amazon Bedrock. The model responses are too long. Which parameter should they adjust to limit the output length?

A.temperature

B.max_tokens

C.stop sequences

D.top_p

AnswerB

Setting max_tokens caps the number of tokens the model generates, directly satisfying the requirement to limit response length. In Amazon Bedrock, this parameter bounds output size independently of the prompt, so the data scientist can truncate verbose answers without altering the input or switching models.

Why this answer

The `max_tokens` parameter directly controls the maximum number of tokens (words or subwords) the model can generate in a single response. By reducing this value, the data scientist caps the output length, preventing overly long responses. Temperature and top_p affect randomness and diversity, not length, while stop sequences define when generation halts but do not enforce a hard token limit.

Exam trap

AWS often tests the distinction between parameters that control output length (`max_tokens`) versus those that control output randomness or diversity (`temperature`, `top_p`), leading candidates to confuse 'limiting length' with 'limiting creativity'.

How to eliminate wrong answers

Option A is wrong because temperature controls the randomness of token selection (higher values increase creativity, lower values make output more deterministic), not the length of the response. Option C is wrong because stop sequences are custom strings (e.g., '###' or 'END') that tell the model to cease generation when encountered, but they do not limit the total number of tokens generated before that point. Option D is wrong because top_p (nucleus sampling) limits the cumulative probability of token choices to a threshold (e.g., 0.9), affecting diversity, not the maximum output length.

Full explanation →

569

MCQmedium

A company uses a foundation model for real-time translation in a chat application. The latency is high. Which optimization would reduce latency the most?

A.Increase batch size

B.Use model distillation to create a smaller model

C.Use a larger model

D.Use a CDN for model weights

AnswerB

Distillation trains a compact student model to mimic the larger teacher, cutting inference compute and memory per token. Fewer parameters mean faster forward passes, directly addressing the real-time chat latency constraint rather than merely tuning prompts or batching.

Why this answer

Model distillation reduces the size of the foundation model by training a smaller 'student' model to mimic the behavior of a larger 'teacher' model. This directly decreases inference latency because the smaller model requires fewer computational resources (FLOPs) per forward pass, which is critical for real-time translation in a chat application where low latency is paramount.

Exam trap

The AIF-C01 exam often tests the distinction between throughput optimization (batch size) and latency optimization (model size/distillation), leading candidates to mistakenly choose increasing batch size when the question explicitly asks for reducing latency.

How to eliminate wrong answers

Option A is wrong because increasing batch size improves throughput (more requests processed per unit time) but does not reduce per-request latency; in fact, it can increase latency for individual requests as the model must wait for the batch to fill. Option C is wrong because using a larger model increases the number of parameters and computational complexity, which would increase latency, not reduce it. Option D is wrong because a CDN for model weights only accelerates the initial download of the model to edge locations, not the inference latency of each translation request; once the model is loaded, inference speed is determined by the model architecture and hardware, not network delivery.

Full explanation →

570

MCQmedium

A financial services company uses Amazon Bedrock to power a customer-facing chatbot that provides investment advice. The company must ensure that the chatbot's responses comply with regulatory standards, meaning that the model should not generate advice that is speculative or promises returns. The company has implemented Bedrock Guardrails with content filters. However, during testing, the chatbot still generates responses that violate the guidelines. A review of the guardrail configuration shows that the content filters are set to the lowest sensitivity. The company wants to enforce stricter filtering without completely blocking legitimate responses. What should the company do?

A.Increase the sensitivity of the content filters in the Bedrock Guardrails configuration.

B.Use a different foundational model that has built-in compliance filters.

C.Configure the chatbot to route all responses to a human reviewer before delivering to the customer.

D.Add a deny topic for investment advice to completely block that topic.

AnswerA

Raising content filter sensitivity tightens the confidence thresholds Bedrock Guardrails apply when classifying prompts and responses against harmful categories, so speculative or return-promising advice is intercepted more aggressively. Because filters remain category-based rather than blanket-blocking, legitimate investment responses still pass, satisfying the requirement for stricter filtering without wholesale denial.

Why this answer

Increasing the sensitivity of the content filters in Bedrock Guardrails directly addresses the issue: the current filters are set to the lowest sensitivity, allowing speculative or promise-based responses to pass through. By raising the sensitivity, the guardrails will block more non-compliant content while still permitting legitimate investment advice, striking the required balance between regulatory compliance and functionality.

Exam trap

The trap here is that candidates may think adding a deny topic (Option D) is the simplest way to enforce compliance, but they overlook that it completely blocks all investment advice, which violates the requirement to allow legitimate responses; the exam tests understanding of granular guardrail tuning versus blunt blocking.

How to eliminate wrong answers

Option B is wrong because switching to a different foundational model does not guarantee built-in compliance filters that meet the specific regulatory standards; models themselves do not enforce content policies—guardrails do. Option C is wrong because routing all responses to a human reviewer introduces latency and scalability issues, and does not solve the underlying guardrail configuration problem; it is a workaround, not a fix. Option D is wrong because adding a deny topic for investment advice would completely block all investment-related queries, which is overly restrictive and prevents the chatbot from providing any legitimate advice, violating the requirement to avoid completely blocking legitimate responses.

Full explanation →

571

MCQhard

Refer to the exhibit. A developer receives an error when trying to invoke the Claude Instant model from an application. The application uses the IAM role 'MyAppRole'. Which IAM policy statement should be added to the role to resolve the error?

A.{"Effect":"Allow","Action":"bedrock:GetFoundationModel","Resource":"*"}

B.{"Effect":"Allow","Action":"bedrock:InvokeModel","Resource":"arn:aws:bedrock:us-east-1::foundation-model/*"}

C.{"Effect":"Allow","Action":"bedrock:InvokeModel","Resource":"arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-instant-v1"}

D.{"Effect":"Allow","Action":"bedrock:*","Resource":"*"}

AnswerC

Bedrock authorises model invocation through the bedrock:InvokeModel action scoped to the specific foundation-model ARN. The role currently lacks this permission, so granting it on the Claude Instant ARN resolves the AccessDenied error without over-permissioning unrelated models or regions.

Why this answer

The error occurs when the application tries to invoke the Claude Instant model via the Bedrock InvokeModel API, and the IAM role 'MyAppRole' lacks the necessary permission. The required action is 'bedrock:InvokeModel', and the resource must be scoped to the specific foundation model ARN, which for Claude Instant in us-east-1 is 'arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-instant-v1'. This grants least-privilege access to invoke exactly that model.

