AWS Certified AI Practitioner (AIF-C01) Exam Questions

A retail company wants a system that can learn from historical sales data and improve its demand forecasts automatically over time, without being explicitly reprogrammed for every new pattern in the data.

Which term BEST describes this capability?

⬜ A. Artificial intelligence (AI)
✅ B. Machine learning (ML)
⬜ C. Business intelligence (BI)
⬜ D. Robotic process automation (RPA)

Explanation:
Machine learning (ML) is a subset of AI in which systems learn patterns from data and improve their performance on a task over time without being explicitly reprogrammed for every scenario.
Incorrect Options:
Option A: AI is the broader field of building systems that can perform tasks that typically require human intelligence; ML is the specific technique described here.
Option C: BI focuses on analyzing historical data to support reporting and decision-making, not on models that learn and improve automatically.
Option D: RPA automates repetitive, rule-based tasks and does not learn from data the way ML does.
Source: What is Machine Learning?

---

Which statement correctly describes the difference between supervised learning and unsupervised learning?

✅ A. Supervised learning uses labeled data to learn a mapping from inputs to known outputs, while unsupervised learning finds patterns or structure in unlabeled data.
⬜ B. Supervised learning only works with image data, while unsupervised learning only works with text data.
⬜ C. Supervised learning requires no training data, while unsupervised learning requires large labeled datasets.
⬜ D. Supervised learning and unsupervised learning are two names for the same technique.

Explanation:
Supervised learning trains a model on labeled data (inputs paired with known correct outputs) to predict outcomes for new data. Unsupervised learning works with unlabeled data and looks for hidden structure, such as clusters or associations, without predefined output labels.
Incorrect Options:
Option B: Both approaches can be applied to any data modality (tabular, text, image, and so on); the distinction is about labeled vs. unlabeled data, not data type.
Option C: Supervised learning requires labeled training data, not the other way around.
Option D: They are distinct learning paradigms with different data requirements and use cases.
Source: What is Machine Learning?

---

A company wants to process a large batch of scanned invoices overnight and does not need results immediately. Which type of inferencing is the MOST cost-effective fit for this use case?

⬜ A. Real-time inferencing
⬜ B. Asynchronous inferencing
✅ C. Batch inferencing
⬜ D. Serverless inferencing with sub-second latency requirements

Explanation:
Batch inferencing processes large volumes of data together on a scheduled or on-demand basis when immediate, low-latency responses are not required, which makes it the most cost-effective choice for overnight, non-time-sensitive workloads like invoice processing.
Incorrect Options:
Option A: Real-time inferencing is optimized for low-latency, interactive responses (e.g., a live chat), which isn’t needed here and typically costs more to keep available.
Option B: Asynchronous inferencing is better suited for large individual payloads (such as large documents or videos) that need queuing but not necessarily a full batch job design.
Option D: There’s no stated sub-second latency requirement, so optimizing for it adds unnecessary cost.
Source: Deploy models for inference

---

A data science team is building a model to predict which of three shipping carriers will deliver a package fastest, based on structured features like distance, weight, and historical delivery times. Which type of ML technique is this?

⬜ A. Regression
✅ B. Classification
⬜ C. Clustering
⬜ D. Reinforcement learning

Explanation:
This is a classification problem because the model must assign each input to one of a fixed set of discrete categories (the three carriers), rather than predicting a continuous numeric value.
Incorrect Options:
Option A: Regression predicts a continuous numeric value (e.g., estimated delivery time in hours), not a category.
Option C: Clustering is an unsupervised technique for grouping similar data points without predefined labels; here the carriers are known, labeled categories.
Option D: Reinforcement learning involves an agent learning through trial-and-error rewards, which doesn’t match this supervised, labeled-outcome scenario.
Source: What is Machine Learning?

---

A company needs to add multi-language customer support to its application by translating incoming support tickets. Which AWS AI service is purpose-built for this use case with the LEAST implementation effort?

⬜ A. Amazon Comprehend
✅ B. Amazon Translate
⬜ C. Amazon Transcribe
⬜ D. Amazon Polly

Explanation:
Amazon Translate is a fully managed neural machine translation service purpose-built for translating text between languages, making it the best fit with the least implementation effort.
Incorrect Options:
Option A: Amazon Comprehend extracts insights like sentiment, entities, and key phrases from text; it does not translate languages.
Option C: Amazon Transcribe converts speech to text; it does not translate between languages.
Option D: Amazon Polly converts text to lifelike speech; it does not translate between languages.
Source: AWS AI Services

---

Which scenario is the LEAST appropriate use case for an AI/ML solution?

