AWS Certified AI Practitioner — Practice Set 1

Question 1

A company wants to build a model that predicts whether a customer will churn next month based on past usage data labeled as 'churned' or 'retained'. Which type of machine learning approach BEST describes this task?

Accepted Answer

Supervised learning trains a model on labeled data — in this case, historical customer records labeled as 'churned' or 'retained' — to predict outcomes for new data. Unsupervised learning finds patterns in unlabeled data. Reinforcement learning trains agents through reward signals, not labeled datasets. Self-supervised learning generates labels from the data itself and is commonly used in large language model pre-training.

Question 2

An organization is evaluating AI services for its data science team. The team wants a fully managed platform to build, train, and deploy machine learning models at scale without managing infrastructure. Which AWS service BEST meets this requirement?

Accepted Answer

Amazon SageMaker AI is AWS's fully managed end-to-end ML platform for building, training, and deploying ML models at scale, covering the entire ML lifecycle with managed infrastructure. Amazon Rekognition is a pre-built computer vision service for image and video analysis. Amazon Comprehend is a pre-built natural language processing (NLP) service. AWS Glue is an ETL data integration service, not an ML training platform.

Question 3

A company trained a machine learning model that performs with 99% accuracy on training data but only 62% accuracy on new production data. Which ML problem BEST describes this situation?

Accepted Answer

Overfitting occurs when a model learns the training data too precisely, including its noise and outliers, and therefore fails to generalize to new unseen data — exactly what the high train/low production accuracy gap indicates. Underfitting is the opposite: the model performs poorly on both training and production data. Data drift refers to changes in input data distribution over time after deployment. Class imbalance is a training data issue where one class is disproportionately represented, not a generalization failure.

Question 4

A company wants to detect fraudulent transactions in real time from its financial data. The data science team does not have labeled examples of fraud. Which ML approach is MOST appropriate for this requirement?

Accepted Answer

Unsupervised anomaly detection identifies data points that deviate significantly from normal patterns without requiring labeled examples of fraud. Supervised binary classification requires labeled fraud/non-fraud examples, which are not available here. Natural language processing deals with text, not financial transaction data. Reinforcement learning trains agents through reward/penalty feedback, not suitable for fraud detection without a reward signal.

Question 5

A company's ML team is evaluating a classification model for medical diagnosis. The team is concerned about the cost of false negatives (missing a real disease) versus false positives (incorrectly flagging healthy patients). Which metric BEST helps the team understand the model's ability to avoid missing actual positive cases?

Accepted Answer

Recall (also called Sensitivity or True Positive Rate) measures the proportion of actual positive cases correctly identified by the model — directly quantifying how well the model avoids false negatives (missed diagnoses). Accuracy measures overall correct predictions and can be misleading with imbalanced classes. Precision measures the proportion of positive predictions that are actually correct, focusing on false positives rather than false negatives. F1 Score is the harmonic mean of precision and recall, useful when both are important, but does not isolate the false negative rate.

Question 6

A company wants to integrate a chatbot into its customer service portal. The chatbot should generate natural, context-aware responses to customer questions. Which AWS service provides access to foundation models for building this type of generative AI application?

Accepted Answer

Amazon Bedrock is a fully managed service that provides access to high-performing foundation models (FMs) from AWS and leading AI companies through a single API, enabling developers to build and scale generative AI applications like chatbots. Amazon Comprehend is an NLP service for analysis tasks like sentiment detection, not text generation. Amazon Transcribe converts speech to text. Amazon Polly converts text to speech.

Question 7

An organization's developer is designing prompts for a large language model (LLM) and wants to understand what determines the LLM's output quality. Which factor MOST directly influences the relevance and accuracy of an LLM's responses?

Accepted Answer

The quality, clarity, and specificity of the input prompt is the most direct factor influencing an LLM's output relevance and accuracy — this is the fundamental principle behind prompt engineering. The number of tokens in the response is an output characteristic, not an input that drives quality. The hosting region affects latency and data residency but not the model's comprehension of the task. File format is an API concern unrelated to output quality.

Question 8

A company wants to use a large language model to generate creative product descriptions and is willing to accept more varied and sometimes unexpected outputs. Which model inference parameter should the team INCREASE to achieve this behavior?

Accepted Answer

Temperature controls the randomness of the model's token selection. A higher temperature value increases randomness and creativity in the output, producing more varied and sometimes unexpected responses. Top-P (nucleus sampling) controls the cumulative probability of token selection and also affects diversity, but temperature is the primary dial for creativity. Maximum token length controls response length, not creativity. Stop sequences define where the model should stop generating, unrelated to output variety.

Question 9

A company wants to customize a foundation model in Amazon Bedrock to recognize its specific product catalog terminology without retraining the model from scratch. The company has a small set of labeled product description examples. Which customization technique is MOST appropriate?

Accepted Answer

Fine-tuning adapts a pre-trained foundation model to domain-specific tasks by continuing training on a small labeled dataset, teaching the model to recognize specialized terminology without the cost of full pre-training. Pre-training from scratch is extremely expensive and unnecessary when a foundation model already exists. Increasing the context window is a model capability parameter, not a customization technique. Changing temperature only affects inference randomness, not what the model knows.