Exam trap

The trap here is that candidates often confuse the 'InvokeModel' action with 'GetFoundationModel' or choose a wildcard resource ARN, thinking it is sufficient, but AWS specifically tests the need for a precise model ARN to enforce least-privilege access in Bedrock IAM policies.

How to eliminate wrong answers

Option A is wrong because 'bedrock:GetFoundationModel' is a read-only action used to retrieve metadata about a foundation model, not to invoke it for inference; the application needs the 'InvokeModel' action. Option B is wrong because while it allows 'bedrock:InvokeModel', the resource ARN uses a wildcard 'foundation-model/*', which grants access to all foundation models in the region, violating the principle of least privilege and potentially allowing unintended model invocations. Option D is wrong because 'bedrock:*' grants full administrative access to all Bedrock actions, which is overly permissive and not a best practice for a specific application role; it would also allow actions like model creation or deletion.

Full explanation →

572

MCQmedium

A developer is using Bedrock Studio to prototype a summarization application. They want to quickly test different foundation models and prompts without writing code. What should they use?

A.Bedrock Knowledge Bases

B.Bedrock Agents

C.Bedrock Guardrails

D.Bedrock Playground

AnswerD

Bedrock Playground provides a console interface for selecting foundation models and iterating on prompts interactively, with no code required. This directly satisfies the developer's need to rapidly compare models and prompt variants while prototyping the summarization application.

Why this answer

Amazon Bedrock Playground provides an interactive, no-code environment where developers can experiment with different foundation models and prompts. It allows quick testing and comparison of model outputs without writing any code, making it ideal for prototyping a summarization application.

Exam trap

AIF-C01 often tests the distinction between Bedrock features, and candidates may confuse the Playground with Knowledge Bases or Agents, especially if they focus on the word 'test' and overlook the no-code aspect.

How to eliminate wrong answers

Option A is wrong because Bedrock Knowledge Bases is a feature for connecting foundation models to proprietary data for retrieval-augmented generation (RAG), not for testing models and prompts interactively. Option B is wrong because Bedrock Agents is used to orchestrate multi-step tasks and call APIs, not for quick prompt testing. Option C is wrong because Bedrock Guardrails is for implementing safeguards and content filtering, not for model experimentation.

Full explanation →

573

MCQhard

A company is building a text classification system using embeddings. They need to choose between Amazon Titan Text Embeddings and Cohere Embed. The documents are in multiple languages, and the team requires strong cross-lingual performance without additional training. Which model is optimized for multilingual use cases?

A.Amazon Titan Text Embeddings

B.Neither model supports multilingual text

C.Both are equally multilingual

D.Cohere Embed

AnswerD

Cohere Embed is trained to map semantically equivalent text across languages into a shared vector space, so cross-lingual similarity comparisons work without fine-tuning. This directly satisfies the stem's requirement for strong multilingual performance without additional training, unlike Amazon Titan Text Embeddings, which targets primarily English-centric workloads.

Why this answer

Cohere Embed is explicitly optimized for multilingual use cases, supporting over 100 languages with strong cross-lingual performance out of the box. Amazon Titan Text Embeddings, while powerful for English-centric tasks, does not offer the same level of native multilingual optimization. Therefore, for a system requiring robust cross-lingual performance without additional training, Cohere Embed is the correct choice.

Exam trap

The trap here is that candidates may assume Amazon Titan Text Embeddings, being a broad AWS service, inherently supports multilingual use cases as well as Cohere Embed (available on AWS Bedrock), but the exam tests specific knowledge of which model is explicitly optimized for cross-lingual performance without additional training.

How to eliminate wrong answers

Option A is wrong because Amazon Titan Text Embeddings is primarily optimized for English and does not provide the same breadth of native multilingual support as Cohere Embed, making it unsuitable for strong cross-lingual performance without additional training. Option B is wrong because Cohere Embed explicitly supports multilingual text across many languages, so the claim that neither model supports multilingual text is factually incorrect. Option C is wrong because the two models are not equally multilingual; Cohere Embed is specifically designed and optimized for multilingual use cases, whereas Amazon Titan Text Embeddings is not.

Full explanation →

574

MCQmedium

An AI team uses the IAM policy shown in the exhibit to control endpoint creation. Why does this policy support responsible AI?

A.It requires human approval before deploying any model

B.It prevents the use of GPU instances to reduce cost

C.It ensures data capture is enabled for model monitoring

D.It restricts endpoints to only use models built in SageMaker

AnswerC

Enabling data capture feeds actual endpoint inputs and outputs into monitoring, satisfying the responsible AI requirement for ongoing oversight of deployed models. Without captured inference data, drift, bias and anomalous behaviour cannot be detected, so the policy enforces traceability rather than leaving monitoring optional.

Why this answer

The IAM policy includes a condition that enforces the `DataCaptureConfig.EnableCapture` parameter to be set to `true` when creating a SageMaker endpoint. This ensures that model monitoring data is automatically collected, which is a key practice for responsible AI as it allows continuous monitoring of model performance, bias detection, and drift analysis. Without data capture, teams cannot audit or validate model behavior in production, undermining accountability and transparency.

Exam trap

The AIF-C01 exam often tests the misconception that IAM policies for responsible AI focus on restricting model sources or instance types, when in fact the key mechanism is enforcing observability through data capture for ongoing monitoring.

How to eliminate wrong answers

Option A is wrong because the IAM policy does not include any condition requiring human approval (e.g., using `sts:AssumeRole` with MFA or a separate approval workflow); it only enforces data capture settings. Option B is wrong because the policy does not restrict instance types (e.g., GPU instances like `ml.p3.2xlarge`); it focuses solely on data capture configuration. Option D is wrong because the policy does not restrict endpoints to models built in SageMaker; it allows any model to be deployed as long as data capture is enabled, and there is no condition referencing model origin.

Full explanation →

575

MCQhard

A financial services company must comply with regulatory requirements that mandate explainability of credit scoring models. They have deployed a model using SageMaker and need to generate reports showing feature importance for each prediction. Which combination of services should they use to automate this?

A.SageMaker Model Monitor + Amazon QuickSight

B.SageMaker Ground Truth + AWS Lambda

C.SageMaker Clarify + SageMaker Pipelines

D.SageMaker Data Wrangler + SageMaker Studio

AnswerC

SageMaker Clarify computes feature attributions, such as SHAP values, for individual predictions, while SageMaker Pipelines orchestrates and automates the report generation workflow. Together they satisfy the regulatory explainability mandate by producing per-prediction feature-importance reports without manual intervention.

