AWS AI Practitioner (AIF-C01) — Free Sample Questions
Explore 30 free sample questions for the AWS Certified AI Practitioner exam (AIF-C01). Each includes a full explanation covering correct and incorrect answers so you learn as you practice.
Exam Domains Covered
- Fundamentals of AI and ML (20%) — supervised vs. unsupervised learning, neural networks, common ML algorithms
- Fundamentals of Generative AI (24%) — transformers, diffusion models, prompt engineering basics
- Applications of Foundation Models (28%) — Amazon Bedrock, model selection, RAG, fine-tuning
- Guidelines for Responsible AI (14%) — bias detection, fairness metrics, transparency requirements
- Security, Compliance, and Governance (14%) — data privacy for AI, model access controls, audit logging
Why These Questions Are Different
The AIF-C01 exam was introduced in 2024 and focuses heavily on generative AI concepts unique to AWS — Bedrock runtime selection, model customization jobs, knowledge bases, guardrails. Our questions test these AWS-specific topics, not generic ML theory.
How to Approach AI Practitioner Prep
- Understand the difference between foundation models, fine-tuning, and RAG
- Know which Bedrock models suit which use cases (Claude for reasoning, Titan for embeddings)
- Study responsible AI guardrails and how AWS implements them
- Practice interpreting scenario-based questions — the exam is never just "define X"
Continue Your Preparation
After these 30 samples, access 600 AI Practitioner questions spanning 30 quizzes and 2 full-length exams — each mapped to specific AIF-C01 domains and task statements.
All 30 Free Sample Questions
Question 1
A system is described as having three characteristics: it can perceive its environment, reason about information, and take actions to achieve goals. Which term best describes this system?
- Artificial Intelligence system — Correct answer
- Machine Learning system
- Neural Network
- Deep Learning system
Explanation:
- Why correct: Artificial Intelligence is the broad field encompassing systems that can perceive their environment, reason about information, and take actions to achieve goals. These three characteristics—perception, reasoning, and action—are the foundational definition of an AI system.
- Why B is wrong: Machine Learning is a subset of AI focused specifically on systems that learn patterns from data. The description covers broader intelligent behavior (perceive, reason, act) that goes beyond just learning from data.
- Why C is wrong: A Neural Network is a specific computational architecture inspired by the brain, used to implement certain ML algorithms. It does not describe the broad concept of perceiving, reasoning, and acting intelligently.
- Why D is wrong: Deep Learning is a narrow subset of ML that uses neural networks with many layers. The scenario describes general intelligent system behavior, not a specific multi-layered architecture.
Question 2
A media company wants to automatically extract text from scanned documents and forms. Which AWS service should they use?
- Amazon Comprehend
- Amazon Rekognition
- Amazon Textract — Correct answer
- Amazon Transcribe
Explanation:
- Why correct: Amazon Textract uses machine learning to automatically extract printed text, handwriting, and structured data (tables, forms) from scanned documents, making it the ideal service for this use case.
- Why A is wrong: Amazon Comprehend is an NLP service that analyzes existing text for sentiment, entities, and key phrases—it cannot extract text from images or scanned documents.
- Why B is wrong: Amazon Rekognition analyzes images for objects, faces, and scenes, but it is not optimized for extracting structured text and form data from documents the way Textract is.
- Why D is wrong: Amazon Transcribe converts speech audio into text. It processes audio input, not scanned document images.
Question 3
A data science team is planning the initial stages of their ML project before model training begins. What is the correct order of these stages?
- Data collection → Exploratory data analysis → Preprocessing → Feature engineering — Correct answer
- Feature engineering → Data collection → Training → Monitoring
- Training → Deployment → Data collection → Evaluation
- Data collection → Preprocessing → Training → Exploratory data analysis
Explanation:
- Why correct: The ML pipeline begins with data collection to gather raw data, followed by exploratory data analysis (EDA) to understand distributions and relationships, then preprocessing to clean and normalize the data, and finally feature engineering to transform data into model-ready inputs.
- Why B is wrong: Feature engineering depends on understanding the data (EDA) and having clean data (preprocessing). It cannot logically precede data collection.
- Why C is wrong: Training must happen before deployment, and data must be collected and prepared before training. This sequence reverses the fundamental pipeline flow.
- Why D is wrong: Preprocessing occurs after data collection and EDA, not after training. Training requires already-preprocessed data as input.
Question 4
A company wants to build a system that learns to identify spam emails by analyzing thousands of labeled examples. What combination of learning type and ML technique is most appropriate?
- Reinforcement learning with rewards
- Unsupervised learning with clustering
- Supervised learning with classification — Correct answer
- Supervised learning with regression
Explanation:
- Why correct: Spam detection with labeled examples (spam/not spam) is a supervised classification problem—predicting categorical outcomes from labeled data.
- Why A is wrong: Reinforcement learning uses an agent that learns through trial-and-error with a reward signal. Email classification from labeled data is not a sequential decision-making problem, making RL inappropriate.
- Why B is wrong: Unsupervised clustering does not use labels, but the scenario explicitly provides thousands of labeled examples (spam/not spam), so a supervised approach is required.
