Knowledge + Practice

AWS Certified AI Practitioner AIF-C01 (AIF-C01) — Questions 226–300

619 questions total · 9pages · All types, answers revealed

Take a mock exam Exam hub

Page 4 of 9

226

Multi-Selecteasy

A data scientist is setting up a RAG pipeline using Amazon Bedrock Knowledge Bases. They need to ingest documents, split them into chunks, generate embeddings, and store them for retrieval. Which TWO steps are part of the ingestion process? (Choose TWO)

Select 2 answers

A.Define a Lambda function for post-processing

B.Perform a similarity search on the vector store

C.Augment the user prompt with retrieved chunks

D.Parse the documents and split them into chunks

E.Generate embeddings for each chunk using a Bedrock embedding model

AnswersD, E

Chunking is a key ingestion step to divide documents into manageable pieces.

Why this answer

Ingestion includes parsing documents, chunking, and generating embeddings. Querying the vector store and augmenting the prompt occur during inference, not ingestion.

Full explanation →

227

MCQmedium

A data governance team wants to enforce fine-grained access control on data in an Amazon S3 data lake used by multiple business units for AI training. Which AWS service should they use to define and manage data permissions at the table and column level?

A.AWS Identity and Access Management (IAM)

B.AWS Lake Formation

C.AWS CloudTrail

D.Amazon Macie

AnswerB

Lake Formation provides column-level and row-level security for data in S3 data lakes.

Why this answer

AWS Lake Formation provides fine-grained access control for data lakes, allowing you to set permissions on tables and columns across AWS analytics and ML services.

Full explanation →

228

MCQhard

Refer to the exhibit. A team is configuring a SageMaker Model Bias job. The baseline job has been completed. However, the bias job fails with a resource not found error. What is the most likely cause?

A.The StoppingCondition is too short

B.The BaseliningJobName is incorrect

C.The instance type ml.m5.large is not supported

D.The IAM role lacks permissions to DescribeBaselineJob

AnswerB

Typo or mismatch in the baseline job name.

Why this answer

The bias job requires a reference to the completed baseline job to compare the training data against. If the BaseliningJobName parameter is incorrect or does not match the actual name of the completed baseline job, SageMaker will throw a 'ResourceNotFound' error because it cannot locate the specified baseline job. The error is not related to timeouts, instance types, or IAM permissions for describing the baseline job.

Exam trap

AWS often tests the distinction between different error types (timeout vs. resource not found vs. permission denied) to see if candidates understand the specific cause-and-effect relationship between misconfigured parameters and the exact error message returned.

How to eliminate wrong answers

Option A is wrong because a StoppingCondition that is too short would cause a timeout error, not a 'resource not found' error. Option C is wrong because ml.m5.large is a supported instance type for SageMaker processing jobs, including bias jobs. Option D is wrong because the IAM role lacking permissions to DescribeBaselineJob would result in an access denied or authorization error, not a 'resource not found' error.

Full explanation →

229

MCQhard

An organization uses a third-party foundation model accessed through Amazon Bedrock. The compliance team requires that all model inputs and outputs be auditable and retained for one year. Which approach should the team implement?

A.Store all prompts and responses in a DynamoDB table via custom code in the application

B.Use VPC Flow Logs to capture network traffic to Bedrock

C.Enable Bedrock model invocation logging and configure destination to S3 with a lifecycle policy to retain logs for one year

D.Enable CloudTrail data events for Bedrock

AnswerC

This logs the inputs and outputs and retains them for the required period.

Why this answer

Bedrock model invocation logging captures all requests and responses and can be sent to CloudWatch Logs or S3 for retention. CloudTrail logs only API-level events (e.g., InvokeModel calls) but not the actual payloads. The other options do not provide payload logging.

Full explanation →

230

MCQmedium

A team is evaluating two generative AI models for a summarization task. They have reference summaries and generated summaries. Which automated metric is BEST suited to measure the overlap of n-grams between the generated and reference summaries?

A.ROUGE

B.BLEU

C.BERTScore

D.Perplexity

AnswerA

Correct. ROUGE is designed for summarization and measures recall of n-grams.

Why this answer

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures n-gram overlap between generated and reference text, and is commonly used for summarization. BLEU is for translation, BERTScore uses embeddings, and perplexity measures language model confidence.

Full explanation →

231

MCQhard

A financial services company uses a foundation model for document analysis. They need to ensure the model does not output sensitive customer information from its training data. What is the most effective mitigation?

A.Implement output filtering using an external service

B.Choose a model that has been fine-tuned on financial data

C.Apply data masking before sending input

D.Use a private endpoint

AnswerA

Output filtering can scan and block responses containing sensitive data.

Why this answer

Output filtering using an external service is the most effective mitigation because it acts as a post-processing layer that can detect and redact sensitive customer information (e.g., PII, account numbers) before the model's response is returned to the user. This approach does not rely on the model's internal training or input modifications, which can be bypassed or incomplete. It provides a robust, policy-driven control that can be updated independently of the model.

Exam trap

The trap here is that candidates often confuse input-side controls (like data masking or fine-tuning) with output-side controls, assuming that protecting the input or training the model on domain data is sufficient to prevent leakage of memorized sensitive information.

How to eliminate wrong answers

Option B is wrong because fine-tuning on financial data does not guarantee the model will not memorize or regurgitate sensitive customer information from its original training data; fine-tuning adjusts the model's behavior but does not erase existing memorized data. Option C is wrong because data masking before sending input only protects the input data, not the model's outputs; the model could still output sensitive information from its training data that was never masked. Option D is wrong because using a private endpoint secures the network connection and access control but does not prevent the model from generating outputs containing sensitive training data; it addresses data-in-transit security, not output content safety.

Full explanation →

232

MCQhard

A company uses a diffusion model on Amazon Bedrock to generate marketing images. They notice that the generated images often contain artifacts and lack fine details, especially when the prompt is complex. The team wants to improve image quality without increasing inference time significantly. Which parameter adjustment is MOST likely to help?

A.Increase the number of inference steps

B.Increase the seed to introduce randomness

C.Decrease the output resolution to reduce artifacts

D.Reduce the guidance scale to allow more creative freedom

AnswerA

More inference steps allow the model to iteratively refine the image, reducing artifacts and improving detail.

Why this answer

Increasing the number of inference steps allows the diffusion model to iteratively refine the generated image, reducing artifacts and improving fine details. This directly addresses the problem of complex prompts producing lower-quality outputs without a significant increase in inference time, as the trade-off is typically sub-linear.

Exam trap

The AIF-C01 exam often tests the misconception that more randomness (seed) or lower resolution can fix quality issues, when the core mechanism for detail refinement in diffusion models is the number of inference steps.

How to eliminate wrong answers

Option B is wrong because increasing the seed only changes the initial random noise, not the quality of the generation; it may produce a different image but does not reduce artifacts or improve detail. Option C is wrong because decreasing output resolution reduces the pixel count, which can mask artifacts but does not eliminate them; it also reduces fine details, contradicting the goal of improving image quality. Option D is wrong because reducing the guidance scale gives the model more creative freedom, which often leads to less adherence to the prompt and can increase artifacts, not reduce them.

Full explanation →

233

MCQhard

A company has deployed a model on Amazon SageMaker and enabled Model Monitor. They notice that the model's prediction accuracy has declined over time. Which type of drift is this, and what should they do?

A.This is bias drift; they should run SageMaker Clarify

B.This is concept drift; they should retrain the model with new data

C.This is model drift; they should delete and redeploy the model

D.This is data drift; they should update the training data

AnswerB

Concept drift occurs when the relationship between input and output changes, leading to accuracy decline. Retraining addresses it.

Why this answer

Model drift refers to the degradation of model performance over time due to changes in data or relationships. SageMaker Model Monitor can detect data drift, but model drift (accuracy decline) requires retraining or updating the model.

Full explanation →

234

MCQmedium

A data scientist is comparing two foundation models for a text classification task. They want to use automated metrics to evaluate performance on a labelled test set. Which metric is most appropriate for a multi-class classification problem?

A.ROUGE-L

B.BLEU

C.Perplexity

D.F1 score

AnswerD

F1 score is suitable for classification tasks.

Why this answer

F1 score is a standard metric for classification, balancing precision and recall. ROUGE and BLEU are for text generation, Perplexity for language modelling.

Full explanation →

235

MCQhard

A bank is deploying a credit scoring model and must comply with regulatory requirements that decisions can be explained to customers. The model is a gradient boosting machine with hundreds of features. Which explainability technique should the team use to provide local explanations for individual loan decisions?

A.SHAP values

B.Global feature importance based on permutation

C.Partial dependence plots

D.LIME

AnswerA

SHAP values assign each feature a contribution to an individual prediction, providing local explanations grounded in game theory.

Why this answer

SHAP values provide a unified measure of feature importance for individual predictions, satisfying local explainability requirements. LIME is also local but is less stable and not game-theoretic. Global feature importance does not explain individual decisions.

Full explanation →

236

MCQeasy

Which AWS service can be used to extract text and data from scanned documents such as invoices and receipts?

A.Amazon Transcribe

B.Amazon Textract

C.Amazon Comprehend

D.Amazon Rekognition

AnswerB

Textract is specifically designed to extract text and data from scanned documents.

Why this answer

Amazon Textract is a machine learning service specifically designed to extract text, handwriting, and structured data (like tables and forms) from scanned documents. It goes beyond simple optical character recognition (OCR) by using ML models to understand the layout and relationships within documents, making it ideal for processing invoices and receipts.

Exam trap

The trap here is that candidates may confuse Amazon Rekognition's ability to detect text in images with the specialized document extraction capabilities of Amazon Textract, but Rekognition lacks the structured data extraction and layout analysis needed for invoices and receipts.

How to eliminate wrong answers

Option A is wrong because Amazon Transcribe is an automatic speech recognition (ASR) service that converts audio to text, not a service for extracting data from scanned documents. Option C is wrong because Amazon Comprehend is a natural language processing (NLP) service that analyzes text for entities, sentiment, and key phrases, but it cannot extract text from images or scanned files. Option D is wrong because Amazon Rekognition is primarily an image and video analysis service for detecting objects, faces, and scenes, and while it can detect text in images via its DetectText API, it is not optimized for extracting structured data from documents like invoices and receipts, lacking the layout analysis and form understanding that Textract provides.

Full explanation →

237

Multi-Selectmedium

A data scientist is building a text classification system using Amazon Bedrock. They want to evaluate different foundation models for accuracy and latency. Which TWO approaches are appropriate for comparing models? (Select TWO.)

Select 2 answers

A.Build a custom model from scratch that combines the outputs of all models

B.Invoke several models with a sample set of queries and compare the outputs manually

C.Read the model documentation and assume the one with the largest parameter count is best

D.Create a test dataset with labeled ground truth, run predictions, and compare accuracy metrics

E.Select the model with the lowest cost per token without testing

AnswersB, D

Manual evaluation on sample queries provides direct insight into output quality.