⬜ A. Forecasting next month’s product demand based on historical sales trends.
⬜ B. Detecting potentially fraudulent credit card transactions in near real time.
✅ C. Calculating the exact income tax owed by a taxpayer under a fixed, published tax formula.
⬜ D. Recommending products to a customer based on browsing history.

Explanation:
Calculating tax owed under a fixed, deterministic formula requires an exact, auditable, and repeatable outcome rather than a statistical prediction. This is best solved with rule-based logic/traditional software, not an AI/ML model, because ML models produce probabilistic predictions, not guaranteed exact answers.
Incorrect Options:
Option A, B, and D: These are all classic AI/ML use cases where the value comes from finding patterns in data (demand forecasting, anomaly detection, and recommendations) rather than needing one deterministic, rule-defined outcome.
Source: What is Artificial Intelligence (AI)?

---

A company is building a fraud-detection pipeline and wants to identify the AWS services and stages typically involved in an end-to-end AI/ML pipeline. Which sequence BEST reflects the stages of a typical ML pipeline?

⬜ A. Model deployment, then data collection, then model monitoring, then data labeling.
✅ B. Data collection and preparation, model building and training, model evaluation, deployment, and ongoing monitoring/retraining.
⬜ C. Model training only, since AWS managed services eliminate the need for data preparation or monitoring.
⬜ D. Data labeling, followed immediately by production deployment, with no evaluation step.

Explanation:
A typical AI/ML pipeline flows through data collection and preparation, model building and training, evaluation, deployment, and then ongoing monitoring and retraining as new data becomes available — these are the core stages AWS services like Amazon SageMaker AI support.
Incorrect Options:
Option A, C, and D: These either reorder the stages incorrectly or skip essential stages such as data preparation, evaluation, or monitoring, all of which are necessary for a production-quality ML pipeline.
Source: Machine Learning Operations (MLOps) on AWS

---

A company has deployed a fraud-detection model and wants to track ongoing model quality metrics like accuracy and precision in production, in addition to business metrics like cost per transaction reviewed. Which concept does this practice belong to?

⬜ A. Model Context Protocol (MCP)
✅ B. MLOps (ML Operations)
⬜ C. Retrieval Augmented Generation (RAG)
⬜ D. Prompt engineering

Explanation:
MLOps encompasses the practices of monitoring models in production, tracking performance and business metrics, and retraining models as needed to keep them accurate and reliable over time.
Incorrect Options:
Option A: MCP is a protocol for connecting AI agents and tools/data sources; it is unrelated to production model monitoring.
Option C: RAG is a technique for grounding generative AI outputs in external data; it doesn’t describe ongoing model-quality monitoring.
Option D: Prompt engineering is about crafting effective inputs for generative models, not monitoring deployed model performance.
Source: Machine Learning Operations (MLOps) on AWS

---

A team evaluating a binary fraud classifier wants a single metric that balances both precision (avoiding false fraud flags) and recall (catching actual fraud). Which metric should they use?

✅ A. F1 score
⬜ B. Mean squared error (MSE)
⬜ C. R-squared
⬜ D. BLEU score

Explanation:
The F1 score is the harmonic mean of precision and recall, making it the standard metric when a team needs to balance false positives and false negatives in a classification task.
Incorrect Options:
Option B and C: MSE and R-squared are regression metrics used to evaluate continuous numeric predictions, not classification outcomes.
Option D: BLEU score is used to evaluate the quality of machine-generated text (such as translation), not classification accuracy.
Source: Monitor model quality metrics

---

A company is deciding between building a custom ML model from scratch versus using a pre-trained foundation model (FM) for a new text-summarization feature. Which factor would MOST justify choosing a pre-trained FM over training a custom model from scratch?

⬜ A. The company has years of proprietary labeled data and unlimited budget for training infrastructure.
✅ B. The company needs a general-purpose capability quickly, without the time and cost of collecting massive training datasets and training a model from zero.
⬜ C. The company requires a model trained exclusively on its own private data, with no exposure to any external data.
⬜ D. The company wants the lowest possible per-inference cost regardless of development timeline.