- Why D is wrong: Supervised regression predicts continuous numerical values (such as prices or temperatures), but spam detection is a categorical binary problem (spam or not spam).
Question 5
A data science team is building a new large language model. They begin by exposing it to billions of web pages, books, and scientific papers so it can learn broad language patterns before any task-specific training. Which stage of the foundation model lifecycle are they currently performing?
- Pre-training — Correct answer
- Fine-tuning
- Evaluation
- Deployment
Explanation:
- Why correct: Pre-training is the first and most resource-intensive stage of the foundation model lifecycle. The model is exposed to massive, diverse datasets so it develops broad language understanding, factual knowledge, and reasoning capabilities before any task-specific specialization.
- Why B is wrong: Fine-tuning comes after pre-training and involves training the already pre-trained model on a smaller, curated, task-specific dataset to specialize it for a particular domain; the scenario describes the initial broad learning phase, not a specialization phase.
- Why C is wrong: Evaluation is the stage where a trained model is benchmarked against held-out test data to measure quality and readiness for deployment; it involves assessing the model, not training it on new data.
- Why D is wrong: Deployment is the production stage where a fully trained and evaluated model is made accessible to applications or end users; all training and evaluation steps must be completed first.
Question 6
A company is choosing between a rule-based system and a GenAI solution for customer support. Their product catalog changes quarterly and they frequently launch new service categories. Which GenAI advantage is the strongest argument for choosing GenAI over the rule-based approach?
- Adaptability — the model handles new products and categories without rewriting rules — Correct answer
- Determinism — the model produces identical answers every time
- Lower infrastructure cost — GenAI always requires fewer servers than rule-based systems
- Guaranteed accuracy — GenAI never makes mistakes in customer interactions
Explanation:
- Why correct: Adaptability is a core GenAI advantage. When a product catalog changes frequently, a rule-based system requires engineers to update rules for every new product or category. GenAI models generalize from their training and can handle new scenarios without manual reprogramming, making them more practical for rapidly changing business needs.
- Why B is wrong: Determinism means producing identical outputs for identical inputs. GenAI is actually nondeterministic, and this property is unrelated to the business advantage of handling frequent catalog changes.
- Why C is wrong: GenAI does not inherently require fewer servers. Infrastructure cost depends on model size, request volume, and deployment choices — it is not a guaranteed advantage over rule-based systems.
- Why D is wrong: GenAI models can hallucinate and produce errors. Guaranteed accuracy is not a GenAI characteristic and would not be a valid reason to choose it over a rule-based system.
Question 7
A company wants to access multiple foundation models from different providers through a single managed service. Which AWS service provides this capability?
- Amazon S3
- Amazon Bedrock — Correct answer
- Amazon SageMaker AI only
- Amazon EC2
Explanation:
- Why correct: Amazon Bedrock provides access to multiple foundation models from various providers — including Anthropic, Meta, AI21 Labs, Cohere, Stability AI, and Amazon — through a single unified managed API. Customers can switch between models without changing infrastructure.
- Why A is wrong: Amazon S3 is an object storage service for storing and retrieving data. It does not provide access to foundation models or AI inference capabilities.
- Why C is wrong: Amazon SageMaker AI is a platform for building, training, and deploying custom ML models. While it provides model hosting, it is not designed as a single managed API for accessing multiple pre-built foundation models from different providers the way Bedrock is.
- Why D is wrong: Amazon EC2 provides general-purpose compute infrastructure (virtual machines). You could manually deploy models on EC2 instances, but it does not offer managed access to multiple foundation models through a unified API.
Question 8
A team is explaining the core architecture of LLM-based GenAI. Which TWO represent the structural building blocks that LLMs are constructed from, as opposed to techniques or settings applied to an already-built model?
- Tokens as basic text processing units — Correct answer
- Prompt engineering techniques
- Foundation models as pre-trained base models — Correct answer
- Vector embeddings for semantic representation
- Temperature as a sampling parameter
Explanation:
- Why correct: Tokens (A) are the fundamental units that LLMs are built to process—text is broken into these discrete pieces before any computation happens, making them a structural requirement of every LLM. Foundation models (C) are the large pre-trained neural networks themselves—without them, there is no GenAI system. Together, tokens and foundation models are what LLM-based GenAI is made of.
- Why B is wrong: Prompt engineering is a useful technique for guiding model outputs, but it is applied to an already-built model. It is not a structural component of the system itself.
- Why D is wrong: Vector embeddings are a representation method used to encode meaning numerically. While important, they are a product of the model's internal processing, not a building block the system is constructed from.
- Why E is wrong: Temperature is a sampling parameter that controls output randomness at inference time. It is a configuration setting applied to the model, not part of the model's structure.
Question 9
A real-time customer service chatbot requires responses within 2 seconds. Which model selection criterion is most critical?
- Model latency and response time — Correct answer
- Training data size only
- Model size only
- Model benchmark accuracy on standard NLP tasks
Explanation:
- Why correct: For a real-time chatbot with a strict 2-second response SLA, model latency and response time is the binding selection constraint. A highly accurate model that consistently responds in 5 seconds fails the requirement entirely, regardless of any other quality.