Why this answer

Invoking several models with a sample set of queries and comparing outputs manually allows for direct, qualitative assessment of response quality and latency under realistic conditions. This approach is practical for initial model selection in Amazon Bedrock, where you can test multiple FMs via the InvokeModel API without committing to a single model.

Exam trap

AWS often tests the misconception that parameter count or cost alone determines model suitability, but the correct approach requires empirical testing with representative data and metrics.

Full explanation →

238

MCQeasy

A company is using Amazon Bedrock to deploy a foundation model. To comply with GDPR, they need to ensure that the model does not generate outputs containing personally identifiable information (PII). Which AWS service can best help detect and redact PII from the model's responses?

A.Amazon Comprehend

B.Amazon GuardDuty

C.Amazon Rekognition

D.AWS WAF

AnswerA

Amazon Comprehend has PII detection and redaction features.

Why this answer

Amazon Comprehend is the correct service because it provides a built-in PII detection and redaction capability that can be integrated with Amazon Bedrock. Using the `DetectPiiEntities` API, you can scan model responses for PII such as names, addresses, and credit card numbers, and then redact or mask those entities before returning the output to the user. This directly addresses the GDPR requirement to prevent PII leakage from generative AI outputs.

Exam trap

The AIF-C01 exam often tests the distinction between security monitoring services (GuardDuty, WAF) and content analysis services (Comprehend), leading candidates to mistakenly choose a network-level security tool for a data content problem.

How to eliminate wrong answers

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 service for detecting or redacting PII in text content. Option C is wrong because Amazon Rekognition is an image and video analysis service that can detect faces, objects, and text in media, but it does not provide PII detection or redaction for text-based model responses. Option D is wrong because AWS WAF is a web application firewall that protects against common web exploits like SQL injection and cross-site scripting, and it has no capability to analyze or redact PII from AI model outputs.

Full explanation →

239

MCQhard

A data scientist wants to perform automatic model tuning (hyperparameter optimization) on SageMaker. They need to find the best hyperparameters for a gradient boosting model. Which strategy is BEST for this task?

A.Random search

B.Grid search

C.Exhaustive search

D.Bayesian optimization

AnswerD

Uses a probabilistic model to select hyperparameters, achieving better results with fewer iterations.

Why this answer

Bayesian optimization is the best strategy for automatic model tuning on SageMaker because it builds a probabilistic model of the objective function and uses it to select the most promising hyperparameters to evaluate next. This approach is far more sample-efficient than random or grid search, making it ideal for expensive-to-evaluate models like gradient boosting, where each training run consumes significant time and compute resources.

Exam trap

The AIF-C01 exam often tests the misconception that exhaustive or grid search is the most thorough and therefore best approach, but the trap is that they ignore the practical constraints of compute cost and time, making Bayesian optimization the superior choice for automatic model tuning in SageMaker.

How to eliminate wrong answers

Option A is wrong because random search, while better than grid search for high-dimensional spaces, does not use past evaluation results to inform future trials, making it less efficient than Bayesian optimization for finding optimal hyperparameters. Option B is wrong because grid search exhaustively evaluates all combinations of a predefined set of hyperparameter values, which is computationally prohibitive for gradient boosting models with many continuous hyperparameters and does not scale well. Option C is wrong because exhaustive search is essentially a synonym for grid search and suffers from the same curse of dimensionality, making it impractical for hyperparameter optimization in SageMaker's automatic model tuning context.

Full explanation →

240

MCQhard

A startup is fine-tuning a large language model (LLM) for code generation using Amazon SageMaker. They are using a p4d.24xlarge instance with a single GPU. The training process is extremely slow, taking over 48 hours for one epoch. The dataset is 10GB of code snippets. The company needs to iterate quickly. Which action would most significantly reduce training time without sacrificing model quality?

A.Enable distributed training using SageMaker’s data parallelism library across multiple GPUs

B.Switch to spot instances to reduce cost, not time

C.Increase the batch size to use GPU memory more efficiently

D.Use a smaller foundation model to reduce compute per step

AnswerA

Distributed training scales across GPUs/nodes, significantly speeding up training while preserving model size.

Why this answer

Distributed training across multiple GPUs and instances dramatically reduces time by parallelizing the workload. Increasing instance count or using a smaller model helps but may not be optimal. Spot instances could be unstable.

Data parallelism is a standard technique for large models.

Full explanation →

241

Multi-Selecthard

Which THREE are best practices for ensuring generated content complies with corporate brand guidelines when using Amazon Bedrock?

Select 3 answers

A.Implement guardrails to restrict tone, topics, and language

B.Use prompt engineering to specify brand voice and style

C.Increase the temperature for more creative outputs

D.Use random prompts to test variability

E.Fine-tune the model on a dataset of brand-compliant content

AnswersA, B, E

Guardrails enforce content policies at inference time.

Why this answer

Amazon Bedrock Guardrails allow you to define policies that restrict the model's output to specific tones, topics, and language, ensuring alignment with corporate brand guidelines. By configuring denied topics and content filters, you can prevent the model from generating off-brand or inappropriate content, directly enforcing compliance at the inference layer.

Exam trap

AWS often tests the misconception that increasing temperature or using random prompts can help enforce brand guidelines, when in fact these actions increase variability and reduce control, directly opposing the goal of compliance.

Full explanation →

242

Multi-Selectmedium

A company is deploying an LLM for generating marketing copy. They want to reduce the risk of hallucinations and ensure the content is factually accurate. Which TWO approaches should they implement?

Select 2 answers

A.Use Retrieval-Augmented Generation (RAG) with a verified knowledge base

B.Implement human review of generated content before publication

C.Use a smaller LLM to reduce the chance of hallucinations

D.Disable all guardrails to avoid interference

E.Increase the temperature parameter to generate more diverse outputs

AnswersA, B

RAG helps provide factual grounding.

Why this answer

RAG grounds the model in retrieved facts, and human review provides oversight to catch inaccuracies.

Full explanation →

243

MCQmedium

A company fine-tunes a foundation model on SageMaker using a custom dataset. They notice the training job takes too long. Which optimization technique is specifically designed to reduce training time for foundation models?

A.Distributed training using SageMaker Data Parallelism

B.Using a smaller instance type

C.Using Spot Instances

D.Reducing batch size

AnswerA

Data parallelism partitions the data and trains across multiple devices, reducing wall-clock time.

Why this answer

SageMaker Data Parallelism distributes the training workload across multiple GPUs or instances, splitting the data and synchronizing gradients using optimized all-reduce algorithms. This specifically reduces training time for large foundation models by enabling parallel computation, which is the most direct technique for accelerating training at scale.

Exam trap

AWS often tests the misconception that cost-saving techniques like Spot Instances or smaller instances also improve performance, but the question specifically asks for optimization to reduce training time, not cost.

How to eliminate wrong answers

Option B is wrong because using a smaller instance type reduces computational capacity, which would increase training time rather than reduce it. Option C is wrong because Spot Instances reduce cost by using spare AWS capacity, but they do not inherently speed up training; they may even cause interruptions that prolong total time. Option D is wrong because reducing batch size can actually slow convergence and increase the number of training steps, potentially increasing overall training time.

Full explanation →

244

Multi-Selectmedium

A company is using Amazon Bedrock to generate marketing content. They want to evaluate the quality of the generated text. Which TWO metrics are most appropriate for evaluating text quality?

Select 2 answers

A.Precision

B.Perplexity

C.Accuracy

D.F1 score

E.BLEU (Bilingual Evaluation Understudy)

AnswersB, E

Perplexity measures how well the model predicts the text.

Why this answer

Perplexity measures how well a language model predicts a sample, with lower values indicating higher confidence and coherence in generated text. BLEU evaluates the overlap between generated text and reference text, making it suitable for assessing fluency and relevance in content generation tasks like marketing copy.

Exam trap

AWS often tests the distinction between classification metrics (precision, accuracy, F1) and generation evaluation metrics (perplexity, BLEU), leading candidates to mistakenly apply classification concepts to text quality assessment.

Full explanation →

245

MCQeasy

A data scientist needs to train a model in Amazon SageMaker using a dataset that contains personally identifiable information (PII). The company policy requires all data at rest to be encrypted with a customer-managed key. Which configuration meets this requirement?

A.Enable default encryption for the S3 bucket using AWS-managed S3 keys (SSE-S3)

B.Specify a KMS customer-managed key when creating the SageMaker training job and enable data encryption for the S3 bucket with the same key

C.Encrypt the data before uploading to S3 using a client-side library

D.Use SageMaker's local mode and store data on the instance's ephemeral storage

AnswerB

This encrypts both the training data and the model artifacts with a customer-managed key.

Why this answer

Using a KMS customer-managed key (CMK) for SageMaker's EBS volumes and S3 buckets ensures encryption at rest with a key the customer controls. The other options either use AWS-managed keys, skip encryption, or are not applicable to SageMaker training jobs.

Full explanation →

246

MCQhard

A financial services company needs to deploy a real-time fraud detection model with sub-100ms inference latency. The model is a large ensemble requiring 8 GB of memory per request. The workload has bursty traffic. Which Amazon SageMaker deployment strategy best meets these requirements?

A.Deploy behind an Application Load Balancer with multiple ml.m5.xlarge EC2 instances running the model

B.Use a single ml.r5.2xlarge instance with an auto-scaling policy based on CPU utilization

C.Use a SageMaker multi-model endpoint with ml.m5.large instances to cache multiple models

D.Use SageMaker asynchronous inference with a large batch size

AnswerB

A real-time endpoint with a large instance and auto-scaling handles bursty traffic and meets latency requirements.

Why this answer

A single ml.r5.2xlarge instance provides 16 GB of memory, which can handle the 8 GB per request requirement, and SageMaker real-time endpoints with auto-scaling based on CPU utilization can dynamically adjust to bursty traffic while maintaining sub-100ms inference latency. This approach avoids the overhead of load balancers or multi-model caching that could introduce latency.

Exam trap

The trap here is that candidates may assume multi-model endpoints (Option C) are suitable for large models, but they are designed for many small models sharing memory, not for a single large ensemble requiring 8 GB per request.

How to eliminate wrong answers

Option A is wrong because deploying behind an Application Load Balancer with multiple ml.m5.xlarge instances adds network hop latency and does not leverage SageMaker's native endpoint routing, potentially exceeding the sub-100ms requirement; also, m5.xlarge instances have only 8 GB of memory, which may not handle the 8 GB per request without memory contention. Option C is wrong because SageMaker multi-model endpoints are designed for serving multiple smaller models from a shared instance, not for a single large ensemble requiring 8 GB per request, and ml.m5.large instances have only 4 GB of memory, insufficient for the workload. Option D is wrong because SageMaker asynchronous inference is intended for non-real-time, large payloads with minutes of latency, not sub-100ms real-time fraud detection, and batching would increase latency beyond the requirement.