Explanation:
Pre-trained foundation models are trained on massive general-purpose datasets and can be used or lightly adapted (via prompting or fine-tuning) quickly, avoiding the significant time, cost, and data requirements of training a model from scratch.
Incorrect Options:
Option A: Having abundant proprietary data and budget actually strengthens the case for a custom-trained model.
Option C: Needing a model trained exclusively on private data points toward custom training rather than a general pre-trained FM.
Option D: Per-inference cost is only one factor and doesn’t by itself justify choosing an FM over a custom model.
Source: What are Foundation Models?

---

Which term refers to the smallest units of text (such as words or sub-word pieces) that a large language model (LLM) processes and generates?

⬜ A. Embeddings
✅ B. Tokens
⬜ C. Vectors
⬜ D. Chunks

Explanation:
Tokens are the basic units of text (words, sub-words, or characters) that an LLM breaks input into and generates output as; token counts also directly drive Amazon Bedrock’s token-based pricing.
Incorrect Options:
Option A: Embeddings are numeric vector representations of meaning, not the raw text units themselves.
Option C: Vectors are the numeric arrays used to represent embeddings, not the text units.
Option D: Chunking refers to splitting larger documents into smaller pieces (often for RAG), which is a different concept from tokenization.
Source: What is a Large Language Model (LLM)?

---

A company wants to build a semantic search feature that finds documents with similar meaning to a user’s query, even when they don’t share the same exact keywords. Which underlying GenAI concept makes this possible?

⬜ A. Prompt templates
✅ B. Embeddings
⬜ C. Diffusion models
⬜ D. Token-based pricing

Explanation:
Embeddings are numeric vector representations of text (or other data) that capture semantic meaning, so texts with similar meaning end up close together in vector space, enabling semantic (meaning-based) search rather than simple keyword matching.
Incorrect Options:
Option A: Prompt templates are reusable prompt structures with placeholders; they don’t provide the vector-based similarity that semantic search relies on.
Option C: Diffusion models are primarily used for image/audio generation, not for semantic text similarity search.
Option D: Token-based pricing is a billing model, unrelated to how semantic search works.
Source: What are Embeddings in Machine Learning?

---

Which AWS service allows a company to access a variety of foundation models (FMs) from multiple providers through a single, unified API to build generative AI applications?

⬜ A. Amazon SageMaker Ground Truth
✅ B. Amazon Bedrock
⬜ C. AWS Glue
⬜ D. Amazon Kinesis

Explanation:
Amazon Bedrock is a fully managed service that provides access to foundation models from Amazon and third-party model providers through a single API, along with capabilities to customize and deploy generative AI applications.
Incorrect Options:
Option A: SageMaker Ground Truth is a data-labeling service, not a foundation-model access service.
Option C: AWS Glue is a data integration/ETL service, unrelated to foundation model access.
Option D: Amazon Kinesis is a service for real-time streaming data, not foundation model access.
Source: Amazon Bedrock

---

A generative AI application’s monthly cost varies significantly depending on the number of input and output tokens processed by the foundation model. Which pricing model does this describe?

⬜ A. Flat monthly subscription
✅ B. Token-based pricing
⬜ C. Per-user licensing
⬜ D. Free tier only, with no usage-based charges

Explanation:
Many foundation models on Amazon Bedrock use token-based pricing, where cost scales with the number of input and output tokens processed, directly affecting both cost and, indirectly, performance considerations for high-volume applications.
Incorrect Options:
Option A, C, and D: These billing models don’t tie cost directly to the volume of tokens processed by the model, which is the defining characteristic of the scenario described.
Source: Amazon Bedrock pricing

---

Which of the following is a well-known DISADVANTAGE of generative AI solutions that a company should account for before deploying a customer-facing chatbot?

⬜ A. Generative AI models are always fully deterministic and produce identical output every time for the same input.
✅ B. Generative AI models can hallucinate, producing plausible-sounding but factually incorrect or fabricated responses.
⬜ C. Generative AI models cannot generate any conversational responses.
⬜ D. Generative AI models require no evaluation before production deployment.

Explanation:
Hallucination — where a model generates confident, plausible-sounding, but incorrect or fabricated content — is a well-documented limitation of generative AI that must be mitigated (for example, with RAG or guardrails) before customer-facing use.
Incorrect Options:
Option A: GenAI models are often nondeterministic, especially at higher temperature settings, which is the opposite of this claim.
Option C: Conversational capability is actually a strength of GenAI, not a limitation.
Option D: Best practice explicitly requires evaluation before production deployment; skipping it is a risk, not a factual characteristic of the technology.
Source: What are Hallucinations in AI Models?