- Why B is wrong: Training data size is a development-time factor that influences model capability, but it does not directly indicate the inference latency the deployed model will achieve on production traffic.
- Why C is wrong: Model size (parameter count) correlates with latency but is not the direct measure of it. With optimized serving infrastructure, larger models can still meet latency targets. The selection criterion should be measured latency, not model size.
- Why D is wrong: Benchmark accuracy on standard NLP tasks measures output quality, not how quickly responses are generated. An accurate model can still fail a 2-second SLA if it is too slow. Latency must be validated separately.
Question 10
A developer is configuring inference parameters for a foundation model and sets top-k to 50. What effect does this setting have on the model's output generation?
- It limits the total response to exactly 50 tokens
- It restricts token selection at each step to the 50 highest-probability candidates — Correct answer
- It causes the model to generate 50 independent responses and return the best one
- It increases the model's context window to 50,000 tokens
Explanation:
- Why correct: Top-k sampling limits the model’s token selection pool at each generation step to only the k highest-probability tokens. With top-k set to 50, the model considers only the 50 most likely next tokens and samples from that restricted set. This reduces the chance of selecting very low-probability tokens while still allowing diversity within the top candidates. It is one of the three primary sampling parameters alongside temperature and top-p.
- Why 'It limits the total response to exactly 50 tokens' is wrong: The parameter that controls response length is max_tokens (maximum output length), not top-k. Top-k operates at the per-token selection level, not the total response level. A response with top-k=50 can be any length up to the max_tokens limit.
- Why 'It causes the model to generate 50 independent responses' is wrong: Top-k does not trigger multiple response generation. It constrains the candidate pool at each individual token-sampling step within a single response. Generating multiple responses would require separate API calls or a specific ‘n’ parameter.
- Why 'It increases the model’s context window to 50,000 tokens' is wrong: The context window is a fixed architectural property of the model determined at training time. Inference parameters like top-k have no effect on how many input tokens the model can process.
Question 11
A company is preparing data for fine-tuning. Which THREE data quality traits are important?
- High quality and accuracy — Correct answer
- Maximized for total volume over all other considerations
- Representative of target scenarios — Correct answer
- Sourced exclusively from a single narrow use case
- Properly curated and filtered — Correct answer
- Collected at the lowest possible cost to maximize quantity
Explanation:
- Why correct: High quality and accuracy (index 0), representativeness of target scenarios (index 2), and proper curation and filtering (index 4) are the three foundational data quality traits required for effective fine-tuning. High-quality data ensures the model learns correct patterns; representativeness ensures the data reflects the full range of real-world scenarios the model will encounter; and curation/filtering removes noise, duplicates, and inappropriate content before training.
- Why index 1 is wrong: Prioritizing total data volume over quality, accuracy, and relevance can introduce noisy, incorrect, or off-topic examples that degrade fine-tuning effectiveness. Volume alone does not ensure the model learns correct patterns.
- Why index 3 is wrong: Data sourced exclusively from a single narrow use case lacks the breadth needed to represent the full range of target scenarios, limiting the model’s ability to generalize across the variety of inputs it will encounter in production.
- Why index 5 is wrong: Minimizing data collection cost often means skipping quality review, curation, and filtering steps. The resulting dataset may contain errors, duplicates, and irrelevant content that hurt model performance.
Question 12
A company recently deployed a foundation model application to production. After launch, the team notices user satisfaction scores are declining but no code changes were made. Which evaluation practice would have detected this drift earliest?
- Evaluate the model once at launch and rely on that baseline
- Continuous evaluation and monitoring in production — Correct answer
- Evaluate only the first 100 production requests
- Only evaluate during development before deployment
Explanation:
- Why correct: Continuous evaluation and monitoring in production detects performance degradation, data drift, and model quality issues over time. FM outputs can degrade as user behavior shifts, input distributions change, or the underlying model is updated — none of which a one-time evaluation can catch.
- Why A is wrong: Evaluating only at launch creates a false sense of security. Production environments introduce real-world variability that pre-launch testing cannot fully replicate. A model that performs well at launch may degrade as user inputs evolve.
- Why C is wrong: Evaluating only on the first 100 requests gives a narrow, potentially non-representative sample. It misses regressions that appear with edge-case inputs or during scaling. Sustained monitoring across diverse traffic is required for reliable quality assurance.
- Why D is wrong: Development-phase evaluation validates the model in a controlled setting, but production traffic is different in volume, diversity, and intent. Only production monitoring can surface issues that emerge from real user interactions.
Question 13
A company wants to improve their foundation model's response quality and alignment by systematically collecting human ratings of model outputs and using those ratings to guide fine-tuning. Which technique should they use?
- Prompt engineering
- Benchmark dataset evaluation
- Continuous pre-training
- Reinforcement Learning from Human Feedback (RLHF) — Correct answer
Explanation:
- Why correct: RLHF (Reinforcement Learning from Human Feedback) is the technique where human raters evaluate model outputs, and those ratings are used to train a reward model. That reward model then guides fine-tuning via reinforcement learning, aligning the FM's responses with human preferences for quality, helpfulness, and safety.
- Why A is wrong: Prompt engineering modifies the input prompt to guide the model's behavior without changing model weights. It cannot systematically incorporate structured human feedback into the model training loop.