Why this answer

SageMaker Clarify provides built-in explainability capabilities, including feature importance for individual predictions via SHAP values, which directly addresses the regulatory requirement for model explainability. SageMaker Pipelines automates the end-to-end workflow, allowing you to schedule and run Clarify processing jobs to generate reports on a recurring basis without manual intervention.

Exam trap

AWS often tests the distinction between monitoring (Model Monitor) and explainability (Clarify), so the trap here is that candidates confuse 'monitoring model performance' with 'explaining individual predictions' and pick Option A.

How to eliminate wrong answers

Option A is wrong because SageMaker Model Monitor is designed for detecting data drift and model quality degradation, not for generating per-prediction feature importance reports; Amazon QuickSight is a visualization tool that cannot compute explainability metrics. Option B is wrong because SageMaker Ground Truth is used for creating labeled datasets via human annotation, not for model explainability, and AWS Lambda is a serverless compute service that lacks the specialized algorithms (e.g., SHAP) needed for feature importance. Option D is wrong because SageMaker Data Wrangler is for data preparation and feature engineering, not for post-hoc model explainability, and SageMaker Studio is an IDE that provides an interface but does not automate report generation.

Full explanation →

576

MCQeasy

Refer to the exhibit. A data scientist is training a model in SageMaker using a KMS-encrypted dataset. The training job fails with the error shown. Which action should be taken to resolve this issue?

A.Add the SageMaker execution role to the KMS key policy with the kms:Decrypt permission.

B.Create a new KMS key and update the bucket policy to use the new key.

C.Attach an IAM policy to the SageMaker execution role that allows kms:Decrypt on the key.

D.Disable server-side encryption on the S3 bucket and use client-side encryption.

AnswerA

The key policy must explicitly grant the execution role the kms:Decrypt permission.

Why this answer

The error indicates that the SageMaker training job cannot access the KMS-encrypted dataset. Since the S3 bucket uses server-side encryption with a customer-managed KMS key (SSE-KMS), the SageMaker execution role must be explicitly granted kms:Decrypt permission in the KMS key policy. Option A correctly adds this permission to the key policy, which is required for SageMaker to decrypt the data during training.

Exam trap

AWS often tests the distinction between IAM policies and KMS key policies; the trap here is that candidates assume attaching an IAM policy to the execution role (Option C) is sufficient, but KMS key policies must explicitly grant access to the role when the key is customer-managed.

How to eliminate wrong answers

Option B is wrong because creating a new KMS key and updating the bucket policy does not resolve the access issue; the training job still needs decrypt permissions on the key, and the bucket policy controls S3 access, not KMS decryption. Option C is wrong because attaching an IAM policy to the SageMaker execution role alone is insufficient if the KMS key policy does not grant the role access; KMS key policies must explicitly allow the principal (the IAM role) to use the key, and IAM policies alone cannot override a restrictive key policy. Option D is wrong because disabling server-side encryption on the S3 bucket is unnecessary and insecure; the correct approach is to grant the SageMaker execution role decrypt access to the existing KMS key, not to bypass encryption.

Full explanation →

577

MCQmedium

A developer is using Amazon Bedrock to generate text responses. They want to reduce the randomness of the output and make the model more deterministic. Which parameter should the developer decrease?

A.stop_sequences

B.Temperature

C.max_tokens

D.top_p

AnswerB

Temperature directly scales the sampling distribution's entropy before token selection, so lowering it sharpens probabilities toward the highest-likelihood tokens and reduces output randomness. This satisfies the stem's determinism constraint, unlike top-p or token limits, which control candidate pool size or length rather than sampling randomness itself.

Why this answer

Temperature directly controls the randomness of the model's output. Lowering the temperature (e.g., from 1.0 to 0.2) reduces the probability of sampling less likely tokens, making the model more deterministic and focused on the highest-probability next token. This is the standard parameter for adjusting output creativity versus determinism in large language models like those on Amazon Bedrock.

Exam trap

The AWS AI Practitioner exam often tests the distinction between temperature (direct logit scaling) and top_p (cumulative probability cutoff), leading candidates to incorrectly choose top_p when the question asks for reducing randomness, because both affect diversity but temperature is the more direct and commonly used parameter for determinism.

How to eliminate wrong answers

Option A is wrong because stop_sequences define tokens or phrases that halt generation, not randomness; they control when output ends, not how tokens are selected. Option C is wrong because max_tokens sets the maximum length of the generated response, not the randomness or determinism of token selection; reducing it truncates output but does not affect sampling behavior. Option D is wrong because top_p (nucleus sampling) controls the cumulative probability threshold for token selection, reducing it makes output more focused but still allows stochastic sampling; temperature is the primary parameter for directly scaling logits to adjust randomness.

Full explanation →

578

MCQeasy

A data scientist wants to quickly build a supervised learning model for binary classification on a tabular dataset with 10,000 rows and 200 features. The dataset has some missing values and requires minimal code. Which AWS service should the data scientist use?

A.Amazon SageMaker Studio Lab

B.Amazon SageMaker Clarify

C.Amazon SageMaker Autopilot

D.Amazon SageMaker JumpStart

AnswerC

SageMaker Autopilot automates algorithm selection, feature engineering and hyperparameter tuning for tabular classification, handling missing values and returning an explainable model with minimal code. It directly meets the binary classification requirement on the 10,000-row, 200-feature dataset.

Why this answer

Amazon SageMaker Autopilot is the correct choice because it automatically performs data preprocessing (including handling missing values), feature engineering, model selection, and hyperparameter tuning for supervised learning tasks like binary classification. It requires minimal code—users can simply point to a tabular dataset in Amazon S3 and specify the target column, and Autopilot will automatically train and evaluate multiple candidate models, making it ideal for quickly building a binary classifier on a 10,000-row, 200-feature dataset with missing values.

Exam trap

The AIF-C01 exam often tests the distinction between automated ML services (Autopilot) and model hosting or development environments (Studio Lab, JumpStart), so the trap here is that candidates may confuse SageMaker Autopilot with SageMaker JumpStart, thinking JumpStart also automates model building, when in fact JumpStart only provides pre-built models and requires manual configuration.