---

A company is comparing several foundation models for a customer-service application and wants to evaluate cross-model performance, return on investment (ROI), and customer satisfaction. Which category of consideration BEST describes this evaluation?

✅ A. Determining the business value and metrics of a GenAI application
⬜ B. Choosing an inference parameter such as temperature
⬜ C. Selecting a vector database for embeddings
⬜ D. Defining a prompt template

Explanation:
Evaluating ROI, cross-domain performance, and customer satisfaction/engagement are examples of determining the business value and metrics of a generative AI application — connecting technical performance to business outcomes.
Incorrect Options:
Option B: Temperature is a technical inference parameter that controls output randomness, not a business-value metric.
Option C: Vector database selection is an infrastructure/technical design decision for RAG, not a business-value evaluation.
Option D: Prompt templates are a prompt-engineering construct, unrelated to business ROI evaluation.
Source: What is Generative AI?

---

A company building an autonomous AI agent needs its agent to call external tools, maintain memory across steps, and communicate with other agents in a standardized way. Which foundational agentic AI concept addresses standardizing how an agent connects to external tools and data sources?

✅ A. Model Context Protocol (MCP)
⬜ B. Token-based pricing
⬜ C. Diffusion models
⬜ D. BLEU score

Explanation:
The Model Context Protocol (MCP) is an open standard that defines how AI agents/applications connect to external tools, data sources, and context in a consistent way, which is foundational to building interoperable agentic AI systems.
Incorrect Options:
Option B: Token-based pricing is a billing concept unrelated to tool/context standardization.
Option C: Diffusion models are a generative modeling technique for images/audio, not an agent-tool integration standard.
Option D: BLEU score is a text-generation evaluation metric, unrelated to agent tool connectivity.
Source: Amazon Bedrock

---

Which AWS capability provides curated, pre-built foundation models and solution templates that developers can quickly deploy within Amazon SageMaker AI to accelerate GenAI application development?

✅ A. SageMaker JumpStart
⬜ B. AWS CloudTrail
⬜ C. Amazon Macie
⬜ D. AWS Config

Explanation:
Amazon SageMaker JumpStart provides pre-built, pre-trained models (including foundation models) and solution templates that can be deployed with minimal effort, accelerating GenAI and ML application development.
Incorrect Options:
Option B: CloudTrail records account activity/API calls for auditing; it doesn’t provide pre-built models.
Option C: Macie discovers and protects sensitive data in S3; it doesn’t provide pre-built models.
Option D: AWS Config tracks resource configuration compliance; it doesn’t provide pre-built models.
Source: SageMaker JumpStart

---

A startup wants to launch a generative AI proof-of-concept quickly, with a low upfront cost and without managing GPU infrastructure. Which is an advantage of using AWS managed GenAI services like Amazon Bedrock for this use case?

⬜ A. It requires the company to build and train a foundation model from scratch.
✅ B. It lowers the barrier to entry by providing managed access to foundation models with pay-as-you-go pricing, without needing to manage the underlying infrastructure.
⬜ C. It eliminates the need for any prompt engineering or evaluation of model output.
⬜ D. It requires committing to a single foundation model provider with no flexibility to switch models.

Explanation:
A key advantage of AWS managed GenAI services like Amazon Bedrock is a lower barrier to entry: companies can access powerful foundation models through a managed API with usage-based pricing, without provisioning or managing GPU infrastructure themselves.
Incorrect Options:
Option A: Bedrock provides access to existing pre-trained FMs; it does not require building a model from scratch.
Option C: Prompt engineering and evaluation remain best practices regardless of which managed service is used.
Option D: Bedrock supports multiple model providers through one API, giving flexibility rather than locking a company into one provider.
Source: Amazon Bedrock

---

A company generating GenAI-produced marketing images at high volume wants to understand the cost tradeoffs of committing to provisioned throughput versus on-demand token-based pricing on Amazon Bedrock. Which factor is the PRIMARY consideration for choosing provisioned throughput?

⬜ A. Provisioned throughput is always cheaper than on-demand pricing regardless of usage volume.
✅ B. Provisioned throughput reserves dedicated model capacity for predictable, high-volume workloads, typically at a different cost structure than on-demand token pricing.
⬜ C. Provisioned throughput is required before any model can be invoked on Amazon Bedrock.
⬜ D. Provisioned throughput only applies to model fine-tuning jobs, not inference.