- Why B is wrong: Continuous pre-training extends a model's training on new raw text data to broaden general knowledge. It does not incorporate human ratings or preference signals to improve response quality and alignment.
- Why C is wrong: Transfer learning adapts a model trained on one task to perform another, typically by fine-tuning on a new dataset. It does not use a human feedback loop or reward model to align outputs with human preferences.
Question 14
A hospital is deploying an AI-assisted diagnostic tool. The project team adds extensive testing, red-teaming, and automated monitoring to ensure the system will not misdiagnose patients or recommend harmful treatments. Which responsible AI principle is the team primarily addressing?
- Fairness
- Safety — Correct answer
- Veracity
- Robustness
Explanation:
- Why correct: Safety is the responsible AI principle that requires AI systems to avoid causing physical, psychological, or material harm to users and third parties. In healthcare, this means the system must not produce outputs (e.g., wrong diagnoses or harmful treatment suggestions) that could injure patients. Testing, red-teaming, and monitoring are all safety mechanisms.
- Why option A is wrong: Fairness addresses whether the model produces equitable outcomes across demographic groups. It is relevant in healthcare, but it does not capture the core concern of preventing direct patient harm described in the scenario.
- Why option C is wrong: Veracity (truthfulness) refers to ensuring the model does not hallucinate or produce false information. While related, the scenario focuses on preventing harm, not specifically on factual accuracy of outputs.
- Why option D is wrong: Robustness refers to the system maintaining reliable performance under varied or adversarial inputs. Robustness is a prerequisite for safety, but the overarching principle being demonstrated — not causing harm — is safety.
Question 15
A data science team is selecting a model type for a credit-scoring application subject to regulatory review. Regulators require that individual decisions be explainable to applicants. Which of the following models would be considered the LEAST transparent and hardest to explain?
- A deep neural network with 50 layers trained on proprietary data — because its internal weight calculations cannot be directly mapped to human-understandable rules — Correct answer
- A decision tree with a depth of 3 — because its branching logic is too complex for non-technical users to follow
- A linear regression model — because it uses mathematical coefficients that stakeholders cannot interpret
- A k-nearest neighbors classifier — because its predictions depend on distance calculations that are invisible to users
Explanation:
- Why A is correct: Deep neural networks (DNNs) are classic examples of opaque or 'black-box' models. Their predictions emerge from millions of learned weights across many layers, making it impossible to directly trace why a specific input produced a specific output. This lack of intrinsic interpretability is precisely why they contrast with transparent models in responsible AI frameworks. Post-hoc methods like SHAP or LIME can partially explain predictions, but the model itself does not provide human-readable decision logic.
- Why B is wrong: A shallow decision tree (depth 3) is one of the most interpretable model types available. Each leaf node represents a clear, traceable decision path based on feature thresholds, making it easy to explain individual predictions to regulators and applicants.
- Why C is wrong: Linear regression is inherently transparent. Its coefficients directly quantify the contribution of each input feature to the output, and the relationship is mathematically explicit. Stakeholders can understand and audit coefficient values.
- Why D is wrong: While k-nearest neighbors uses distance calculations, its predictions can still be explained at an instance level by showing the k most similar training examples. It is more interpretable than a deep neural network in practice.
Question 16
A company publishes a SageMaker Model Card for their new credit-scoring AI system before releasing it to business stakeholders. Which outcome does this most directly support?
- Grants regulatory approval to deploy the model in production
- Enables stakeholders to understand the model's capabilities, limitations, and appropriate use cases — Correct answer
- Continuously monitors the model for data drift and alerts the team when production distributions shift
- Automatically generates SHAP values to explain individual credit-scoring decisions to applicants
Explanation:
- Why correct: SageMaker Model Cards provide structured, standardized documentation that captures a model's intended use, performance metrics across demographic groups, training data details, known limitations, and responsible AI considerations. Publishing a model card before release ensures stakeholders understand what the model can and cannot do, reducing the chance of misuse and building informed trust — a core transparency tool per AWS responsible AI guidance.
- Why A is wrong: A model card documents performance characteristics and limitations but does not itself grant regulatory approval to deploy. Regulatory approval depends on a formal compliance and risk-governance process. A model card is an input to that process, but publishing it does not substitute for it.
- Why C is wrong: SageMaker Model Monitor (not Model Cards) is used to detect production data drift and model quality degradation. Model Cards are documentation artifacts; they have no real-time monitoring or alerting capability and do nothing to catch distribution shifts after deployment.
- Why D is wrong: SHAP values, produced by Amazon SageMaker Clarify's explainability processing job, explain individual predictions by quantifying each input feature's contribution. A model card is a static documentation artifact describing the model at a high level — it does not generate per-prediction feature attribution values.
Question 17
A company wants to ensure their Amazon SageMaker training jobs and inference endpoints are not accessible from the public internet. Which AWS service is the foundation for achieving this network isolation?