How to eliminate wrong answers

Option A is wrong because Amazon SageMaker Studio Lab is a free, no-code ML development environment that provides JupyterLab notebooks and limited compute resources, but it does not automate model building or handle missing values—it requires the user to write all code manually. Option B is wrong because Amazon SageMaker Clarify is designed for bias detection, model explainability, and fairness analysis, not for building or training supervised learning models; it cannot handle missing values or perform automated model selection. Option D is wrong because Amazon SageMaker JumpStart provides pre-built models and solutions for transfer learning and fine-tuning, but it does not automatically preprocess missing values or perform automated model selection for tabular binary classification—it requires the user to select and configure a model manually.

Full explanation →

579

MCQmedium

A company is using Amazon Bedrock to power a customer-facing chatbot. The security team wants to ensure that the chatbot does not respond to prompts that attempt to elicit inappropriate or off-topic responses, such as discussing competitors or providing medical advice. The company also wants to monitor and log all blocked prompts for analysis. Which AWS service or feature should they use to achieve this?

A.AWS WAF with custom rules to block inappropriate prompts

B.Amazon SageMaker Ground Truth to label and filter prompts

C.Amazon Comprehend to analyze and filter prompts

D.Amazon Bedrock Guardrails with denied topics and content filters

AnswerD

Amazon Bedrock Guardrails allows you to define denied topics to block discussions on specific subjects like competitors or medical advice. It also provides content filters to block harmful content. Guardrails can be configured to log blocked prompts to CloudWatch or S3, enabling monitoring and analysis. This directly meets the requirements.

Why this answer

Amazon Bedrock Guardrails provides configurable safeguards, including denied topics and content filters, that can block inappropriate or off-topic responses in real time. It also supports logging of blocked prompts, which allows the security team to monitor and analyze attempts. This is the most direct and effective solution for the chatbot's requirements.

Exam trap

The trap here is thinking that AWS WAF can inspect the semantic content of prompts; WAF operates at the network layer, not the application layer for natural language.

Full explanation →

580

MCQmedium

A company wants to automatically detect anomalies in server metrics. Which algorithm is most appropriate?

A.XGBoost

B.One-class SVM

C.Linear SVM

D.K-Means

AnswerB

One-class SVM learns a decision boundary describing normal behaviour from only normal samples, flagging deviations as anomalies. Server metrics rarely contain labelled anomalies, so this unsupervised approach suits the scenario better than supervised classifiers requiring labelled attack examples.

Why this answer

One-class SVM is specifically designed for anomaly detection, as it learns a boundary around the normal data points in the feature space and identifies any point falling outside this boundary as an anomaly. This makes it ideal for detecting unusual patterns in server metrics without requiring labeled anomaly examples.

Exam trap

The AIF-C01 exam often tests the distinction between supervised and unsupervised learning, and the trap here is that candidates may choose XGBoost or Linear SVM because they are familiar with them for classification, forgetting that anomaly detection typically requires a one-class approach when only normal data is available.

How to eliminate wrong answers

Option A is wrong because XGBoost is a supervised ensemble learning algorithm used for classification and regression, not for unsupervised anomaly detection; it requires labeled training data and is not designed to identify outliers without prior examples. Option C is wrong because Linear SVM is a supervised binary classifier that separates data into two classes using a hyperplane, and it cannot perform one-class anomaly detection without negative samples. Option D is wrong because K-Means is an unsupervised clustering algorithm that partitions data into clusters based on distance, but it does not inherently detect anomalies; while outliers can be inferred from cluster distances, it is not a dedicated anomaly detection method and lacks the statistical boundary learning of one-class SVM.

Full explanation →

581

MCQhard

A media company uses Amazon Bedrock to generate article summaries. They notice that for long articles, the model sometimes ignores instructions placed at the beginning of the prompt. The company wants to improve the model's adherence to instructions without changing the model or increasing cost significantly. Which prompt engineering technique should they apply?

A.Use a higher temperature value to make the model more creative in following instructions.

B.Split the article into smaller chunks and summarize each chunk separately, then concatenate the summaries.

C.Increase the maxTokenCount parameter to allow the model to process more of the article.

D.Move the instructions to the end of the prompt and repeat them after the article text.

AnswerD

Placing instructions at the end of the prompt, after the long context, leverages the model's tendency to pay more attention to recent tokens. Repeating key instructions after the article text reinforces them, improving adherence without changing the model or adding significant cost. This is a known prompt engineering technique for long-context scenarios.

Why this answer

The model's tendency to overlook early instructions in long contexts is a known limitation. By moving instructions to the end of the prompt and repeating them after the article, the company places them where the model's attention is strongest. This prompt engineering technique improves adherence without retraining or increasing cost, making it the most effective solution.

Exam trap

The trap here is thinking that increasing maxTokenCount or temperature will fix instruction-following, when the real issue is prompt structure and attention placement.

Full explanation →

582

Multi-Selecteasy

Which TWO actions are essential for ensuring accountability in AI systems according to AWS responsible AI guidelines?

Select 2 answers

A.Automate all decisions to ensure consistency

B.Establish clear human oversight and decision-making authority

C.Maintain detailed documentation and version control for models

D.Remove all human review processes to eliminate bias

E.Share raw training data publicly for transparency

AnswersB, C

Accountability requires identifiable humans who own outcomes, so assigning clear oversight and decision-making authority ensures someone is answerable for the system's behaviour. Without named responsibility, no governance control can enforce remediation when the model causes harm.

Why this answer

Option B is correct because AWS responsible AI guidance treats accountability as requiring identifiable human ownership: clear human oversight and defined decision-making authority ensure a person or role can be held responsible for a model's outcomes, including escalation and override paths. Option C is correct because accountability depends on traceability — detailed documentation (data provenance, design decisions, evaluation results) plus version control of models and artifacts lets auditors reproduce, explain, and attribute what a given system version did and why. Option A is wrong because full automation removes the human accountability chain rather than creating it, and consistency alone is not accountability.

Option D is wrong because eliminating human review destroys oversight and can worsen unchecked bias, the opposite of responsible AI. Option E is wrong because publishing raw training data raises privacy, consent, and IP risks and is not an accountability mechanism; transparency is served through appropriate documentation and disclosures, not indiscriminate data release.

Exam trap

The trap here is that candidates may confuse 'transparency' (Option E) with accountability, overlooking that raw data sharing introduces privacy and compliance risks, while proper accountability requires controlled documentation and human oversight, not unrestricted disclosure.

Full explanation →

583

MCQhard

Refer to the exhibit. A developer is optimizing latency for a generative AI model deployed on SageMaker. Based on the exhibit, which change would most likely reduce per-token latency?