Question 10

An organization wants to generate AI-powered images from text descriptions for its marketing campaigns. Which AWS service provides text-to-image generation capability using foundation models?

Accepted Answer

Amazon Bedrock provides access to Stability AI image generation models (such as Stable Diffusion) that generate high-quality images from text prompts. Amazon Rekognition is a computer vision service for analyzing existing images, not generating new ones. Amazon SageMaker AI Canvas is a no-code ML tool for tabular and structured data predictions. AWS DeepComposer is an AI-powered music composition tool, not an image generator.

Question 11

A company's generative AI application occasionally produces responses that contain factually incorrect information presented with high confidence. The engineering team refers to this behavior by a specific AI term. What term BEST describes this phenomenon?

Accepted Answer

Hallucination refers to a generative AI model producing plausible-sounding but factually incorrect or fabricated information. This occurs because LLMs generate statistically likely token sequences rather than retrieving verified facts. Model overfitting is a training data generalization failure. Prompt injection is an attack where malicious instructions are embedded in user input to manipulate model behavior. Data poisoning involves corrupting training data to cause harmful model behavior.

Question 12

An organization wants to compare the performance of several foundation models available through Amazon Bedrock before choosing one for its customer support use case. The team wants to evaluate models on their own proprietary test prompts without running a full deployment. Which Amazon Bedrock feature MOST directly supports this workflow?

Accepted Answer

Amazon Bedrock Model Evaluation allows customers to evaluate, compare, and select foundation models based on custom prompts and automatic or human-review metrics before committing to a specific model for production. Amazon Bedrock Knowledge Bases connects FMs to proprietary data sources for retrieval-augmented generation. Amazon Bedrock Agents builds autonomous AI agents that execute multi-step tasks. Amazon Bedrock Guardrails enforces content filters and safety policies on deployed models.

Question 13

A company wants to build a question-answering system that answers questions using its internal documents stored in Amazon S3. The system should not hallucinate answers that are not in the documents. Which architecture pattern on Amazon Bedrock BEST addresses this requirement?

Accepted Answer

Retrieval-Augmented Generation (RAG) with Amazon Bedrock Knowledge Bases grounds LLM responses in authoritative retrieved documents, significantly reducing hallucinations by only letting the model answer based on what it retrieves from the knowledge base. Fine-tuning teaches the model patterns from documents during training but does not ensure real-time grounding to specific retrieved passages. Higher temperature increases creativity and randomness, which would worsen hallucination. A custom SageMaker AI model does not inherently address document-grounded Q&A.

Question 14

A company wants to automate its customer support workflow so that an AI assistant can look up order status, issue refunds, and send confirmation emails — all in response to a single customer message. Which Amazon Bedrock feature enables this multi-step automation capability?

Accepted Answer

Amazon Bedrock Agents enables the creation of autonomous AI agents that can break down complex tasks, call external APIs and AWS services, and complete multi-step workflows — such as looking up orders, issuing refunds, and sending emails — based on a single natural language request. Guardrails enforces content policies but does not execute multi-step tasks. Model Evaluation compares models. Knowledge Bases retrieves documents but does not call APIs or execute actions.

Question 15

An organization wants to use a foundation model via Amazon Bedrock to summarize lengthy legal contracts. The team notices the model sometimes truncates the summary before capturing key clauses. Which model parameter should the team adjust to allow the model to produce longer responses?

Accepted Answer

Maximum tokens (or max_gen_len) sets the upper limit on the number of tokens the model can generate in a single response. Increasing this value allows the model to produce longer, more complete summaries. Temperature and Top-P control randomness in token selection, not response length. Top-K limits the vocabulary considered at each step for diversity control, also unrelated to output length.

Question 16

A company is building an AI coding assistant using Amazon Bedrock. The team wants the model to follow a specific response format — always returning a JSON object with code and an explanation field. Which prompting technique BEST ensures consistent structured output from the model?

Accepted Answer

Few-shot prompting provides the model with concrete examples of the desired input/output format in the prompt — in this case, showing example JSON responses — which strongly guides the model to follow the same structure consistently. Zero-shot prompting gives no examples and is less reliable for strict formatting. Increasing temperature to 1.0 increases randomness, making output less consistent. Using a smaller model does not reliably enforce output structure.

Question 17

A company wants to transcribe customer service phone calls and then automatically extract the sentiment and key action items from the transcription. Which combination of AWS AI services provides the MOST direct solution with LEAST operational overhead?

Accepted Answer

Amazon Transcribe converts speech to text from the phone call recordings, and Amazon Comprehend then analyzes the transcribed text to extract sentiment and key phrases. Both are fully managed services requiring no ML expertise, making this the lowest-overhead solution. Amazon Polly is text-to-speech (the reverse of what is needed). Amazon Rekognition analyzes images and video, not audio. Amazon Lex builds conversational interfaces. Amazon Kendra is an enterprise search service.