- AWS PrivateLink
- Amazon Virtual Private Cloud (VPC) — Correct answer
- AWS WAF
- Amazon GuardDuty
Explanation:
- Why correct: Amazon VPC enables you to deploy AWS resources in a logically isolated private network. For AI workloads, this means that components like SageMaker training jobs, inference endpoints, and data preprocessing pipelines can be placed in private subnets with no direct internet exposure. Combined with VPC security groups and NACLs, VPC is the foundational AWS service for network-level isolation.
- Why A is wrong: AWS PrivateLink enables private connectivity between VPCs and AWS services over the AWS backbone without traversing the public internet — but it complements VPC rather than replacing it. PrivateLink is used to securely access service endpoints privately, while VPC provides the broader network isolation boundary.
- Why C is wrong: AWS WAF (Web Application Firewall) filters HTTP/HTTPS traffic at the application layer, protecting against web exploits like SQL injection and XSS. It does not provide network isolation for backend AI workloads running in private infrastructure.
- Why D is wrong: Amazon GuardDuty is a threat detection service that monitors for suspicious activity using CloudTrail, VPC Flow Logs, and DNS logs. It detects threats but does not isolate network resources or create private network boundaries.
Question 18
A company deploys a customer-facing chatbot powered by Amazon Bedrock. The product team is concerned that users may not realize they are interacting with an AI system rather than a human agent. Which AI governance protocol element addresses this concern?
- Encrypt all chatbot conversation logs using AWS KMS
- Implement transparency standards requiring disclosure of AI-generated interactions — Correct answer
- Configure role-based access control to restrict chatbot modifications
- Deploy automated CI/CD pipelines for rapid chatbot updates
Explanation:
- Why correct: Transparency standards are a core governance protocol element that requires organizations to disclose when users are interacting with an AI system. This includes labeling AI-generated content, publishing model cards that describe system capabilities and limitations, and ensuring users can make informed decisions about their interactions. Leading AI governance frameworks — including the NIST AI Risk Management Framework and the EU AI Act — identify transparency as a foundational requirement for trustworthy AI.
- Why A is wrong: Encryption of conversation logs using AWS KMS is a security control that protects data confidentiality at rest. It does not address whether users know they are interacting with an AI system — an encrypted conversation is still an undisclosed AI interaction.
- Why C is wrong: Role-based access control (RBAC) restricts who can modify or manage the chatbot internally. It is an access management control, not a mechanism for disclosing AI use to external end users.
- Why D is wrong: Automated CI/CD pipelines accelerate deployment of chatbot updates but do not address user-facing transparency. A faster deployment pipeline does not inform users that they are communicating with AI rather than a human.
Question 19
A security engineer wants to continuously scan the EC2 instances and container images that support an AI inference pipeline for known software vulnerabilities (CVEs) and unintended network exposure. Which AWS service is designed for this purpose?
- Amazon Inspector — Correct answer
- Amazon GuardDuty
- AWS Security Hub
- AWS Trusted Advisor
Explanation:
- Why correct: Amazon Inspector is a vulnerability management service that automatically and continuously scans EC2 instances and Amazon ECR container images for software vulnerabilities (CVEs) and unintended network accessibility. It integrates with AWS Security Hub to centralize findings and supports the continuous vulnerability management lifecycle required for AI infrastructure protection.
- Why B is wrong: Amazon GuardDuty is a threat detection service that analyzes CloudTrail, VPC Flow Logs, and DNS logs to identify malicious or unauthorized behavior such as reconnaissance, compromised credentials, or unusual data access patterns. GuardDuty detects active threats but does not perform CVE scanning or assess patch compliance — that is Inspector's role.
- Why C is wrong: AWS Security Hub aggregates, normalizes, and prioritizes security findings from multiple services (including Inspector and GuardDuty). It is a single-pane-of-glass tool for viewing security posture, not a scanning engine itself — it depends on Inspector to generate the vulnerability findings.
- Why D is wrong: AWS Trusted Advisor provides best-practice recommendations across cost, security, fault tolerance, and performance. Its security checks cover things like open S3 buckets or unrestricted security group ports, but it does not perform continuous vulnerability assessment or CVE scanning of EC2 instances and container images.
Question 20
A healthcare company uses an ML model to classify medical images. Because errors could have serious consequences, they need a workflow that automatically routes low-confidence predictions to qualified human reviewers before results are acted upon. Which AWS service provides this human-in-the-loop review capability?
- Amazon SageMaker Ground Truth
- Amazon SageMaker Model Monitor
- Amazon Augmented AI (A2I) — Correct answer
- Amazon Rekognition Custom Labels
Explanation:
- Why correct: Amazon Augmented AI (A2I) is purpose-built to add human review to ML prediction workflows. You define confidence thresholds that automatically route low-confidence outputs to a human workforce (private team, Mechanical Turk, or AWS Marketplace vendors), and A2I manages the review process and result collection—all without building custom infrastructure.
- Why Amazon SageMaker Ground Truth is wrong: Ground Truth manages human labeling of training datasets to build and improve ML models; it operates before deployment to create labeled data, not to review live model predictions in a production pipeline.
- Why Amazon SageMaker Model Monitor is wrong: Model Monitor detects statistical drift and data quality degradation in deployed models over time; it alerts teams to model degradation but does not route individual low-confidence predictions for human review.