A.Use a CPU instance

B.Reduce model size through quantization

C.Switch to a larger instance type

D.Increase batch size to 10

AnswerB

Quantization reduces weight precision, shrinking the model so each forward pass performs fewer and cheaper memory-bound operations per token. This directly lowers per-token latency, satisfying the exhibit's constraint of reducing inference time without retraining or changing the endpoint's instance type.

Why this answer

Reducing model size through quantization directly decreases the computational and memory requirements per inference step, which lowers the time to generate each token. This is especially effective on GPU instances where smaller models fit better in GPU memory and reduce memory bandwidth bottlenecks, leading to lower per-token latency.

Exam trap

Candidates often think that larger instances always reduce latency, when in fact they may increase latency due to higher memory latency and inter-chip communication, while quantization directly addresses the memory bandwidth bottleneck in autoregressive decoding.

How to eliminate wrong answers

Option A is wrong because CPU instances lack the parallel processing capabilities needed for efficient generative AI inference, resulting in significantly higher per-token latency compared to GPU instances. Option C is wrong because switching to a larger instance type may increase throughput but does not necessarily reduce per-token latency; it can even increase latency due to higher memory access times and inter-chip communication overhead. Option D is wrong because increasing batch size to 10 increases the total tokens processed per batch, which can improve throughput but typically increases per-token latency due to longer queueing and processing times for each batch.

Full explanation →

584

MCQeasy

Which vector store is a fully managed AWS service that can be used with Amazon Bedrock Knowledge Bases for semantic search?

A.Amazon DynamoDB

B.Amazon RDS for MySQL

C.Amazon S3

D.Amazon OpenSearch Serverless

AnswerD

Amazon OpenSearch Serverless provides a fully managed vector engine with k-NN search, and Bedrock Knowledge Bases integrates with it natively as a vector store. This satisfies the stem's constraint of a managed AWS service supporting semantic retrieval without provisioning servers.

Why this answer

Amazon OpenSearch Serverless is a fully managed AWS service that provides a vector store capability, which is required for semantic search in Amazon Bedrock Knowledge Bases. It supports vector indexing and similarity search, enabling efficient retrieval of relevant documents based on embedding vectors. Other options like DynamoDB, RDS for MySQL, and S3 are not purpose-built vector stores and lack the native vector search functionality needed for this use case.

Exam trap

The trap here is that candidates often confuse general-purpose databases or storage services (like DynamoDB, RDS, or S3) with purpose-built vector stores, assuming any database can perform semantic search if it stores data, but AWS specifically requires a vector store with native ANN indexing for Bedrock Knowledge Bases.

How to eliminate wrong answers

Option A is wrong because Amazon DynamoDB is a NoSQL key-value and document database that does not natively support vector indexing or similarity search; it would require external libraries or custom implementations to perform semantic search. Option B is wrong because Amazon RDS for MySQL is a relational database that lacks built-in vector search capabilities; while MySQL can store vectors as blobs, it cannot efficiently perform the distance-based queries required for semantic search without significant overhead. Option C is wrong because Amazon S3 is an object storage service, not a database or vector store; it cannot execute search queries or index vectors natively, making it unsuitable for semantic search in Bedrock Knowledge Bases.

Full explanation →

585

MCQhard

An enterprise wants to ensure that generative AI applications built on AWS comply with data privacy regulations. They need to prevent the model from using customer data in future training. Which feature of Amazon Bedrock should they enable?

A.Policy-based data governance

B.Opt-out of model improvement

C.Data encryption at rest

D.Model customization with customer data

AnswerB

Opting out of model improvement prevents Amazon from using the customer's inputs and outputs to train or refine foundation models. This satisfies the data privacy requirement by contractually and technically excluding customer data from future training.

Why this answer

Amazon Bedrock's opt-out of model improvement feature allows customers to prevent AWS from using their data (including prompts, completions, and associated metadata) for model training or service improvement. This is essential for compliance with data privacy regulations like GDPR or CCPA, as it ensures customer data is not retained or used beyond the immediate inference request.

Exam trap

Candidates often confuse data protection mechanisms (encryption, access control) with data usage controls (opt-out of model improvement) on Amazon Bedrock.

How to eliminate wrong answers

Option A is wrong because policy-based data governance (e.g., using AWS Lake Formation or IAM policies) controls access and permissions to data, but does not prevent the model provider from using customer data for future training. Option C is wrong because data encryption at rest (e.g., using AWS KMS or SSE) protects data confidentiality during storage but has no effect on whether the model uses that data for training. Option D is wrong because model customization with customer data (e.g., fine-tuning or continued pre-training) explicitly involves using customer data to improve the model, which is the opposite of preventing data usage for training.

Full explanation →

586

MCQmedium

A data scientist is using Amazon SageMaker to train a model and wants to understand the contribution of each feature to individual predictions. Which technique should they use to generate local explanations?

A.Permutation feature importance

B.Global feature importance

C.SHAP values

D.Partial dependence plots

AnswerC

SHAP values derive from cooperative game theory, assigning each feature a contribution to a single prediction while accounting for feature interactions. This yields consistent, locally accurate explanations per instance, unlike global impurity-based importance, which summarises the whole model rather than individual predictions.

Why this answer

SHAP (SHapley Additive exPlanations) values are the correct choice because they provide local explanations by decomposing a prediction into the additive contribution of each feature, based on cooperative game theory. This allows the data scientist to understand exactly how each feature influenced a specific individual prediction, unlike global methods that summarize behavior across the entire dataset.

Exam trap

The trap here is that candidates often confuse global feature importance (e.g., permutation importance) with local explanation methods, mistakenly thinking that a global ranking can explain individual predictions, when in fact only techniques like SHAP or LIME provide per-instance feature contributions.

How to eliminate wrong answers

Option A is wrong because permutation feature importance measures the decrease in model performance when a feature's values are randomly shuffled, which yields a global importance score across all predictions, not a per-instance local explanation. Option B is wrong because global feature importance aggregates feature contributions over the entire dataset (e.g., average absolute SHAP values), providing a single ranking per feature rather than explaining individual predictions. Option D is wrong because partial dependence plots show the average marginal effect of a feature on the predicted outcome across the dataset, which is a global interpretation technique and does not decompose a single prediction into feature-level contributions.

Full explanation →

587

MCQmedium

A company is using Amazon Rekognition to detect objects in images. They find that the service sometimes mislabels objects. What is the best way to improve accuracy for their specific use case?