Question 18

A company is deploying a generative AI application using Amazon Bedrock and wants to prevent the model from generating responses on topics outside the company's approved use cases, block harmful content, and redact sensitive personal information from outputs. Which Amazon Bedrock feature provides ALL of these controls in a single configuration?

Accepted Answer

Amazon Bedrock Guardrails provides a single configurable layer that can block denied topics, filter harmful content across categories, redact sensitive PII from inputs and outputs, and prevent prompt attacks — all in one configuration applied to any model on Bedrock. Agents with action groups define what tools an agent can call, not content safety policies. Knowledge Bases metadata filtering limits which documents are retrieved, not what the model can say. Amazon SageMaker AI Model Monitor tracks model quality and data drift in production ML endpoints, not generative AI content safety.

Question 19

A company's AI team wants to use Amazon Bedrock to build an application that requires the model to retain context across many turns of a long conversation. The team is concerned that very long conversations will exceed the model's context window. Which approach MOST effectively mitigates this limitation?

Accepted Answer

Selecting a model with a larger context window allows more conversation history to fit in a single prompt. When that is not sufficient, implementing conversation summarization — replacing older turns with a compressed summary — preserves key context while reducing token count. Increasing temperature controls randomness, not context capacity. Model Evaluation compares models but does not solve the context window problem in production. Text-to-speech converts output format and is irrelevant to context window limitations.

Question 20

A company is evaluating the fairness of its AI model used to screen job applicants. The team discovers the model recommends male candidates at a significantly higher rate than equally qualified female candidates. Which responsible AI principle does this finding violate?

Accepted Answer

Fairness in AI requires that models do not produce biased outcomes that discriminate against individuals based on protected characteristics such as gender. A model that favors male candidates over equally qualified female candidates violates this principle. Transparency relates to explainability of how a model makes decisions. Privacy concerns protecting individuals' personal data. Veracity (or accuracy/robustness) relates to producing correct and reliable outputs.

Question 21

A company is deploying an AI model that makes loan approval decisions. Regulators require the company to be able to explain why a specific loan application was denied. Which responsible AI requirement does this represent?

Accepted Answer

Explainability (or interpretability) requires that AI systems can provide understandable reasons for their decisions — a regulatory requirement for high-stakes decisions such as loan denials under laws like the Equal Credit Opportunity Act. Fairness ensures the model does not discriminate. Robustness ensures the model performs reliably under varied conditions. Privacy concerns protecting personal data used in training or inference.

Question 22

An organization wants to use Amazon SageMaker AI to monitor its deployed production ML model for signs that the model's predictions are degrading over time because real-world input data is shifting. Which AWS capability addresses this need?

Accepted Answer

Amazon SageMaker AI Model Monitor continuously monitors deployed models for data drift, model quality drift, bias drift, and feature attribution drift, alerting teams when production data deviates from training baselines. Amazon SageMaker AI Debugger helps debug training jobs. Amazon SageMaker AI Clarify detects bias in datasets and models at training time and explains model predictions. Amazon SageMaker AI Experiments tracks and compares ML training runs.

Question 23

A company is training an AI image recognition model for a healthcare application. The AI team discovers that the training dataset contains significantly more images of lighter-skinned patients than darker-skinned patients. Which responsible AI risk does this imbalance MOST directly introduce?

Accepted Answer

Training data imbalance across demographic groups causes the model to learn more accurately for over-represented groups and perform poorly for under-represented groups — a classic case of algorithmic bias leading to disparate impact and potential harm in a healthcare context. Prompt injection is an adversarial attack on generative AI models. Model hallucination refers to LLMs generating false information. Underfitting relates to the model's complexity being insufficient to learn the training data patterns, not demographic representation.

Question 24

A company is building a generative AI application using Amazon Bedrock and must ensure that users cannot manipulate the AI to bypass safety controls by embedding malicious instructions inside documents the AI is asked to process. Which type of attack does this describe, and which Amazon Bedrock feature MOST directly helps defend against it?

Accepted Answer

Prompt injection is an attack where malicious instructions are embedded in input content (such as user-provided documents) to manipulate the LLM into ignoring its system prompt or safety rules. Amazon Bedrock Guardrails helps defend against prompt injection by detecting and blocking attempts to circumvent model instructions. Data poisoning corrupts training data, not runtime inputs. Model inversion is a privacy attack to reconstruct training data. DDoS attacks target service availability, not LLM behavior.

Question 25

A financial services company is using Amazon Bedrock to process customer financial data. The compliance team requires that data sent to foundation models is not used to train or improve the underlying models and that all data remains within the company's AWS environment. Which approach BEST meets both requirements?

Accepted Answer

Amazon Bedrock does not use customer inputs or outputs to train the underlying foundation models, satisfying the data privacy compliance requirement. Using Amazon VPC endpoints (AWS PrivateLink) ensures that all traffic between the company's AWS environment and Amazon Bedrock stays on the private AWS network and never traverses the public internet. Option A's default API settings allow public internet traffic. Option C (self-hosted EC2 LLM) achieves isolation but requires significant infrastructure management. Option D (SageMaker AI JumpStart) requires downloading and managing model weights, adding significant operational overhead.

AIF-C01 — AWS Certified AI Practitioner