- Why Amazon Rekognition Custom Labels is wrong: Rekognition Custom Labels trains custom computer vision classifiers using your labeled images; it is a model-building service and does not orchestrate a human review layer for live prediction outputs.
Question 21
A company runs AI inference workloads on Amazon EC2 instances and Amazon SageMaker endpoints. The security team wants to automatically detect if any of these resources communicate with known malicious IP addresses or exhibit unexpected behaviors such as cryptocurrency mining. Which AWS service should they enable?
- Amazon GuardDuty — Correct answer
- Amazon Macie
- AWS Security Hub
- Amazon Inspector
Explanation:
- Why correct: Amazon GuardDuty is a managed intelligent threat detection service that continuously monitors AWS accounts, workloads, and data using machine learning, anomaly detection, and integrated threat intelligence feeds. It detects threats such as instances communicating with known malicious IP addresses, unexpected cryptocurrency mining activity, and compromised credential usage — without requiring agents or additional infrastructure changes.
- Why B is wrong: Amazon Macie uses machine learning to discover and classify sensitive data (such as PII) stored in Amazon S3. It focuses on data visibility and classification within storage, not on detecting behavioral anomalies or active threats originating from compute workloads like EC2 or SageMaker endpoints.
- Why C is wrong: AWS Security Hub aggregates and centralizes security findings from multiple services — including GuardDuty, Inspector, and Macie — into a single dashboard and checks your environment against security standards. It does not independently perform threat detection; it requires underlying services to generate the findings it consolidates.
- Why D is wrong: Amazon Inspector automatically assesses EC2 instances, Lambda functions, and container images for software vulnerabilities (CVEs) and unintended network exposure. It performs static vulnerability and reachability assessment, not real-time behavioral threat detection based on network traffic and API call patterns.
Question 22
A company stores structured ML training metadata and model performance logs and needs to run complex analytical SQL queries across terabytes of historical data with fast, consistent response times. The data will be queried frequently by analysts using BI tools. Which AWS service is best suited as the data warehouse for this workload?
- Amazon Athena
- AWS Glue
- Amazon Redshift — Correct answer
- Amazon EMR
Explanation:
- Why correct: Amazon Redshift is a fully managed, petabyte-scale data warehouse optimized for complex analytical queries using columnar storage, result caching, and massively parallel processing (MPP). It integrates natively with BI tools and delivers consistent high performance for frequent, complex workloads on large structured datasets.
- Why A is wrong: Amazon Athena is a serverless query engine that runs ad-hoc SQL directly against data in S3. While cost-effective for occasional queries, it lacks the persistent storage, query performance optimizations (sort keys, distribution keys, result caching), and consistent throughput of a dedicated data warehouse under heavy, frequent analytical workloads.
- Why B is wrong: AWS Glue is a serverless ETL service used to prepare and transform data before loading it into an analytics store. It is not a data warehouse for storing and running analytical queries.
- Why D is wrong: Amazon EMR runs distributed big data processing frameworks such as Spark and Hadoop. While EMR supports SQL-on-Hadoop tools like Hive and Presto, it is not a persistent, optimized data warehouse—it requires managing cluster infrastructure and lacks the columnar storage engine, sort keys, distribution keys, and result caching that make Redshift ideal for consistent, high-performance analytical workloads with BI tools.
Question 23
An AI team is building a Retrieval-Augmented Generation (RAG) pipeline and needs to store dense vector embeddings and perform approximate nearest-neighbor (ANN) similarity searches at query time. Which TWO AWS services are purpose-built to support vector storage and similarity search?
- Amazon RDS for PostgreSQL with pgvector — Correct answer
- Amazon DynamoDB with a composite sort key
- Amazon OpenSearch Service with vector engine — Correct answer
- Amazon Aurora MySQL with full-text index
- Amazon S3 with S3 Select
Explanation:
- Why correct: Amazon RDS for PostgreSQL with the pgvector extension (option A) adds a native vector data type and IVFFlat/HNSW indexes, enabling cosine, L2, and inner-product similarity search directly inside a relational database. Amazon OpenSearch Service with the k-NN vector engine (option C) provides ANN search at scale using HNSW indexes across billions of vectors, and is a supported vector store for Amazon Bedrock Knowledge Bases.
- Why B is wrong: Amazon DynamoDB is a key-value and document store. Its composite sort key enables range scans and sorting on scalar values, but it has no vector data type, no embedding index, and no ANN search capability. Similarity search requires comparing high-dimensional floating-point vectors, which DynamoDB cannot do natively.
- Why D is wrong: Amazon Aurora MySQL supports full-text indexes that enable keyword-based MATCH…AGAINST queries, but keyword search operates on tokenized text, not on dense vector embeddings. It cannot compute cosine similarity or Euclidean distance between embedding vectors.
- Why E is wrong: Amazon S3 with S3 Select allows querying structured data (CSV, JSON, Parquet) using SQL-like expressions, but it is an object storage service with no vector index, no embedding data type, and no nearest-neighbor search capability.
Question 24
A company wants recommendations on cost optimization, security, and best practices for their AI workloads. What AWS service should they use?