A.Use a larger image size

B.Contact AWS support

C.Increase the confidence threshold

D.Use Amazon SageMaker to build a custom model

AnswerD

Rekognition's pre-trained labels cannot be tuned to niche classes, so mislabelling persists. SageMaker lets you train a custom model on your own annotated images, matching the exact object categories and visual conditions of your use case, which directly addresses the accuracy constraint.

Why this answer

Amazon Rekognition is a pre-trained service that may not perform optimally for specialized or domain-specific use cases. By using Amazon SageMaker to build a custom model, you can train a model on your own labeled dataset, which directly addresses the mislabeling issue by tailoring the model to your specific images and objects.

Exam trap

The trap here is that candidates often assume increasing the confidence threshold is a universal fix for accuracy issues, but the AIF-C01 exam tests the understanding that pre-trained services have limitations and that custom training (via SageMaker) is required for domain-specific improvements.

How to eliminate wrong answers

Option A is wrong because using a larger image size does not inherently improve Rekognition's detection accuracy; the service already resizes images to a standard input size, and larger images may only increase processing time without correcting mislabeling. Option B is wrong because contacting AWS support will not modify the underlying pre-trained model or improve its accuracy for your specific use case; support can only assist with service configuration or bugs, not model retraining. Option C is wrong because increasing the confidence threshold reduces false positives but does not fix systematic mislabeling; it may cause the service to return fewer results, potentially missing correct detections, without addressing the root cause of incorrect object identification.

Full explanation →

588

Multi-Selecthard

Which THREE considerations are essential when deploying a generative AI application in a regulated industry such as healthcare?

Select 3 answers

A.Lowest possible inference latency for real-time responses.

B.Full audit trail of model inputs and outputs for accountability.

C.Robust content filtering to block harmful or inaccurate outputs.

D.Maximum creative freedom for the model to generate diverse responses.

E.Data privacy and compliance with regulations like HIPAA.

AnswersB, C, E

Healthcare regulators require traceability of every AI-driven decision. Logging each model input and output creates the audit trail needed for accountability, incident investigation and post-market surveillance, satisfying the stem's regulated-industry constraint that opaque inference cannot meet.

Why this answer

In a regulated industry such as healthcare, option B (full audit trail of model inputs and outputs) is essential because accountability and traceability are required to investigate incidents, demonstrate compliance during audits, and reconstruct how a clinical or patient-facing decision was produced. Option C (robust content filtering) is essential because generative models can hallucinate or emit harmful advice, and filtering/guardrails are needed to prevent unsafe or inaccurate outputs from reaching patients or clinicians. Option E (data privacy and HIPAA compliance) is essential because healthcare data is protected health information, so the deployment must enforce safeguards such as access controls, encryption, and Business Associate Agreements to avoid regulatory violations.

Option A is not essential in this context because lowest possible inference latency is a performance optimization, not a regulatory or safety requirement, and accuracy, privacy, and auditability take precedence. Option D is not appropriate because maximum creative freedom increases variability and hallucination risk, which directly conflicts with the determinism, safety, and compliance needs of regulated healthcare use.

Exam trap

The trap here is that candidates may prioritize performance metrics like latency (Option A) over compliance requirements, mistakenly assuming that speed is always critical in healthcare, whereas AWS services in regulated industries must prioritize data privacy and auditability as non-negotiable, as mandated by regulations like HIPAA.

Full explanation →

589

MCQhard

A company is building a resume screening model and discovers that the training data contains only resumes from one gender, leading to biased predictions. Which type of bias does this represent, and what is the most effective mitigation strategy?

A.Aggregation bias; mitigate by using a single model for all groups

B.Representation bias; mitigate by collecting more diverse training data

C.Measurement bias; mitigate by using more precise measurement tools

D.Historical bias; mitigate by removing sensitive attributes from the model

AnswerB

Training data drawn from a single gender under-represents the population the model scores, so the model learns skewed patterns. Collecting diverse resumes that reflect the real applicant distribution corrects the sampling gap at source, satisfying the need to remove the bias rather than mask it post hoc.

Why this answer

Representation bias occurs when certain groups are underrepresented in the training data. Mitigation includes collecting more diverse data or using techniques like re-weighting or synthetic data generation.

Full explanation →

590

MCQmedium

A data science team is using Amazon SageMaker to train multiple models with different hyperparameters. They want to track metrics, compare runs, and reproduce the best result. Which SageMaker feature should they use?

A.SageMaker Model Registry

B.SageMaker Debugger

C.SageMaker Autopilot

D.SageMaker Experiments

AnswerD

SageMaker Experiments groups training runs into experiments and trials, automatically logging hyperparameters, metrics and artefacts for each job. This directly satisfies the team's need to track metrics across runs, compare them side by side, and reproduce the best result by retrieving its exact configuration.

Why this answer

SageMaker Experiments is the correct feature because it is specifically designed to track, organize, and compare machine learning training runs (trials) with different hyperparameters and metrics. It allows data scientists to log parameters, metrics, and artifacts for each run, compare results across runs, and retrieve the exact configuration needed to reproduce the best-performing model.

Exam trap

The trap here is that candidates often confuse SageMaker Experiments with SageMaker Model Registry, mistakenly thinking that model versioning and run tracking are the same feature, when in fact Experiments focuses on the iterative training process and Registry focuses on the final model lifecycle.

How to eliminate wrong answers

Option A is wrong because SageMaker Model Registry is a catalog for managing and versioning trained models, not for tracking and comparing individual training runs or hyperparameter experiments. Option B is wrong because SageMaker Debugger monitors training jobs in real time for issues like vanishing gradients or overfitting, but it does not provide a structured way to log, compare, or reproduce runs with different hyperparameters. Option C is wrong because SageMaker Autopilot automatically explores different algorithms and hyperparameters to find the best model, but it does not give the team the ability to manually track, compare, and reproduce their own custom runs with specific hyperparameters.

Full explanation →

591

MCQhard

A team is developing a real-time code completion feature using an LLM deployed on Amazon SageMaker. They observe high latency under load. Which optimization technique should they prioritize?

A.Increase batch size

B.Switch to a larger instance type

C.Increase instance count with Auto Scaling

D.Use model quantization

AnswerD

Quantization reduces weight precision, shrinking model size and memory bandwidth demand, which lowers per-token inference latency and increases throughput under concurrent load. This directly addresses the high-latency constraint for real-time code completion without retraining or architectural changes.