Full explanation →

247

MCQhard

A developer sends the above request to Amazon Bedrock with Anthropic Claude. The model returns a response that stops before reaching 500 tokens. What is the most likely reason?

A.The temperature is set too high

B.The model is not trained on this topic

C.The model reached a stop sequence

D.The token limit is exceeded

AnswerC

The model can stop early when it identifies a natural endpoint.

Why this answer

The model stopped before reaching 500 tokens because the request likely included a stop sequence (e.g., `\n\nHuman:` or a custom stop token) that matched the generated output. When a stop sequence is encountered, Bedrock immediately halts generation, even if the token limit has not been reached. This is the most direct explanation for a premature stop.

Exam trap

AWS often tests the distinction between a stop sequence and a token limit; the trap here is that candidates confuse a premature stop with exceeding the token limit, but a stop sequence causes an early halt while a token limit would cause truncation at the limit.

How to eliminate wrong answers

Option A is wrong because a high temperature increases randomness and can cause the model to generate more tokens or diverge, not stop early. Option B is wrong because Bedrock's Claude models are trained on a broad corpus and can generate responses on any topic; lack of training would produce low-quality or repetitive text, not a stop before the token limit. Option D is wrong because if the token limit were exceeded, the model would truncate the response at the limit, not stop before reaching it.

Full explanation →

248

MCQeasy

Which of the following is a key principle of inclusive design in AI?

A.Deploy AI systems without user testing

B.Use only data from a single demographic group

C.Ensure the AI system is usable by people with diverse abilities and backgrounds

D.Maximize model complexity for better performance

AnswerC

Inclusive design focuses on accessibility and diversity.

Why this answer

Inclusive design aims to create products that are usable by people with diverse abilities and backgrounds, considering accessibility and user needs.

Full explanation →

249

MCQhard

A company uses Amazon SageMaker for model training. To comply with data residency requirements, they must ensure that the training data never leaves a specific AWS region. However, during training, the SageMaker service might use resources in other regions for auto-scaling. Which configuration should they use to enforce data residency?

A.Configure the training job to use only local spot instances and enable network isolation.

B.Use Amazon SageMaker's inter-container traffic encryption and disable cross-region data transfer.

C.Use AWS Organizations to create an SCP that denies access to SageMaker resources in other regions.

D.Use a VPC with a VPC endpoint for SageMaker and restrict the training job to use only local resources.

AnswerC

SCPs can explicitly deny SageMaker actions in non-compliant regions, enforcing data residency.

Why this answer

AWS Organizations Service Control Policies (SCPs) can explicitly deny access to SageMaker resources in any region outside the allowed one. By attaching an SCP that denies `sagemaker:*` actions when the `aws:RequestedRegion` condition key does not match the permitted region, the company can enforce data residency at the account level, preventing SageMaker from provisioning resources in other regions even if auto-scaling would otherwise trigger cross-region activity.

Exam trap

The trap here is that candidates often assume VPC endpoints or network isolation are sufficient to enforce regional boundaries, but they do not control the SageMaker control plane's ability to launch resources in other regions; only an SCP or IAM policy with a region condition can enforce that restriction at the API level.

How to eliminate wrong answers

Option A is wrong because using local spot instances and network isolation only restricts the instance type and network access; it does not prevent SageMaker from launching training resources in other regions for auto-scaling or data processing. Option B is wrong because inter-container traffic encryption secures data in transit between containers but does not control the geographic location of the compute resources; disabling cross-region data transfer is not a configurable SageMaker setting. Option D is wrong because a VPC with a VPC endpoint for SageMaker restricts network traffic to the service endpoint within the VPC, but SageMaker can still launch training jobs in other regions if the training job configuration or service backend decides to use resources outside the local region; the VPC endpoint does not enforce regional boundaries on SageMaker's resource provisioning.

Full explanation →

250

Multi-Selectmedium

A data scientist is preparing to fine-tune an Amazon Titan model for a domain-specific text classification task. Which THREE components are essential for the fine-tuning process on Amazon Bedrock? (Choose THREE.)

Select 3 answers

A.Hyperparameter configuration (e.g., learning rate, batch size)

B.A separate inference endpoint for testing during training

C.Training dataset in JSONL format with prompt and completion

D.Validation dataset for model evaluation

E.Model architecture code for the Titan model

AnswersA, C, D

Hyperparameters define the training process and are required for fine-tuning.

Why this answer

Hyperparameters such as learning rate, batch size, and number of epochs directly control the optimization process during fine-tuning on Amazon Bedrock. These settings determine how the model updates its weights based on the training data, and they must be configured to achieve convergence without overfitting or underfitting.

Exam trap

The trap here is that candidates may confuse the need for a separate inference endpoint with the ability to evaluate during training, but Bedrock allows evaluation using the validation dataset without an extra endpoint, and they may also mistakenly think they need to provide model architecture code when Bedrock abstracts that away as a managed service.

Full explanation →

251

MCQmedium

A company is using Amazon Textract to extract text from scanned documents stored in an S3 bucket. The security team requires that all access to the documents be logged and that the documents be encrypted at rest using a customer-managed key. What should the company do to meet these requirements?

A.Use S3 default encryption and enable Textract logging

B.Enable S3 server-side encryption with AWS KMS (SSE-KMS) and enable CloudTrail data events for the S3 bucket

C.Enable S3 server access logs and use S3 SSE-KMS

D.Use S3 server-side encryption with S3-managed keys (SSE-S3) and enable S3 access logs

AnswerB

SSE-KMS provides encryption with customer-managed keys; CloudTrail data events log access to objects.

Why this answer

Enabling S3 server-side encryption with AWS KMS (SSE-KMS) satisfies the requirement for encryption at rest using a customer-managed key, and enabling CloudTrail data events for the S3 bucket captures all access to the documents (including GetObject, PutObject, etc.) for logging. This combination meets both security requirements precisely.

Exam trap

The trap here is that candidates often confuse S3 server access logs with CloudTrail data events, assuming both provide equivalent logging, but only CloudTrail data events offer reliable, real-time, and comprehensive object-level access logging required for security audits.

How to eliminate wrong answers

Option A is wrong because S3 default encryption uses either SSE-S3 or SSE-KMS, but it does not specify customer-managed keys, and enabling Textract logging only logs Textract API calls, not S3 data access events. Option C is wrong because S3 server access logs provide access logs but are delivered on a best-effort basis with potential delays and do not capture all API-level details like CloudTrail data events; also, while SSE-KMS is used, the logging mechanism is insufficient for comprehensive audit requirements. Option D is wrong because SSE-S3 uses AWS-managed keys, not customer-managed keys, and S3 access logs are not as granular or reliable as CloudTrail data events for logging all access.

Full explanation →

252

Multi-Selecthard

A data science team is fine-tuning a foundation model on Amazon SageMaker. Which THREE steps are part of the best practice? (Choose three.)

Select 3 answers

A.Increase model size to improve performance.

B.Monitor for catastrophic forgetting during fine-tuning.

C.Use early stopping to prevent overfitting.

D.Deploy the model to production immediately after fine-tuning.

E.Use a diverse dataset representing various scenarios.

AnswersB, C, E

Catastrophic forgetting can cause loss of original capabilities; monitoring helps adjust training.

Why this answer

Catastrophic forgetting is a known risk when fine-tuning foundation models, where the model loses previously learned knowledge while adapting to new data. Monitoring for this during fine-tuning on SageMaker allows the team to detect performance degradation on the original task and adjust training accordingly, ensuring the model retains its general capabilities.

Exam trap

AWS often tests the misconception that fine-tuning always requires a larger model or immediate deployment, while the real best practices focus on validation, monitoring, and data diversity to maintain model robustness.

Full explanation →

253

MCQhard

A developer implements a RAG pipeline with Amazon Bedrock and Amazon OpenSearch Serverless. Users report that the chatbot sometimes returns off-topic responses. Investigation shows the retrieved chunks are semantically unrelated to the queries. What is the MOST likely cause?

A.The LLM used for generation has too few parameters

B.The agent's action group has misconfigured Lambda functions

C.The vector store index is not encrypted at rest

D.The chunking strategy is not optimised for the document type, or the embedding model is not well-suited

AnswerD

Inappropriate chunking or a weak embedding model leads to poor semantic search results.

Why this answer

Poor embedding quality or chunking strategy can cause irrelevant retrieval. The other options relate to generation or permission issues.

Full explanation →

254

MCQmedium

An e-commerce company uses Amazon Bedrock to generate product descriptions. They notice the descriptions are too long and contain repetitive phrases. Which parameter adjustment can help?

A.Increase frequency penalty

B.Increase temperature

C.Increase top_p

D.Decrease presence penalty

AnswerA

Frequency penalty reduces the likelihood of repeating tokens.

Why this answer

Increasing the frequency penalty reduces the likelihood of the model repeating the same phrases or tokens, directly addressing the issue of repetitive language in generated product descriptions. This parameter penalizes tokens that have already appeared in the text, encouraging more diverse output and naturally shortening overly long descriptions by avoiding redundant loops.

Exam trap

AWS often tests the distinction between frequency penalty and presence penalty, where candidates confuse 'penalizing repetition' with 'reducing randomness' and incorrectly choose temperature or top_p adjustments.

How to eliminate wrong answers

Option B is wrong because increasing temperature makes the model more random and creative, which could actually worsen verbosity and introduce more irrelevant phrases rather than reducing repetition. Option C is wrong because increasing top_p (nucleus sampling) expands the set of possible next tokens, which may increase diversity but does not specifically penalize repeated tokens and can still produce long, repetitive text. Option D is wrong because decreasing presence penalty would reduce the penalty for tokens that have already appeared, making the model more likely to repeat itself, which is the opposite of what is needed.

Full explanation →

255

MCQmedium

A research organization is using Amazon SageMaker Studio to collaborate on building machine learning models. The security policy requires that all data and code remain within a VPC and cannot be accessed from the public internet. Additionally, the organization wants to enforce that only approved base images are used for the Studio environment. How should the organization configure SageMaker Studio to meet these requirements?

A.Use an AWS Transit Gateway to connect the VPC, and enforce HTTPS for all traffic.

B.Configure the Studio domain in VPC Only mode and use a SageMaker Studio lifecycle configuration to restrict the list of available base images to those in a private Amazon ECR repository.

C.Configure the Studio domain to disable direct internet access, and let users choose any base image from the SageMaker public registry.

D.Use AWS CloudFormation to create a Studio domain in a VPC and rely on individuals to use approved images.

AnswerB

VPC Only ensures no public internet access, and lifecycle configurations can restrict images to approved ones from a private ECR.

Why this answer

Configuring the SageMaker Studio domain in VPC Only mode ensures that all data and code remain within the VPC and cannot be accessed from the public internet. Additionally, using a lifecycle configuration script to restrict available base images to those in a private Amazon ECR repository enforces the policy that only approved base images are used, as the lifecycle configuration can modify the Jupyter Server settings to limit the image registry.