- AWS Cost Explorer for per-service spending analysis and visualization
- AWS Well-Architected Tool for structured six-pillar workload reviews
- AWS Trusted Advisor — Correct answer
- Amazon CloudWatch for operational metrics and performance monitoring
Explanation:
- Why correct: AWS Trusted Advisor analyzes your AWS environment in real time and provides automated recommendations across six categories: cost optimization, security, fault tolerance, performance, service limits, and operational excellence—making it the right fit for ongoing AI workload best-practice guidance without manual review effort.
- Why A is wrong: AWS Cost Explorer provides detailed spending analysis and visualization by service, resource tag, and time period, but it does not proactively surface configuration risks or flag security and reliability issues. It shows where money is going, not whether your setup follows best practices.
- Why B is wrong: The AWS Well-Architected Tool conducts a structured six-pillar review and produces a personalized improvement plan, but it requires a manual review session and generates a one-time workload assessment rather than continuous automated checks across your entire environment.
- Why D is wrong: Amazon CloudWatch collects operational metrics, creates dashboards, and triggers alarms based on thresholds, but it monitors runtime system health rather than evaluating architectural decisions for cost efficiency, security posture, or service limit risks.
Question 25
A data science team stores ML training datasets in Amazon S3 in us-east-1. Regulations require that a copy of all training data be maintained in a separate AWS Region for disaster recovery. The team wants this replication to happen automatically whenever new data is uploaded. Which S3 feature meets this requirement with the least operational overhead?
- S3 Cross-Region Replication (CRR) with a replication rule — Correct answer
- S3 Same-Region Replication (SRR) targeting a second bucket in us-east-1
- A scheduled AWS Lambda function that copies objects to a bucket in another region nightly
- S3 Versioning without any replication configuration
Explanation:
- Why A is correct: S3 Cross-Region Replication (CRR) automatically copies objects to a destination bucket in a different AWS Region every time a new object is written. You configure a replication rule once, and S3 handles it without additional infrastructure. CRR is the purpose-built feature for cross-region disaster recovery of S3 data.
- Why B is wrong: S3 Same-Region Replication (SRR) replicates objects within the same region. It does not satisfy a requirement to store a copy in a separate AWS Region, so it would not meet the disaster recovery regulatory requirement.
- Why C is wrong: A nightly Lambda function introduces operational complexity, latency (up to 24 hours of data could be missing in a disaster), and potential failure points. CRR is real-time and serverless with no custom code to maintain.
- Why D is wrong: S3 Versioning preserves multiple versions of an object within the same bucket and region. It protects against accidental overwrites or deletes, but it does not replicate data to another region for geographic disaster recovery.
Question 26
A security team manages an AI training environment with dozens of EC2 instances across multiple subnets in a VPC. A threat intelligence feed identifies a malicious IP range that must be blocked immediately. The team needs to create an explicit deny rule that applies to all current and future instances in a specific subnet without modifying each instance's individual firewall configuration. Which VPC component should they configure?
- Security Groups attached to each EC2 instance
- AWS WAF rules on the Application Load Balancer
- VPC Route Tables controlling traffic paths
- Network Access Control Lists (NACLs) — Correct answer
Explanation:
- Why correct: Network ACLs (NACLs) are stateless, subnet-level firewalls applied to all traffic entering or leaving a subnet. They evaluate rules in order and can explicitly allow or deny traffic by IP range, port, and protocol. Because they operate at the subnet boundary, they provide a consistent policy layer across all resources in that subnet without requiring per-instance configuration.
- Why A is wrong: Security Groups are stateful firewalls applied at the instance (ENI) level, not at the subnet level. They cannot control traffic between subnets—traffic must reach the instance before Security Groups evaluate it, which is too late for broad subnet-level isolation.
- Why B is wrong: AWS WAF operates at the application layer (Layer 7) and inspects HTTP/HTTPS requests. It is attached to resources like ALBs or CloudFront distributions, not to individual subnets, and does not control network-level (Layer 3/4) traffic flows between subnets.
- Why C is wrong: Route Tables control where traffic is forwarded (e.g., to an IGW, NAT Gateway, or VPC peering connection) but do not filter or block traffic based on ports or IPs. They are routing directives, not firewalls.
Question 27
A data science team is running multiple Amazon SageMaker training jobs and Amazon Bedrock inference calls. Management wants visibility into monthly AI spending and automated notifications when costs approach budget limits. Which TWO AWS services should they use together? (Select TWO.)
- AWS Budgets for cost alerts and thresholds — Correct answer
- Amazon CloudWatch for billing metric dashboards
- AWS Cost Explorer for cost analysis and trends — Correct answer
- AWS CloudTrail for tracking API usage
- AWS Trusted Advisor for performance checks
Explanation:
- Why correct: AWS Budgets (A) lets you set custom spending thresholds and sends SNS or email alerts when actual or forecasted costs approach or exceed those thresholds. AWS Cost Explorer (C) provides interactive charts for analyzing cost and usage trends over time, with filters by service, linked account, and tag—giving the team the visibility management requested.
- Why B is wrong: Amazon CloudWatch monitors operational metrics such as CPU, memory, and custom application metrics. While a billing metric alarm exists, it only provides a single threshold alert and lacks the rich cost analysis, filtering, and forecasting capabilities of Cost Explorer and Budgets together.