Why this answer

Model quantization reduces the precision of the model's weights (e.g., from FP32 to INT8), which decreases memory footprint and computational requirements, leading to lower latency per inference. This is the most effective optimization for real-time code completion because it directly reduces the time to generate each token without requiring additional infrastructure changes.

Exam trap

The AWS exam often tests the distinction between throughput optimization (batch size, scaling) and latency optimization (quantization, pruning), and the trap here is assuming that scaling out or increasing resources always solves latency issues, when in fact the bottleneck is per-inference computation time.

How to eliminate wrong answers

Option A is wrong because increasing batch size increases throughput but also increases latency for individual requests, which is counterproductive for real-time code completion that requires low per-request latency. Option B is wrong because switching to a larger instance type may reduce latency but at a higher cost and without addressing the fundamental computational bottleneck; it is a brute-force approach rather than an optimization. Option C is wrong because increasing instance count with Auto Scaling improves scalability and handles more concurrent requests, but it does not reduce the latency of individual inference calls, which is the core issue under load.

Full explanation →

592

MCQmedium

A team has created a knowledge base in Amazon Bedrock for a Q&A application. After updating the source documents, they notice that the model still returns old information. What is the MOST likely cause?

A.The chunking strategy is incorrect

B.The foundation model has a limited context window

C.The guardrails are filtering out the new information

D.The knowledge base has not been resynchronized after updating the documents

AnswerD

Amazon Bedrock knowledge bases only reflect source changes after an ingestion job re-synchronises the data source; editing documents alone leaves the existing vector embeddings untouched. Because the stem describes stale answers following a document update, resynchronisation is the missing step that regenerates embeddings and restores accurate retrieval.

Why this answer

When source documents in an Amazon Bedrock knowledge base are updated, the knowledge base must be resynchronized to ingest the changes. Without resynchronization, the vector embeddings and index still reflect the old documents, so the model continues to retrieve outdated information.

Exam trap

AIF-C01 often tests the operational aspects of Bedrock Knowledge Bases, and candidates may overlook the need for manual or automated resynchronization, instead blaming model limitations or guardrails.

How to eliminate wrong answers

Option A is wrong because an incorrect chunking strategy would affect the quality of retrieval but would not cause the model to return old information after documents are updated; it would affect all queries consistently. Option B is wrong because a limited context window would truncate input but would not cause stale data to be returned; the model would still use the most recent embeddings if they were updated. Option C is wrong because guardrails filter content based on policies, not based on document freshness; they would not selectively filter out new information while allowing old information.

Full explanation →

593

Multi-Selecthard

A company is using Amazon SageMaker to manage the lifecycle of their machine learning models. They need to implement a governance framework that includes model versioning, monitoring for drift, and decommissioning of outdated models. Which THREE AWS services or features should they use together to meet these requirements? (Select THREE.)

Select 3 answers

A.AWS CloudTrail

B.SageMaker Pipelines

C.SageMaker Model Registry

D.SageMaker Role Manager

E.SageMaker Model Monitor

AnswersB, C, E

SageMaker Pipelines orchestrates the end-to-end ML workflow as code, automating retraining, evaluation and conditional deployment steps. Within a governance framework it enforces repeatable, auditable lifecycle transitions, complementing Model Registry versioning and Model Monitor drift detection rather than duplicating them.

Why this answer

SageMaker Model Registry (C) is the correct service for model versioning and governance, as it maintains a catalog of model versions with metadata, approval statuses, and lineage, enabling tracking and controlled promotion of models through their lifecycle. SageMaker Model Monitor (E) is correct because it continuously monitors deployed endpoints for data drift, model quality drift, bias drift, and feature attribution drift, alerting when the model deviates from baseline behavior. SageMaker Pipelines (B) is correct because it provides CI/CD orchestration to automate the ML workflow, including training, evaluation, registration to the Model Registry, and deployment, which supports governance and decommissioning through repeatable, versioned pipeline executions.

AWS CloudTrail (A) only records API activity for auditing and does not provide model versioning, drift monitoring, or lifecycle decommissioning, so it does not meet the requirements. SageMaker Role Manager (D) merely helps create and manage IAM roles and permissions for SageMaker personas; it does not provide versioning, drift detection, or model decommissioning capabilities.

Full explanation →

594

MCQeasy

A company is deploying a machine learning model on Amazon SageMaker. The compliance team requires that the model's predictions be explainable and that the company can provide documentation on how the model makes decisions. Which SageMaker feature should the company use to meet this requirement?

A.Amazon SageMaker Autopilot

B.Amazon SageMaker Debugger

C.Amazon SageMaker Clarify

D.Amazon SageMaker Model Monitor

AnswerC

SageMaker Clarify provides tools to detect bias and explain model predictions. It generates feature attribution explanations using SHAP values, which show how each input feature contributes to the model's output. This directly meets the requirement for explainability and documentation of model decisions, helping the company comply with regulations that demand transparency.

Why this answer

SageMaker Clarify is specifically designed to provide explainability for machine learning models. It generates feature attribution reports that show the contribution of each input feature to the model's predictions, which can be used to document and explain model decisions. This meets the compliance requirement for explainable AI.

Exam trap

The trap here is confusing model monitoring with model explainability; Model Monitor tracks performance, while Clarify explains predictions.

Full explanation →

595

MCQeasy

A company is using Amazon SageMaker to deploy a machine learning model for credit scoring. The compliance team requires that the model's predictions be explainable to customers, and that the company can demonstrate which features contributed most to a decision. Which SageMaker feature should be used to meet this requirement?

A.SageMaker Debugger

B.SageMaker Experiments

C.SageMaker Model Monitor

D.SageMaker Clarify

AnswerD

SageMaker Clarify provides feature attribution explanations using SHAP values, which show the contribution of each feature to a model's prediction. This allows the company to explain individual credit decisions to customers and auditors. Clarify can be integrated with SageMaker endpoints to generate explanations in real time, meeting the compliance requirement.

Why this answer

The requirement is to explain individual predictions by identifying which features contributed most to a credit decision. SageMaker Clarify offers SHAP-based feature attributions that can be generated for real-time predictions. This enables the company to provide explanations to customers and demonstrate compliance.

Other SageMaker features focus on monitoring, debugging, or experiment tracking, not on model explainability.

Exam trap

The trap here is confusing model monitoring or debugging with explainability, when only Clarify provides feature-level attribution for individual predictions.

Full explanation →

596

MCQmedium

A retailer wants to group its customers into distinct behavioral segments for targeted marketing, but it has no predefined segment labels and no historical outcomes to learn from. Which machine learning approach should the retailer use?