Exam trap

The trap here is that candidates often confuse disabling direct internet access with fully restricting image choices, not realizing that without a lifecycle configuration to filter the image registry, users can still select any public SageMaker image that is available within the VPC's network scope.

How to eliminate wrong answers

Option A is wrong because AWS Transit Gateway is used to connect multiple VPCs or on-premises networks, not to enforce VPC-only access for SageMaker Studio, and enforcing HTTPS for all traffic does not restrict base images or prevent public internet access. Option C is wrong because disabling direct internet access alone does not restrict users from choosing any base image from the SageMaker public registry; it only prevents outbound internet traffic, but users could still select public images that are cached or accessible within the VPC. Option D is wrong because using AWS CloudFormation to create a Studio domain in a VPC does not enforce the use of approved images; relying on individuals to use approved images is a manual process that does not meet the security policy requirement.

Full explanation →

256

MCQhard

A team is building an AI chatbot that will be used by customers with visual impairments. Which design practice best supports inclusive accessibility for this user group?

A.Add captions to all video content

B.Provide a voice-only interface without any text output

C.Ensure the chatbot is fully navigable via keyboard and compatible with screen readers

D.Use high-contrast color schemes and large fonts

AnswerC

Keyboard navigation and screen reader compatibility are essential for users who are blind or have low vision, allowing them to interact with the interface using assistive technologies.

Why this answer

Ensuring full keyboard navigability and compatibility with screen readers is a fundamental accessibility requirement for users with visual impairments, enabling them to interact without relying on visual cues.

Full explanation →

257

MCQmedium

A company runs a chatbot using a large language model on Amazon Bedrock. They notice high latency during peak hours. Which action would be MOST effective to reduce latency without degrading response quality?

A.Increase the number of concurrent invocations

B.Switch to a smaller model

C.Decrease the maxTokens parameter

D.Use Provisioned Throughput for model inference

AnswerD

Provisioned Throughput ensures reserved capacity, reducing latency variability.

Why this answer

Provisioned Throughput on Amazon Bedrock reserves dedicated capacity for model inference, ensuring consistent low latency even during peak hours. This eliminates the variability caused by resource contention in the on-demand tier, directly addressing high latency without altering model size or output quality.

Exam trap

AWS often tests the misconception that reducing model size or output length is the primary way to reduce latency, but the real bottleneck in peak-hour scenarios is often infrastructure contention, which Provisioned Throughput resolves without sacrificing quality.

How to eliminate wrong answers

Option A is wrong because increasing concurrent invocations without dedicated capacity can exacerbate resource contention, leading to throttling and higher latency. Option B is wrong because switching to a smaller model reduces response quality (e.g., lower accuracy or coherence), which degrades the chatbot's performance. Option C is wrong because decreasing maxTokens truncates responses, degrading output quality by cutting off reasoning or context, and does not address the root cause of latency from infrastructure contention.

Full explanation →

258

MCQmedium

A data scientist is building a binary classifier to predict customer churn. The dataset is highly imbalanced (95% non-churn, 5% churn). Which metric should be prioritized for model evaluation?

A.F1 score

B.Root Mean Squared Error (RMSE)

C.Mean Absolute Error (MAE)

D.Accuracy

AnswerA

F1 score combines precision and recall, providing a balanced view of model performance on the minority class.

Why this answer

In imbalanced datasets, accuracy can be misleading because a model that always predicts the majority class achieves 95% accuracy but fails to identify churners. The F1 score balances precision and recall, making it more informative for the minority class.

Full explanation →

259

MCQeasy

A company uses AWS Lake Formation to manage data lakes for analytics. They want to ensure that only authorized users can access specific columns in a table containing sensitive data used for ML training. Which Lake Formation feature should they use?

A.Row-level security filters

B.Cell-level security

C.S3 bucket policies

D.Column-level permissions

AnswerD

Lake Formation supports granting permissions on specific columns.

Why this answer

Lake Formation column-level permissions allow fine-grained access control to specific columns within a table.

Full explanation →

260

Multi-Selectmedium

A company uses an LLM to generate medical advice. They are concerned about hallucinations and want to implement safeguards. Which TWO techniques should they prioritize?

Select 2 answers

A.Reduce the temperature parameter to 0

B.Fine-tune the LLM on a general dataset of medical texts

C.Implement Retrieval-Augmented Generation (RAG) using a verified medical knowledge base

D.Set up a human-in-the-loop workflow with Amazon A2I to review high-risk responses

E.Use Bedrock Guardrails with a content filter for medical topics

AnswersC, D

RAG retrieves factual information to ground the LLM's responses, reducing hallucinations.

Why this answer

RAG grounds the model in verified medical documents, and human-in-the-loop (A2I) provides oversight for high-risk outputs. Fine-tuning on general data is not sufficient to prevent hallucinations on specific medical topics. Temperature adjustment is a weak control.

Stricter content filtering may help but is not specifically designed to address factual accuracy.

Full explanation →

261

MCQeasy

A developer is using Amazon Bedrock to build a chatbot that answers customer queries. The chatbot must only respond based on the provided company documentation. Which approach best meets this requirement?

A.Use prompt engineering to instruct the model to only use documentation.

B.Use a RAG architecture with the company documentation as the knowledge base.

C.Fine-tune a foundation model on the company documentation.

D.Use a text classification model to filter responses.

AnswerB

RAG ensures responses are based on retrieved documents.

Why this answer

Retrieval-Augmented Generation (RAG) architecture retrieves relevant chunks from the company documentation at query time and injects them into the prompt, ensuring the model's response is grounded solely in the provided documents. This approach prevents the model from relying on its internal training data or generating information outside the documentation, which is critical for a closed-domain chatbot.

Exam trap

The AIF-C01 exam often tests the distinction between prompt engineering and RAG, where candidates mistakenly believe a well-crafted prompt can fully control model behavior without a retrieval mechanism, overlooking the fact that foundation models inherently generate responses from their training data unless explicitly grounded via external knowledge retrieval.

How to eliminate wrong answers

Option A is wrong because prompt engineering alone cannot guarantee the model will ignore its pre-trained knowledge; the model may still hallucinate or use information not present in the documentation, as it has no mechanism to enforce retrieval of specific content. Option C is wrong because fine-tuning a foundation model on the company documentation embeds the data into the model's weights, which can lead to outdated or incomplete responses and does not allow dynamic retrieval of the latest documentation; it also risks overfitting and does not scale well with changing content. Option D is wrong because a text classification model filters responses after generation, but it cannot ensure the response is based on the documentation; it only labels or rejects outputs, which is insufficient for generating accurate, document-grounded answers.

Full explanation →

262

MCQhard

A company is developing an AI system to screen job applications. They want to ensure the system does not discriminate against candidates based on gender. The dataset used contains historical hiring decisions that may reflect past biases. Which type of bias is MOST likely present in this scenario?

A.Historical bias

B.Measurement bias

C.Representation bias

D.Aggregation bias

AnswerA

Historical bias arises from past inequalities reflected in the training data.

Why this answer

Historical bias occurs when the data reflects past societal biases, such as discriminatory hiring practices. This is the most likely bias given the context.

Full explanation →

263

MCQhard

A company is building a chatbot that must provide accurate answers based on internal documents without retraining the model. Which approach should they use?

A.Reinforcement learning from human feedback (RLHF)

B.Fine-tuning the model on internal documents

C.Model distillation to a smaller model

D.Prompt engineering with retrieval-augmented generation (RAG)

AnswerD

RAG retrieves relevant documents at inference time, providing up-to-date answers.

Why this answer

Retrieval-augmented generation (RAG) allows the chatbot to fetch relevant internal documents at inference time and incorporate them into the prompt, providing accurate, up-to-date answers without retraining the model. This approach combines prompt engineering with a retrieval step, ensuring the model's responses are grounded in the company's specific knowledge base while keeping the base model frozen.

Exam trap

The trap here is that candidates may confuse fine-tuning (which requires retraining) with RAG (which does not), or mistakenly think RLHF or distillation can inject new factual knowledge without retraining, when in fact they address alignment, efficiency, or behavior, not dynamic knowledge retrieval.

How to eliminate wrong answers

Option A is wrong because reinforcement learning from human feedback (RLHF) is a training technique used to align model behavior with human preferences, not a method for injecting new factual knowledge without retraining. Option B is wrong because fine-tuning the model on internal documents would require modifying the model's weights through additional training, which contradicts the requirement of not retraining the model. Option C is wrong because model distillation compresses a large model into a smaller one for efficiency, but it does not enable the model to answer questions based on new internal documents without retraining.

Full explanation →

264

MCQhard

A research lab uses Amazon SageMaker to train a deep learning model for medical diagnosis. They need to ensure the model's decisions are interpretable to clinicians. Which SageMaker feature provides local and global feature importance?

A.SageMaker Model Monitor

B.SageMaker Experiments

C.SageMaker Clarify

D.SageMaker Debugger

AnswerC

Clarify provides explainability metrics.

Why this answer

SageMaker Clarify is the correct answer because it is specifically designed to provide both local and global feature importance for machine learning models. Local feature importance explains individual predictions (e.g., why a specific patient was diagnosed), while global feature importance shows which features most influence the model overall. This directly supports interpretability for clinicians, as required in the question.

Exam trap

The trap here is that candidates confuse SageMaker Debugger's ability to monitor training metrics with model interpretability, but Debugger does not compute feature importance or explain predictions.

How to eliminate wrong answers

Option A is wrong because SageMaker Model Monitor is used for detecting data drift, bias drift, and model quality degradation over time, not for computing feature importance. Option B is wrong because SageMaker Experiments is a tool for tracking, organizing, and comparing machine learning training runs, not for model interpretability or feature importance. Option D is wrong because SageMaker Debugger is designed to monitor training jobs for issues like vanishing gradients or overfitting by capturing tensors and metrics, but it does not provide local or global feature importance.

Full explanation →

265

MCQmedium

A data engineer needs to ensure that all data uploaded to an S3 bucket for SageMaker training is automatically encrypted with a customer-managed key. Which S3 feature should they enable?

A.Object lock with compliance mode.

B.Default encryption with SSE-KMS using an AWS managed key.

C.Default encryption with SSE-KMS using a customer managed key (CMK).

D.Default encryption with SSE-S3.

AnswerC

CMK provides customer control over encryption keys.

Why this answer

The requirement specifies a customer-managed key (CMK). Default encryption with SSE-KMS allows you to specify a CMK, ensuring all objects uploaded to the S3 bucket are automatically encrypted using that key. This satisfies the data engineer's need for control over the encryption key.

Exam trap

The AIF-C01 exam often tests the distinction between AWS managed keys and customer managed keys (CMKs) in SSE-KMS, where candidates mistakenly select an AWS managed key option when the question explicitly requires a customer-managed key.

How to eliminate wrong answers

Option A is wrong because Object Lock with compliance mode is designed to prevent object deletion or overwrites for a fixed retention period, not to enforce encryption. Option B is wrong because it uses an AWS managed key, not a customer-managed key as required. Option D is wrong because SSE-S3 uses Amazon S3-managed keys, which do not provide the customer control required by the scenario.