- Why D is wrong: AWS CloudTrail records API calls for auditing and compliance, not for analyzing or alerting on cost.
- Why E is wrong: AWS Trusted Advisor provides recommendations across cost optimization, security, performance, and fault tolerance, but it does not offer the interactive spending analysis or customizable budget alerts the scenario requires.
Question 28
A company is containerizing their ML inference pipeline. Their engineering team has extensive Kubernetes experience from on-premises clusters. Which AWS container orchestration service should they choose to leverage their existing skills?
- Amazon ECS with Fargate, because managed containers are equivalent to managed Kubernetes
- AWS Lambda with container image support, because it can run any containerized workload
- Amazon ECS with the EC2 launch type for full control over container orchestration
- Amazon EKS, because it provides managed Kubernetes compatible with existing Kubernetes tooling — Correct answer
Explanation:
- Why correct: Amazon EKS (Elastic Kubernetes Service) is a managed service that runs certified upstream Kubernetes. Teams with existing Kubernetes expertise can reuse their manifests, Helm charts, and kubectl tooling with minimal changes, reducing migration effort and retraining costs.
- Why A is wrong: Amazon ECS with Fargate is a managed container service, but it uses its own task definition format and APIs—not the Kubernetes API. Managed containers and managed Kubernetes are not equivalent; migrating to ECS requires rewriting all orchestration configurations from Kubernetes-native formats.
- Why B is wrong: AWS Lambda supports deploying functions packaged as container images, but it executes individual function invocations—not Kubernetes-orchestrated workloads. Lambda does not support Kubernetes manifests, Helm charts, pods, services, or deployments.
- Why C is wrong: Amazon ECS with the EC2 launch type gives you control over the underlying EC2 instances, but it still uses ECS-native task definitions and scheduling—not the Kubernetes API. Choosing the EC2 launch type does not add Kubernetes compatibility.
Question 29 — Domain 2: Fundamentals of Generative AI
A startup is building its first GenAI-powered customer support chatbot. The product manager asks the ML team to set expectations with leadership about what GenAI can and cannot do. Which TWO statements accurately describe GenAI capabilities and limitations?
- GenAI models produce deterministic outputs, so the same prompt always returns the identical response
- GenAI eliminates the need for subject matter experts because it can accurately answer any domain-specific question
- GenAI can generate human-like text, code, and creative content based on patterns learned during training — Correct answer
- GenAI models fully understand the meaning and intent behind every user query they receive
- GenAI outputs are always factually accurate when the model is large enough
- GenAI can hallucinate—producing confident-sounding responses that contain fabricated or incorrect information — Correct answer
Explanation:
- Why C is correct: GenAI models, particularly large language models (LLMs), learn patterns from training data and use them to generate text, code, summaries, and creative content that often appears human-written. This generative capability is the core strength of GenAI.
- Why F is correct: Hallucination is a well-documented limitation of GenAI. Models can produce plausible-sounding but factually wrong, fabricated, or unsupported statements because they generate text based on statistical patterns, not verified knowledge retrieval.
- Why A is wrong: GenAI models are generally non-deterministic by default. Sampling-based decoding (controlled by parameters like temperature) means the same prompt can produce different outputs on repeated runs, not identical ones.
- Why B is wrong: GenAI does not replace subject matter experts. Models can produce incorrect domain answers and lack the verified expertise that human specialists provide. Human oversight remains essential for accuracy and safety.
- Why D is wrong: Current GenAI models do not truly understand meaning or intent. They predict statistically likely next tokens based on training patterns. This can appear like understanding but is fundamentally pattern matching, not comprehension.
- Why E is wrong: Model size does not guarantee factual accuracy. Even the largest models hallucinate. Factual grounding requires additional techniques such as RAG or human review, not simply scaling model parameters.
Question 30
A financial services company’s AI lending model must comply with fair lending regulations. During evaluation, the team finds that approval rates differ significantly across demographic groups even when applicants have similar creditworthiness. Which responsible AI principle is MOST directly at risk?
- Transparency and explainability
- Fairness and non-discrimination — Correct answer
- Robustness and reliability
- Privacy and data protection
Explanation:
- Why correct: Fairness and non-discrimination require that AI systems produce equitable outcomes across demographic groups. When similarly qualified applicants receive different approval rates based on demographic characteristics, the model exhibits discriminatory bias, directly violating fairness principles under responsible AI frameworks and regulations like the Equal Credit Opportunity Act.
- Why A is wrong: Transparency and explainability concern whether stakeholders can understand how the model reaches decisions. While important, the core issue described—differential approval rates for similar applicants—is a fairness problem, not a transparency problem.
- Why C is wrong: Robustness and reliability address whether a model performs consistently under varying conditions and adversarial inputs. The scenario describes systematic bias across groups, not inconsistent or unreliable performance.
- Why D is wrong: Privacy and data protection concern safeguarding personal data from unauthorized access or misuse. The issue described is unequal treatment outcomes, which is a fairness concern rather than a data protection concern.
JavaScript is required to use this application.
Please enable JavaScript in your browser to access the full NestedCerts platform.
Contact: support@nestedcerts.com