A.Clustering

B.Reinforcement learning

C.Regression

D.Supervised classification

AnswerA

Clustering is an unsupervised technique that groups similar records without predefined labels. The retailer only has customer attributes and behavior, and wants to discover natural segments, so an algorithm such as k-means can partition customers by similarity. This directly matches the goal of finding structure in unlabeled data.

Why this answer

Because no segment labels exist and the objective is to discover natural groupings among customers, this is an unsupervised learning problem best solved with clustering. Classification, regression, and reinforcement learning all depend on labels, targets, or reward signals that the retailer does not have, so clustering is the only approach that fits the described situation.

Exam trap

The trap here is reaching for classification because the business wants 'segments', when the absence of predefined labels is precisely what makes this an unsupervised clustering problem.

Full explanation →

597

MCQeasy

A company uses a generative AI model to create marketing copy. They want to ensure that customers know the content is AI-generated. Which practice directly addresses this transparency requirement?

A.Implement AWS CloudTrail to log all model inference calls

B.Add a watermark or disclaimer stating 'This content was generated by AI'

C.Store all generated content in Amazon S3 with versioning enabled

D.Use a more powerful model to generate more natural-sounding text

AnswerB

A visible watermark or disclaimer explicitly labels the marketing copy as AI-generated, so customers immediately recognise its origin. This directly satisfies the transparency requirement, whereas internal documentation or model selection does not inform the audience viewing the published content.

Why this answer

Transparency in AI-generated content involves clearly disclosing to users that the content was produced by AI. This can be done through disclaimers, labeling, or other communication methods.

Full explanation →

598

Multi-Selectmedium

A company is deploying a generative AI application using Amazon Bedrock and needs to optimize costs for a high-volume, latency-tolerant workload. Which TWO strategies should they implement? (Select TWO.)

Select 2 answers

A.Use Batch Inference for asynchronous processing

B.Deploy a large model and fine-tune it

C.Use a smaller, more efficient foundation model

D.Enable Provisioned Throughput for guaranteed capacity

E.Implement model caching to avoid redundant inferences

AnswersA, C

Batch Inference submits large volumes of asynchronous requests at a lower per-token price than on-demand invocation, and results are returned within a target completion window. This satisfies the stem's latency-tolerant constraint while reducing cost for high-volume workloads.

Why this answer

Option A is correct because Amazon Bedrock Batch Inference lets you submit large volumes of prompts as an asynchronous job, which is priced at a discount (typically 50%) compared to on-demand inference and is ideal for latency-tolerant workloads that don't need immediate responses. Option C is correct because selecting a smaller, more efficient foundation model reduces the number of parameters processed per token, directly lowering per-token input/output costs while still meeting the application's quality needs. Option B is not appropriate because fine-tuning a large model adds training costs and does not reduce per-inference pricing for a high-volume workload.

Option D is wrong because Provisioned Throughput reserves dedicated capacity at a fixed hourly commitment, which is more expensive for spiky or latency-tolerant batch workloads than on-demand or batch pricing. Option E is incorrect because Bedrock does not offer a built-in response/model caching feature that avoids redundant inference charges, so it is not a valid cost-optimization strategy here.

Full explanation →

599

MCQhard

A company wants to use Amazon Bedrock to generate personalized marketing emails. They have thousands of customer profiles with demographic data. To generate tailored content efficiently, the application must dynamically insert customer-specific information into prompts. Which prompt management technique is BEST suited for this?

A.Using a larger foundation model to understand the entire customer base

B.Prompt flows with variables

C.Fine-tuning the model on all customer profiles

D.Prompt versioning

AnswerB

Prompt flows with variables let the application substitute each customer's demographic values into a reusable template at runtime, satisfying the requirement to insert customer-specific information dynamically across thousands of profiles. Unlike static prompts, variables parameterise the prompt, so one flow generates tailored content per customer without manual rewriting.

Why this answer

Prompt flows with variables allow the application to dynamically insert customer-specific data (e.g., name, age, purchase history) into a base prompt template at runtime. This technique avoids retraining or switching models for each customer, enabling efficient, personalized content generation without modifying the underlying foundation model.

Exam trap

The trap here is that fine-tuning adapts the model to a general pattern (e.g., tone or style) and cannot handle per-customer dynamic data insertion—prompt variables are the correct, lightweight solution.

How to eliminate wrong answers

Option A is wrong because using a larger foundation model does not inherently enable dynamic insertion of customer-specific data; it only increases computational cost and latency without solving the need for variable substitution. Option C is wrong because fine-tuning the model on all customer profiles is impractical, expensive, and unnecessary—fine-tuning adapts the model to a domain or style, not to individual customer records, and would require retraining for every new profile. Option D is wrong because prompt versioning tracks changes to prompt templates over time but does not provide a mechanism to inject runtime variables into prompts.

Full explanation →

600

MCQmedium

A company is evaluating the performance of a summarization model using Amazon Bedrock model evaluation. They want an automated metric that measures how well the generated summary captures the meaning of the reference summary. Which metric is MOST suitable?

A.BLEU

B.ROUGE-L

C.Accuracy

D.BERTScore

AnswerD

BERTScore computes token-level semantic similarity using contextual embeddings, so it rewards summaries that preserve meaning even when wording differs from the reference. That directly satisfies the requirement to measure how well generated summaries capture the reference summary's meaning, unlike surface-overlap metrics.

Why this answer

BERTScore is the most suitable metric because it uses contextual embeddings from BERT to compute semantic similarity between generated and reference summaries, capturing meaning rather than exact n-gram overlap. This makes it ideal for summarization evaluation where paraphrasing and semantic equivalence are critical.

Exam trap

The trap is that candidates often default to ROUGE-L because it is commonly used in summarization tasks, but for Amazon Bedrock's automated evaluation, BERTScore is preferred as it captures semantic meaning using contextual embeddings from BERT.

How to eliminate wrong answers

Option A is wrong because BLEU measures precision of n-gram overlap, which penalizes valid paraphrasing and does not capture semantic meaning, making it unsuitable for summarization. Option B is wrong because ROUGE-L measures longest common subsequence (LCS) overlap, focusing on surface-level lexical similarity rather than deep semantic meaning. Option C is wrong because Accuracy is a classification metric that requires exact match of discrete labels, not applicable to free-text generation evaluation.

Full explanation →

AWS Certified AI Practitioner AIF-C01 (AIF-C01) — Questions 526–600