Full explanation →

266

MCQmedium

A developer is building an Amazon Bedrock Agent to automate multi‑step workflows like booking flights and hotels. The Agent needs to call an external flight reservation API. How should the developer expose this API to the Agent?

A.Create an action group that includes an OpenAPI schema describing the flight API and a Lambda function to invoke it

B.Add the API documentation as a context document in the Agent's instructions

C.Configure the Agent to use Amazon Kendra to index the API documentation

D.Hard‑code the API endpoint in the Agent's prompt using a system message

AnswerA

Action groups define the tools available to an Agent. The Agent uses the OpenAPI schema to understand the API and triggers the Lambda to execute the call.

Why this answer

Amazon Bedrock Agents use action groups to define external APIs that the agent can invoke. An action group requires an OpenAPI schema (to describe the API operations and parameters) and a Lambda function (to execute the actual API call). This allows the agent to dynamically select and call the flight reservation API during multi-step workflows.

Exam trap

The trap here is that candidates may think providing API documentation as context (Option B) is sufficient, but they miss that Bedrock Agents require an action group with both a schema and a Lambda function to actually invoke the API.

How to eliminate wrong answers

Option B is wrong because adding API documentation as a context document in the Agent's instructions does not enable the agent to actually call the API; it only provides static text that the agent cannot execute. Option C is wrong because Amazon Kendra is a search service for indexing documents, not a mechanism for exposing APIs to an agent; it cannot handle API invocation. Option D is wrong because hard-coding the API endpoint in the agent's prompt using a system message does not provide the structured schema or invocation logic needed for the agent to parse parameters and make the call; it is a fragile, non-scalable approach.

Full explanation →

267

MCQmedium

A company uses Amazon Bedrock to generate marketing copy. The summaries are too verbose. Which parameter should be decreased to directly limit the length of the output?

A.max_tokens

B.temperature

C.top_p

D.frequency_penalty

AnswerA

max_tokens sets the maximum number of tokens in the response.

Why this answer

The `max_tokens` parameter directly controls the maximum number of tokens (words or subwords) in the generated output. By decreasing this value, you explicitly cap the length of the marketing copy, making it less verbose. This is the most direct way to limit output length in Amazon Bedrock and other LLM APIs.

Exam trap

The trap here is that candidates confuse parameters that affect output style (temperature, top_p, frequency_penalty) with the one that directly controls output length (max_tokens), leading them to pick a parameter that changes how the model writes rather than how much it writes.

How to eliminate wrong answers

Option B (temperature) is wrong because it controls the randomness of token selection, not the length of the output; lowering temperature makes output more deterministic but does not shorten it. Option C (top_p) is wrong because it sets a cumulative probability threshold for nucleus sampling, affecting diversity of token choices, not the total number of tokens generated. Option D (frequency_penalty) is wrong because it penalizes tokens based on their frequency in the generated text, reducing repetition but not directly limiting the overall length of the response.

Full explanation →

268

Multi-Selectmedium

An AI system is used to screen job applications. The team finds that the model has a higher false positive rate for male applicants than female applicants. Which fairness metrics should they compute to quantify this disparity? (Choose two.)

Select 2 answers

A.Equalized odds

B.Individual fairness

C.Predictive parity

D.Disparate impact (80% rule)

E.Demographic parity

AnswersA, D

Equalized odds requires that both true positive rate and false positive rate are equal across groups, directly addressing the false positive disparity.

Why this answer

Equalized odds requires equal true positive and false positive rates across groups. Disparate impact compares selection rates. Demographic parity does not consider error rates; predictive parity compares positive predictive values.

Full explanation →

269

MCQmedium

A healthcare company is using Amazon Bedrock to summarize patient notes. The compliance team requires that no patient data is used to improve the underlying foundation model. Which configuration should the team choose?

A.Enable data encryption in transit and at rest.

B.Use a different foundation model from a different provider.

C.Disable model training data logging in the AWS console.

D.Configure a VPC endpoint for Amazon Bedrock.

AnswerC

This setting prevents prompts and completions from being used for model improvement.

Why this answer

Disabling model training data logging in the AWS console prevents Amazon Bedrock from using customer inference data to improve the underlying foundation model. This setting ensures compliance with the requirement that no patient data is used for model training, as Bedrock offers a specific toggle to opt out of data sharing for model improvement.

Exam trap

AWS often tests the misconception that encryption or network controls (like VPC endpoints) are sufficient for data privacy compliance, when the actual requirement is about preventing data usage for model improvement, which is a separate policy control.

How to eliminate wrong answers

Option A is wrong because enabling data encryption in transit and at rest protects data confidentiality but does not prevent the foundation model provider from using the data for training or improvement. Option B is wrong because using a different foundation model from a different provider does not inherently guarantee that patient data will not be used for model improvement; the compliance requirement is about data usage policy, not model origin. Option D is wrong because configuring a VPC endpoint for Amazon Bedrock controls network access and data exfiltration but does not affect whether inference data is logged or used for model training.

Full explanation →

270

MCQeasy

A team is developing an AI system and wants to document key information such as intended use, performance benchmarks, and limitations. According to AWS best practices for responsible AI, what should they create?

A.A whitepaper

B.A business requirement document

C.A technical blog

D.Model cards

AnswerD

Model cards provide a structured summary of model characteristics, intended use, fairness, and limitations.

Why this answer

Model cards are a structured documentation framework recommended by AWS for responsible AI. They provide a standardized way to communicate key information such as intended use, performance benchmarks, limitations, and ethical considerations, ensuring transparency and accountability.

Exam trap

The trap here is that candidates may confuse a general-purpose document like a whitepaper or blog with the specific, structured artifact (model card) that AWS mandates for responsible AI documentation, overlooking the need for standardized transparency fields.

How to eliminate wrong answers

Option A is wrong because a whitepaper is a lengthy, narrative document often used for marketing or high-level overviews, not the standardized, concise format AWS recommends for responsible AI documentation. Option B is wrong because a business requirement document (BRD) focuses on business needs and functional requirements, not on technical performance, limitations, or ethical AI details. Option C is wrong because a technical blog is an informal, narrative publication for sharing insights or tutorials, lacking the structured, mandatory fields required for responsible AI transparency.

Full explanation →

271

MCQhard

An administrator reviews a CloudTrail log entry for a CreateModel API call. Which security concern should they investigate?

A.The model name is not encrypted in the log

B.The model data URL uses HTTP instead of HTTPS

C.The source IP address is an external IP

D.The execution role ARN is visible in the log

AnswerB

HTTPS should be used to encrypt data in transit; using HTTP is a security risk.

Why this answer

Using HTTP instead of HTTPS for the model data URL exposes the data in transit to potential interception or tampering. CloudTrail logs record the URL as provided, and if it uses HTTP, the data transferred from that URL to SageMaker is unencrypted, violating security best practices for data in transit. This is a direct security concern that should be investigated and remediated by using HTTPS.

Exam trap

The AIF-C01 exam often tests the distinction between data in transit encryption (HTTPS vs HTTP) and data at rest encryption, leading candidates to incorrectly focus on log encryption or ARN visibility instead of the actual security risk in the API call parameters.

How to eliminate wrong answers

Option A is wrong because model names are not sensitive data that require encryption in CloudTrail logs; CloudTrail logs are encrypted at rest by default using AWS KMS, but the content of the log entry (including the model name) is not individually encrypted. Option C is wrong because an external source IP address is not inherently a security concern; CloudTrail logs all API calls regardless of source, and external IPs are expected for calls made from outside AWS. Option D is wrong because the execution role ARN is a standard part of the CreateModel API call and is not sensitive; it is necessary for auditing who performed the action and does not expose credentials.

Full explanation →

272

Multi-Selecthard

A company is developing an AI recruitment tool that screens candidates. They want to minimize bias and ensure compliance with emerging regulations like the EU AI Act. Which THREE measures should they implement? (Select THREE.)

Select 3 answers

A.Disclose to candidates that AI is used in the screening process

B.Regularly audit the model for bias using fairness metrics

C.Remove all human involvement from the decision process

D.Create model cards documenting intended use, performance, and limitations

E.Use a single demographic group for training data

AnswersA, B, D

Transparency is required under the EU AI Act; candidates have the right to know they are being evaluated by AI.

Why this answer

Conducting bias audits, documenting model cards, and ensuring transparency when AI is used are key practices for fairness, governance, and compliance with regulations like the EU AI Act which classifies recruitment as high-risk.

Full explanation →

273

MCQhard

A healthcare organization is using Amazon Bedrock to analyze medical images and generate radiology reports. They need to comply with HIPAA regulations and ensure patient data is not used for model training. Which configuration should they use?

A.Fine-tune the model using a custom dataset and deploy as a custom model

B.Use a provisioned throughput model with data isolation

C.Use the on-demand model through Amazon Bedrock

D.Use a third-party model hosted outside of AWS

AnswerB

Provisioned throughput ensures data is not used for training and meets compliance requirements.

Why this answer

Provisioned Throughput with data isolation in Amazon Bedrock ensures that the customer's inference data (including patient medical images and reports) is not used for any model training or service improvement, and it provides a dedicated, isolated environment that meets HIPAA compliance requirements. This configuration guarantees that patient data remains within the customer's AWS account and is not shared with other customers or used to improve the base model.

Exam trap

The trap here is that candidates often assume fine-tuning (Option A) is the only way to customize models for healthcare, but they overlook that HIPAA prohibits using PHI for training, making Provisioned Throughput with data isolation the correct choice for compliant inference.

How to eliminate wrong answers

Option A is wrong because fine-tuning a model with a custom dataset would use patient data to train the model, which violates HIPAA requirements that patient data must not be used for model training. Option C is wrong because the on-demand model through Amazon Bedrock does not provide data isolation; inference data may be used for service improvement and model training, which is not HIPAA-compliant for protected health information. Option D is wrong because using a third-party model hosted outside of AWS would require the healthcare organization to manage HIPAA compliance independently, and it does not leverage AWS's HIPAA-eligible services or the data isolation guarantees provided by Bedrock's Provisioned Throughput.

Full explanation →

274

MCQhard

A healthcare startup is using an LLM to summarize patient medical records. They are concerned about hallucinations where the model may invent symptoms or treatments. Which combination of techniques should they implement to reduce hallucinations while maintaining accuracy?

A.Apply prompt engineering with few-shot examples and set temperature to 0

B.Use a larger LLM and increase the temperature parameter to improve creativity

C.Implement RAG with a trusted medical knowledge base and use Bedrock Guardrails with contextual grounding check

D.Fine-tune the LLM on a large corpus of medical literature and use a higher temperature for diversity

AnswerC

RAG provides factual grounding by retrieving relevant documents, and Bedrock Guardrails' contextual grounding check validates response accuracy against the retrieved sources, significantly reducing hallucinations.

Why this answer

Grounding the model with Retrieval-Augmented Generation (RAG) using verified medical databases and applying Bedrock Guardrails with contextual grounding checks directly reduces hallucinations by ensuring responses are based on retrieved facts and filtered for accuracy.

Full explanation →

275

MCQhard

A company uses Amazon SageMaker to train a model on customer transaction data. The data is stored in an S3 bucket encrypted with SSE-KMS. The data scientists are using a SageMaker notebook instance to preprocess the data. The security team notices that the notebook instance can access the KMS key directly through the SageMaker console, even when the data scientists do not have explicit KMS permissions. What is the MOST likely reason for this behavior?

A.The KMS key policy allows the SageMaker service to decrypt using the key, and the console displays this permission globally.

B.The S3 bucket has a bucket policy that grants s3:GetObject access using the SageMaker service principal, bypassing KMS permissions.

C.The SageMaker execution role attached to the notebook instance has permission to use the KMS key for the S3 bucket, and the SageMaker console uses that role to perform decryption operations on behalf of the user.

D.The SageMaker notebook instance is using an IAM role that has the AWS managed policy AmazonSageMakerFullAccess, which includes kms:Decrypt for all keys.

AnswerC

The SageMaker execution role is assumed by the notebook instance and has KMS permissions. The console uses the role to show decryption status.

Why this answer

The SageMaker notebook instance uses an execution IAM role that is assumed when performing operations. If that execution role has permissions to use the KMS key (e.g., kms:Decrypt, kms:GenerateDataKey) for the S3 bucket, the SageMaker console can leverage that role to decrypt data on behalf of the user, even if the user's own IAM identity lacks explicit KMS permissions. The console acts as a proxy, using the notebook's execution role to access the key.

Exam trap

The trap here is that candidates often assume the user's own IAM permissions are the only factor, but AWS services like SageMaker can proxy requests using the execution role attached to the resource, effectively decoupling the user's identity from the data access permissions.

How to eliminate wrong answers

Option A is wrong because the KMS key policy granting the SageMaker service principal decryption permissions would allow the service to use the key, but the console does not display permissions globally; it reflects the effective permissions of the assumed role. Option B is wrong because an S3 bucket policy granting s3:GetObject to the SageMaker service principal does not bypass KMS permissions; accessing SSE-KMS encrypted objects still requires kms:Decrypt or kms:GenerateDataKey permissions on the KMS key. Option D is wrong because the AWS managed policy AmazonSageMakerFullAccess does not include kms:Decrypt for all keys; it only includes limited KMS actions (e.g., kms:ListKeys) and does not grant decrypt on arbitrary customer-managed KMS keys.

Full explanation →

276

MCQmedium

A company is using Amazon Comprehend to extract entities from customer support tickets. The compliance team requires that the text sent to Comprehend be encrypted in transit and that Comprehend does not store any data beyond the processing time. How should the company configure the API call?

A.Encrypt the text using AWS KMS before sending it to Comprehend

B.Use the AWS SDK with server-side encryption enabled for the API call

C.Create a VPC endpoint for Comprehend and send requests over the private network

D.Use the HTTPS endpoint for the DetectEntities API and rely on Comprehend's stateless design

AnswerD

HTTPS provides encryption in transit; Comprehend does not store data after processing.

Why this answer

Amazon Comprehend's DetectEntities API is inherently stateless—it does not store any text or results beyond the processing time required to generate the response. By using the HTTPS endpoint, the company ensures encryption in transit via TLS, satisfying the compliance requirement without additional configuration. Comprehend does not persist data from synchronous API calls, so no data is retained after the response is returned.

Exam trap

The trap here is that candidates confuse encryption in transit with encryption at rest or data retention controls, leading them to overcomplicate the solution with KMS, SSE, or VPC endpoints when the simple use of HTTPS and the stateless nature of the API already meet the requirements.

How to eliminate wrong answers

Option A is wrong because encrypting the text with AWS KMS before sending it would render the data unreadable to Comprehend, as the service requires plaintext input to perform entity extraction; KMS encryption is for data at rest, not for protecting data in transit. Option B is wrong because server-side encryption (SSE) is a feature for data at rest in services like S3, not for API calls to Comprehend; the AWS SDK does not have a 'server-side encryption' setting for API requests, and encryption in transit is handled by TLS, not SSE. Option C is wrong because a VPC endpoint for Comprehend enables private connectivity over the AWS network but does not inherently encrypt data in transit; encryption still requires TLS, and the VPC endpoint does not affect Comprehend's data retention behavior.

Full explanation →

277

MCQeasy

A startup is building a recommendation engine for their e-commerce platform. They need a fully managed service that can generate personalized product recommendations based on user behavior. Which AWS service should they use?

A.Amazon Personalize

B.Amazon Rekognition

C.Amazon Forecast

D.Amazon Comprehend

AnswerA

Personalize is designed specifically for personalization and recommendations.

Why this answer

Amazon Personalize is a fully managed machine learning service specifically designed to generate real-time personalized product recommendations by processing user behavior data (e.g., clicks, purchases, views) and item metadata. It uses the same technology that powers Amazon.com's recommendation engine, making it the correct choice for this e-commerce use case.

Exam trap

The AIF-C01 exam often tests the distinction between AWS AI services by presenting a use case that sounds like 'forecasting' or 'analysis' but actually requires personalization, leading candidates to confuse Amazon Forecast (time-series) with Amazon Personalize (recommendations).

How to eliminate wrong answers

Option B (Amazon Rekognition) is wrong because it is a computer vision service for image and video analysis (e.g., object detection, facial recognition), not for generating product recommendations. Option C (Amazon Forecast) is wrong because it is a time-series forecasting service for predicting future metrics (e.g., demand, sales), not for personalized recommendations based on user behavior. Option D (Amazon Comprehend) is wrong because it is a natural language processing (NLP) service for extracting insights from text (e.g., sentiment, entities), not for recommendation generation.

Full explanation →

278

Multi-Selecteasy

A company is using Amazon Bedrock to generate code snippets. They want to ensure the generated code is secure. Which TWO practices should they implement?

Select 2 answers

A.Increase the max token limit to generate longer code.

B.Use guardrails to block insecure code patterns.

C.Set the temperature to 0 for deterministic output.

D.Review and test all generated code before deployment.

E.Use a larger model for better accuracy.

AnswersB, D

Guardrails can filter out harmful content.

Why this answer

Amazon Bedrock Guardrails allow you to define policies that filter or block generated content containing insecure code patterns, such as SQL injection or hardcoded credentials, before the output is returned. This provides a proactive security layer that prevents insecure code from reaching the user, directly addressing the requirement to ensure generated code is secure.

Exam trap

The AIF-C01 exam often tests the misconception that model parameters like temperature or token limits can substitute for explicit security controls, when in fact only guardrails and human review directly address code security.

Full explanation →

279

Multi-Selecteasy

Which TWO services can be used to preprocess data for machine learning in AWS? (Choose two.)

Select 2 answers

A.AWS Glue

B.Amazon Athena

C.Amazon SageMaker Data Wrangler

D.Amazon Redshift

E.AWS Lambda

AnswersA, C

Glue provides ETL capabilities suitable for preprocessing.

Why this answer

AWS Glue is a fully managed ETL service that can be used to preprocess data for machine learning by cleaning, transforming, and enriching raw data before feeding it into ML models. It provides built-in transforms and can handle both structured and semi-structured data, making it suitable for preparing large datasets for training.

Exam trap

The AIF-C01 exam often tests the distinction between data querying services (like Athena) and data preprocessing services, leading candidates to mistakenly choose Athena because it can 'process' data via SQL, but it lacks the ML-specific transformation capabilities required for preprocessing.

Full explanation →

280

MCQmedium

A financial services company is deploying a machine learning model to approve loan applications. To comply with regulatory requirements, they must ensure the model does not discriminate based on race. They have historical data that may contain bias. Which AWS service can help detect and measure bias in the dataset and model predictions?

A.AWS CloudTrail

B.Amazon SageMaker Model Monitor

C.AWS Identity and Access Management (IAM)

D.Amazon SageMaker Clarify

AnswerD

SageMaker Clarify provides bias detection and explainability for datasets and models.

Why this answer

Amazon SageMaker Clarify provides tools for bias detection and explainability, including pre-training and post-training bias metrics.

Full explanation →

281

MCQmedium

A data scientist is using Amazon SageMaker Clarify to generate a bias report for a binary classification model. They want to understand which features most influence the model's predictions. Which feature of Clarify should they use?

A.Partial dependence plots

B.Bias metrics report

C.Feature importance from SHAP values

D.Data quality report

AnswerC

Clarify computes SHAP values to show the contribution of each feature to predictions, enabling interpretability.

Why this answer

SageMaker Clarify provides SHAP-based feature importance values to explain model predictions. It can compute both local and global feature importance.

Full explanation →

282

MCQhard

A data scientist is fine-tuning a large language model on Amazon SageMaker for a text summarization task. The training loss decreases steadily but the validation loss starts increasing after a few epochs. What should the scientist do to address this issue?

A.Reduce the batch size

B.Increase the learning rate

C.Increase the number of training epochs

D.Use early stopping based on validation loss

AnswerD

Early stopping prevents overfitting by halting training when validation loss stops improving.

Why this answer

The validation loss increasing while training loss decreases is a classic sign of overfitting. Early stopping based on validation loss halts training when the validation loss stops improving, preventing overfitting and saving computational resources. This is a standard technique in SageMaker's built-in training algorithms and custom training scripts.

Exam trap

The AIF-C01 exam often tests the distinction between overfitting and underfitting; the trap here is that candidates may mistakenly think increasing epochs (Option C) always improves performance, ignoring the validation loss divergence that signals overfitting.

How to eliminate wrong answers

Option A is wrong because reducing batch size introduces more noise into gradient estimates, which can actually worsen generalization and does not directly address overfitting. Option B is wrong because increasing the learning rate can cause the optimizer to overshoot minima, leading to divergence or unstable training, not reduced overfitting. Option C is wrong because increasing the number of training epochs would exacerbate overfitting, as the model would continue to memorize the training data beyond the point where validation loss degrades.

Full explanation →

283

Multi-Selectmedium

A company wants to use Amazon Bedrock to build a multilingual customer support chatbot. The chatbot must answer questions in English, Spanish, and French. Which TWO actions should the company take to achieve this? (Select TWO.)

Select 2 answers

A.Use system prompts to specify the desired output language based on user input

B.Fine-tune a separate model for each language

C.Translate all user queries to English before sending to the model

D.Use a different embedding model for each language

E.Select a foundation model that supports multilingual input and output, such as Anthropic Claude

AnswersA, E

System prompts can instruct the model to respond in a specific language, leveraging its built-in multilingual ability.

Why this answer

Many foundation models on Bedrock support multiple languages natively, so choosing a multilingual model is key. Prompting in the user's language is essential for generating responses in that language.

Full explanation →

284

MCQmedium

A company is using Amazon Rekognition to detect objects in images. They need to detect custom objects that are specific to their domain. What should they do?

A.Use Amazon Rekognition's built-in labels

B.Use Amazon SageMaker Object Detection algorithm

C.Use Amazon Rekognition Custom Labels

D.Use Amazon Comprehend

AnswerC

Custom Labels enables training a custom model with labeled images.

Why this answer

Amazon Rekognition Custom Labels allows you to train a custom model using your own labeled images to detect domain-specific objects that are not covered by Rekognition's built-in labels. This is the correct service for custom object detection without needing to build a model from scratch.

Exam trap

The trap here is that candidates may confuse Amazon Rekognition Custom Labels with SageMaker Object Detection, not realizing that Custom Labels is a managed service specifically designed for custom image analysis without requiring ML expertise.

How to eliminate wrong answers

Option A is wrong because built-in labels are pre-trained on general categories and cannot detect custom domain-specific objects. Option B is wrong because Amazon SageMaker Object Detection algorithm requires you to build, train, and deploy a custom model from scratch, which is more complex and not the recommended approach when Rekognition Custom Labels can handle the task with less effort. Option D is wrong because Amazon Comprehend is a natural language processing (NLP) service for text analysis, not for image object detection.

Full explanation →

285

MCQmedium

A company wants to use a third-party foundation model in Bedrock and is concerned about the provider's data handling policies. Which action should they take to ensure their data is not used for model training by the provider?

A.Use Bedrock's content filters to block sensitive data

B.Encrypt all data with KMS before sending to Bedrock

C.Use a custom model hosted on SageMaker instead

D.Review the provider's data privacy policy and enable the opt-out option for training data use if available

AnswerD

Reviewing the policy and opting out ensures data is not used for training.

Why this answer

AWS provides model access approval settings where customers can review and accept data handling policies; many providers offer opt-out options for training data use.

Full explanation →

286

MCQeasy

A developer invokes an Amazon Bedrock model and receives the above response. What does the 'stopReason' field indicate?

A.The model encountered an error.

B.The model reached a defined stop sequence.

C.The model hit the maximum token limit.

D.The model stopped due to a safety filter.

AnswerB

'stop_sequence' indicates the model encountered a user-defined stop sequence.

Why this answer

The 'stopReason' field in an Amazon Bedrock response indicates why the model stopped generating tokens. When set to 'stop', it means the model encountered a defined stop sequence (such as a special token like <|endoftext|> or a user-specified string) and halted generation normally. This is the expected behavior for a successful, complete response.

Exam trap

The trap here is that candidates confuse 'stop' (normal completion via stop sequence) with 'length' (token limit reached), as both end generation but have different implications for response completeness and cost.

How to eliminate wrong answers

Option A is wrong because a model error would typically result in an HTTP error code or a different field like 'error' or 'failure', not a 'stopReason' of 'stop'. Option C is wrong because hitting the maximum token limit would produce a 'stopReason' of 'length', not 'stop'. Option D is wrong because a safety filter intervention would produce a 'stopReason' of 'content_filtered' or similar, not 'stop'.

Full explanation →

287

Multi-Selecthard

A data scientist is using Amazon SageMaker Clarify to generate a model explainability report. They want to include both global and local feature importance. Which TWO techniques does SageMaker Clarify support for these purposes?

Select 2 answers

A.SHAP values

B.Partial dependence plots

C.Permutation feature importance

D.PDP-based interaction indices

E.LIME

AnswersA, C

SHAP provides local feature importance for individual predictions.

Why this answer

SageMaker Clarify supports SHAP for local explanations and feature importance (e.g., permutation-based) for global explanations.

Full explanation →

288

MCQhard

A developer attached this IAM policy to a role used by an application that invokes Claude v2 in us-east-1. The application receives an access denied error. What is the MOST likely cause?

A.The Allow statement does not include a condition on the region

B.The Deny statement is blocking requests because the condition does not match the resource ARN's region

C.The Deny statement uses StringNotEquals instead of StringEquals

D.The resource ARN in the Allow statement is incorrect

AnswerB

The Deny condition checks aws:RequestedRegion, which may differ from the region in the resource ARN if requests are made to a different region.

Why this answer

The Deny statement uses a `StringNotEquals` condition on `aws:RequestedRegion` set to `us-east-1`. This means the Deny applies to any request where the requested region is NOT `us-east-1`. Since the resource ARN in the Deny statement is `arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-v2`, the condition does not match the resource's region (the resource ARN itself is in us-east-1), but the Deny is triggered when the request is made to a different region, blocking the call.

The application is likely invoking the model from a region other than us-east-1, causing the Deny to take effect.

Exam trap

The AIF-C01 exam often tests the subtle interaction between Allow and Deny statements with condition operators, where candidates mistakenly think the Deny is blocking because of a region mismatch on the resource ARN itself, rather than understanding that the Deny's condition evaluates the request's region, not the resource's region.

How to eliminate wrong answers

Option A is wrong because the Allow statement does not need a region condition; the Allow grants access to the specific resource ARN, and the Deny is the one causing the issue. Option C is wrong because using `StringNotEquals` is correct for this pattern—it denies requests that are NOT in the specified region; `StringEquals` would deny only requests in us-east-1, which is not the intended behavior. Option D is wrong because the resource ARN in the Allow statement (`arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-v2`) is correct for Claude v2 in us-east-1; the error is not due to an incorrect ARN but due to the Deny statement's condition logic.

Full explanation →

289

MCQmedium

A company is deploying an AI-powered document summarization system using Amazon Bedrock. They must ensure that the model only uses information from provided source documents and does not generate unsupported claims. Which Bedrock Guardrails feature should they enable?

A.Content filtering with category-based harmful content filters

B.Contextual grounding check

C.Word filters

D.Topic restrictions

AnswerB

Grounding checks validate that the model's output is based on the provided source documents.

Why this answer

Contextual grounding checks ensure that model responses are factually supported by the reference sources, reducing hallucinations.

Full explanation →

290

MCQeasy

A developer receives the above response from invoking a Bedrock model. Which field indicates that the model completed its response normally?

A.output

B.stop_reason

C.text

D.role

AnswerB

stop_reason 'end_turn' signals normal conversation end.

Why this answer

The `stop_reason` field in the Bedrock response indicates why the model stopped generating text. A value of `"stop"` or `"end_turn"` (depending on the model) signals that the model completed its response normally, as opposed to hitting a token limit, content filter, or other interruption.

Exam trap

The trap here is that candidates confuse the `output` container or the `text` field with the completion indicator, overlooking the dedicated `stop_reason` field that explicitly signals normal termination.

How to eliminate wrong answers

Option A is wrong because `output` is a container object that holds the generated content, not a field that indicates the completion status. Option C is wrong because `text` is a field within the output that contains the actual generated string, but it does not convey why generation stopped. Option D is wrong because `role` indicates the conversational role (e.g., user or assistant) in a multi-turn context, not the model's completion state.

Full explanation →

291

MCQeasy

A company uses Amazon Bedrock to build a conversational AI. They want to enforce role-based access to the model. Which AWS service should they use?

A.AWS Config

B.AWS Identity and Access Management (IAM)

C.AWS CloudTrail

D.AWS Organizations

AnswerB

IAM policies can control which users or roles can invoke specific Bedrock models.

Why this answer

AWS Identity and Access Management (IAM) is the correct service because it enables fine-grained, role-based access control (RBAC) to Amazon Bedrock models. You can define IAM policies that specify which principals (users, groups, or roles) are allowed to invoke specific foundation models, ensuring that only authorized roles can interact with the conversational AI.

Exam trap

The trap here is that candidates often confuse AWS Config (which audits configurations) or CloudTrail (which logs actions) with IAM, mistakenly thinking that logging or compliance tools can enforce access control, when in fact only IAM provides the authorization layer for Bedrock model invocation.

How to eliminate wrong answers

Option A is wrong because AWS Config is a service for evaluating and auditing resource configurations against compliance rules, not for enforcing role-based access to Bedrock models. Option C is wrong because AWS CloudTrail records API activity for auditing and governance, but it does not control or enforce access permissions. Option D is wrong because AWS Organizations manages multi-account governance and policy inheritance across accounts, but it does not provide the granular, per-model role-based access control needed for Bedrock.

Full explanation →

292

MCQeasy

A company needs to store large amounts of unstructured training data (images, videos) in a cost-effective manner while ensuring low-latency retrieval for training jobs running on Amazon SageMaker. Which storage solution should be used?

A.Amazon EFS

B.Amazon S3

C.Amazon RDS

D.Amazon EBS

AnswerB

S3 is the best fit for storing unstructured data with low-latency access via S3 endpoints.

Why this answer

Amazon S3 is the correct choice because it is designed for cost-effective, scalable storage of unstructured data (images, videos) and integrates natively with Amazon SageMaker for low-latency data retrieval during training jobs. S3 provides high throughput and can be accessed directly from SageMaker training instances without the need for file system mounting, making it ideal for large-scale ML workloads.

Exam trap

The trap here is that candidates often confuse the need for low-latency retrieval with the need for a mounted file system (EFS or EBS), not realizing that S3's direct integration with SageMaker provides both low latency and high throughput for training workloads without the cost and complexity of file storage.

How to eliminate wrong answers

Option A is wrong because Amazon EFS is a file system that provides shared access for EC2 instances but is not optimized for the high-throughput, cost-effective storage of large unstructured datasets like images and videos; it also incurs higher costs per GB compared to S3 and can introduce latency overhead when used with SageMaker. Option C is wrong because Amazon RDS is a relational database service designed for structured data with SQL queries, not for storing unstructured training data such as images and videos, and it would be prohibitively expensive and inefficient for large-scale blob storage. Option D is wrong because Amazon EBS provides block-level storage volumes attached to a single EC2 instance, which is not suitable for sharing large datasets across multiple SageMaker training jobs and lacks the cost efficiency and scalability of object storage for unstructured data.

Full explanation →

293

MCQhard

A media company uses a generative AI model to automatically create image captions for user-uploaded photos. During quality assurance, testers discover that the model sometimes generates captions that include stereotypes based on gender and race, even when the photos do not contain people. For example, a photo of a kitchen produces captions like 'woman cooking,' and a photo of a sports car generates 'man driving.' The company wants to launch the feature soon but recognizes the reputational risk. They have a limited budget and need to implement a solution that reduces harmful stereotypes without overly restricting the captions' creativity. The team has access to the model's training data, which is a large public dataset of image-caption pairs. Which approach should the team prioritize?

A.Replace the generative model with a simpler classification model that only describes objects

B.Use a different pre-trained generative model that is larger and more accurate

C.Filter the training data to remove or downweight pairs with stereotypes, then fine-tune the model

D.Add a post-processing filter that checks captions for known stereotype patterns and blocks them

AnswerC

Cleaning training data and fine-tuning directly reduces learned stereotypes.

Why this answer

To reduce stereotypical associations, the most effective approach is to filter and rebalance the training data by removing or downweighting caption pairs that contain stereotypes, and then fine-tune the model. This directly addresses the data source of bias. Adding a post-hoc filter may catch some stereotypes but could also remove valid captions.

Using a different model without data changes may not help. Limiting to object-only descriptions reduces creativity and may not meet product goals.

Full explanation →

294

Multi-Selecthard

Which THREE considerations are important when implementing responsible AI for a production NLP system? (Choose three.)

Select 3 answers

A.Obtain FDA approval for the model

B.Continuously monitor model outputs for bias and drift

C.Apply encryption at rest for all training code

D.Publish model cards detailing intended use, performance, and limitations

E.Include bias detection in the CI/CD pipeline for every model update

AnswersB, D, E

Production models require ongoing monitoring to ensure fairness over time.

Why this answer

Continuous monitoring of model outputs for bias and drift is essential for maintaining responsible AI in production NLP systems. As language patterns and data distributions evolve over time, a model that was initially fair and accurate can develop harmful biases or performance degradation, so automated monitoring ensures ongoing alignment with ethical standards and regulatory requirements.

Exam trap

AWS often tests the distinction between general security practices (like encryption) and the specific pillars of responsible AI (fairness, transparency, accountability, and robustness), leading candidates to confuse data protection with ethical AI governance.

Full explanation →

295

MCQeasy

A company wants to build a generative AI application that can automatically classify customer feedback into positive, neutral, or negative sentiments. Which foundation model capability is BEST suited for this task?

A.Text classification

B.Translation

C.Text generation

D.Text summarization

AnswerA

Classifying text into categories (positive, neutral, negative) is text classification, a standard NLP task for FMs.

Why this answer

Text classification is a common use case for foundation models, which can be prompted to output a label. Sentiment analysis is a specific type of text classification.

Full explanation →

296

MCQhard

A company uses Amazon Bedrock to generate content. They need to guarantee that the model does not produce toxic or biased outputs and that any violations are logged for review. Which service should they integrate with Bedrock?

A.Amazon Comprehend

B.AWS WAF

C.AWS CloudTrail

D.Amazon Bedrock Guardrails

AnswerD

Guardrails provides content filters and toxicity detection.

Why this answer

Amazon Bedrock Guardrails is the correct service because it is specifically designed to enforce content policies, filter toxic or biased outputs, and log violations for review within Bedrock workflows. It provides configurable thresholds for harmful categories (e.g., hate, insults, sexual content) and integrates directly with Bedrock models to intercept and redact or block undesired responses, with all violations recorded in Amazon CloudWatch Logs.

Exam trap

The trap here is that candidates may confuse AWS CloudTrail (which logs API calls) with content-level logging, or assume Amazon Comprehend can be used for real-time content filtering, when in fact Bedrock Guardrails is the only integrated service that both enforces content policies and logs violations directly within the Bedrock invocation flow.

How to eliminate wrong answers

Option A is wrong because Amazon Comprehend is a natural language processing (NLP) service for extracting insights (e.g., sentiment, entities) from text, but it does not provide real-time content filtering or enforcement of safety policies within Bedrock model outputs; it operates as a separate analysis tool, not an integrated guardrail. Option B is wrong because AWS WAF is a web application firewall that protects HTTP/HTTPS endpoints from common web exploits (e.g., SQL injection, cross-site scripting), not a service for monitoring or filtering AI-generated content for toxicity or bias. Option C is wrong because AWS CloudTrail records API activity for auditing (e.g., who called Bedrock APIs), but it does not inspect or log the content of model outputs for toxic or biased violations; it captures metadata, not the payload of responses.

Full explanation →

297

MCQeasy

After deploying a model, a company notices that the distribution of the input features has shifted compared to the training data. Which feature of Amazon SageMaker Model Monitor can alert them to this change?

A.Model quality monitoring

B.Bias drift monitoring

C.Feature importance drift

D.Data quality monitoring

AnswerD

Data quality monitors for drift in input features (baseline vs. live).

Why this answer

Amazon SageMaker Model Monitor's data quality monitoring feature is specifically designed to detect changes in the distribution of input features compared to the training data. It uses statistical tests (e.g., Kolmogorov-Smirnov, Chi-squared) to compare baseline and live data distributions, alerting when drift is detected. This directly addresses the scenario of input feature distribution shift.

Exam trap

The trap here is confusing 'data quality monitoring' (input feature drift) with 'model quality monitoring' (prediction performance metrics), as both involve 'quality' but address entirely different aspects of the ML pipeline.

How to eliminate wrong answers

Option A is wrong because model quality monitoring tracks metrics like accuracy or precision of predictions, not input feature distributions. Option B is wrong because bias drift monitoring focuses on changes in model bias (e.g., demographic parity) over time, not general feature distribution shifts. Option C is wrong because feature importance drift monitors changes in the relative importance of features to model predictions, not the distribution of the feature values themselves.

Full explanation →

298

MCQeasy

A company wants to use Amazon SageMaker to train a model on sensitive healthcare data. What is the MOST secure way to ensure that the training data is encrypted at rest and in transit, and that only authorized users can access the model artifacts?

A.Enable server-side encryption with AWS KMS managed keys for the S3 buckets, SageMaker notebook instances, and training jobs, and use IAM roles to limit access.

B.Disable encryption for faster training, and rely on IAM policies to restrict access.

C.Use client-side encryption for all data, and store encryption keys in AWS Secrets Manager.

D.Use an AWS CloudHSM to store encryption keys and configure SageMaker to use it.

AnswerA

SageMaker supports KMS encryption for all resources, and IAM roles enforce least privilege. This meets the security requirements.

Why this answer

It provides end-to-end encryption for sensitive healthcare data by enabling server-side encryption with AWS KMS managed keys for S3 buckets (data at rest), SageMaker notebook instances (data at rest and in transit via TLS), and training jobs (data at rest on ephemeral storage and in transit between services). IAM roles enforce least-privilege access to model artifacts, ensuring only authorized users can read or write them. This combination meets compliance requirements like HIPAA by encrypting data at rest (AES-256) and in transit (TLS 1.2+), while IAM policies control access.

Exam trap

AWS often tests the misconception that client-side encryption (Option C) or hardware security modules (Option D) are inherently more secure, but the exam expects you to know that SageMaker's native integration with AWS KMS provides the simplest and most secure end-to-end encryption for training jobs, data, and model artifacts without custom code or unsupported integrations.

How to eliminate wrong answers

Option B is wrong because disabling encryption violates security best practices and compliance mandates (e.g., HIPAA) for sensitive healthcare data, and IAM policies alone cannot protect data at rest or in transit from physical or network-level exposure. Option C is wrong because client-side encryption requires managing encryption keys outside AWS (even if stored in Secrets Manager), which adds complexity and risk of key exposure, and SageMaker does not natively support client-side encryption for training jobs or model artifacts without custom code. Option D is wrong because AWS CloudHSM provides hardware-based key storage but SageMaker does not directly integrate with CloudHSM for encryption of training data or model artifacts; SageMaker uses AWS KMS for encryption, and CloudHSM would require custom integration via KMS custom key store, which is not the most straightforward or recommended approach for this scenario.

Full explanation →

299

MCQhard

A company fine-tunes a foundation model on SageMaker JumpStart for sentiment analysis. After deployment, the model shows bias toward positive sentiment. Which action should be taken to mitigate bias?

A.Use a different foundation model

B.Add more positive examples to training data

C.Increase training epochs

D.Perform RLHF (Reinforcement Learning from Human Feedback) to align outputs

AnswerD

RLHF uses human feedback to reduce undesirable biases.

Why this answer

RLHF (Reinforcement Learning from Human Feedback) is the correct approach because it directly addresses the misalignment between the model's outputs and desired human values. By collecting human feedback on model outputs and using it to train a reward model, RLHF fine-tunes the foundation model to reduce biased behavior, such as the over-prediction of positive sentiment, without simply reweighting the training data.

Exam trap

AWS often tests the misconception that bias is solely a data quantity issue, leading candidates to incorrectly choose adding more examples (Option B) instead of recognizing that alignment techniques like RLHF are required to correct model behavior after training.

How to eliminate wrong answers

Option A is wrong because simply switching to a different foundation model does not guarantee the removal of bias; the new model may have its own biases or the same underlying training data issues. Option B is wrong because adding more positive examples would exacerbate the existing bias toward positive sentiment, not mitigate it. Option C is wrong because increasing training epochs does not correct bias; it risks overfitting the model to the existing biased distribution, making the bias worse.

Full explanation →

300

MCQhard

A company is deploying a generative AI model on Amazon Bedrock. The model is accessed by an application running on Amazon ECS Fargate. The security team requires that all data sent to and from the Bedrock model be encrypted in transit, and that the ECS task does not have internet access. The ECS tasks are launched in a private VPC subnet with a VPC endpoint for Bedrock configured. However, when the application attempts to invoke the Bedrock model, the call fails with a timeout error. The VPC endpoint policy is set to allow all actions from all principals. What is the most likely cause of the failure?

A.The ECS tasks are not configured to use a NAT gateway to reach the Bedrock endpoint

B.The VPC endpoint policy does not grant the ECS task IAM role permission to invoke Bedrock

C.The ECS task IAM role does not have the bedrock:InvokeModel permission

D.The security group associated with the VPC endpoint does not allow inbound HTTPS traffic from the ECS tasks

AnswerD

The security group must allow inbound HTTPS from the ECS tasks' security group; otherwise, traffic is blocked.

Why this answer

The timeout error indicates a network connectivity issue, not an authentication or authorization problem. Since the ECS tasks are in a private subnet without internet access and a VPC endpoint for Bedrock is configured, the most likely cause is that the security group associated with the VPC endpoint does not allow inbound HTTPS (port 443) traffic from the ECS tasks. Without this inbound rule, the TCP handshake fails, resulting in a timeout.

Exam trap

AWS often tests the distinction between network-level failures (timeouts) and authorization failures (access denied), and the trap here is that candidates confuse VPC endpoint policy with security group rules, assuming a permissive policy is sufficient for connectivity.

How to eliminate wrong answers

Option A is wrong because a NAT gateway is used for internet access, but the VPC endpoint allows private connectivity to Bedrock without internet; a NAT gateway is not required and would actually violate the requirement that the ECS task has no internet access. Option B is wrong because the VPC endpoint policy is set to allow all actions from all principals, so it does not restrict the ECS task IAM role; the issue is a network timeout, not a policy denial. Option C is wrong because the IAM role lacking bedrock:InvokeModel would cause an access denied error (HTTP 403), not a timeout; the timeout indicates the request never reached Bedrock.

Full explanation →

Page 4 of 9

All pages

1 2 3 4 5 6 7 8 9

Practice AIF-C01 by domain

Target a specific domain to shore up weak areas.

Applications of Foundation Models Security, Compliance, and Governance for AI Solutions Fundamentals of AI and ML Fundamentals of Generative AI Guidelines for Responsible AI

See all domains with question counts →