Explanation:
Provisioned throughput on Amazon Bedrock lets a customer reserve dedicated capacity for a model, which is a cost/performance tradeoff best suited to predictable, high-volume workloads that need consistent throughput, versus the flexible but variable on-demand token-based pricing model.
Incorrect Options:
Option A: Cost-effectiveness depends on usage patterns; provisioned throughput isn’t universally cheaper.
Option C: On-demand token-based invocation is available without provisioned throughput.
Option D: Provisioned throughput applies to inference capacity, not just fine-tuning.
Source: Amazon Bedrock pricing

---

Which statement BEST describes a foundation model (FM)?

⬜ A. A small, task-specific model trained only on one company’s internal dataset for a single narrow use case.
✅ B. A large model pre-trained on a broad range of data that can be adapted to many different downstream tasks through prompting or fine-tuning.
⬜ C. A rule-based expert system with no learned parameters.
⬜ D. A model that can only process image data, never text.

Explanation:
A foundation model is a large model pre-trained on a broad, diverse dataset, giving it general capabilities that can then be adapted to many different downstream tasks — such as summarization, Q&A, or code generation — via prompting, fine-tuning, or other customization techniques.
Incorrect Options:
Option A: This describes a narrow, task-specific model, which is the opposite of the broad adaptability that defines a foundation model.
Option C: FMs are learned, data-driven models, not rule-based expert systems.
Option D: FMs can be multi-modal, handling text, images, audio, and more, not just one modality.
Source: What are Foundation Models?

---

A company is evaluating whether generative AI is the right approach for a use case where they need a highly precise, exact numeric output every time (such as a legal filing deadline calculation), rather than a fluent but potentially varying response. Which GenAI limitation is MOST relevant here?

⬜ A. Adaptability
✅ B. Nondeterminism
⬜ C. Conversational capability
⬜ D. Content generation ability

Explanation:
Nondeterminism means a generative model can produce different outputs for the same input across runs, which makes it a poor fit for use cases that demand one exact, repeatable, precise answer every time.
Incorrect Options:
Option A, C, and D: Adaptability, conversational capability, and content-generation ability are advantages of GenAI, not limitations relevant to this precision-critical scenario.
Source: What is Generative AI?

---

A company wants its customer-support chatbot to answer questions using the most current internal knowledge-base articles, without retraining or fine-tuning the underlying foundation model every time an article changes. Which technique BEST meets this requirement?

✅ A. Retrieval Augmented Generation (RAG)
⬜ B. Full model pre-training from scratch
⬜ C. Increasing the temperature parameter
⬜ D. Reducing the maximum output token length

Explanation:
RAG retrieves relevant, up-to-date information from an external knowledge source (such as a knowledge base) at query time and provides it as context to the foundation model, so answers reflect current data without needing to retrain the model.
Incorrect Options:
Option B: Pre-training from scratch is extremely costly and impractical for keeping answers current with frequently changing articles.
Option C and D: Temperature and output length are inference parameters that control response style and size; they don’t give the model access to current external knowledge.
Source: Knowledge bases for Amazon Bedrock
Source: What is Retrieval Augmented Generation (RAG)?

---

Which AWS service allows a company to connect Amazon Bedrock foundation models to external data sources such as Amazon S3 for a Retrieval Augmented Generation (RAG) architecture, including automatically converting documents into embeddings?

✅ A. Knowledge Bases for Amazon Bedrock
⬜ B. AWS Direct Connect
⬜ C. Amazon CloudFront
⬜ D. AWS Snowball

Explanation:
Knowledge Bases for Amazon Bedrock is a fully managed capability that connects foundation models to a company’s data sources (like Amazon S3), automatically handling ingestion, chunking, and embedding generation to support RAG.
Incorrect Options:
Option B: AWS Direct Connect provides a dedicated network connection to AWS; it is unrelated to RAG or knowledge bases.
Option C: Amazon CloudFront is a content delivery network (CDN), unrelated to RAG.
Option D: AWS Snowball is a physical data-transfer device, unrelated to RAG.
Source: Knowledge bases for Amazon Bedrock

---

A company implementing a RAG solution on AWS needs to store vector embeddings so they can be efficiently searched for semantic similarity. Which of the following is a valid AWS option for storing and querying vector embeddings?

✅ A. Amazon OpenSearch Service
⬜ B. Amazon Simple Notification Service (SNS)
⬜ C. AWS Step Functions
⬜ D. Amazon Simple Queue Service (SQS)

Explanation:
Amazon OpenSearch Service supports vector search capabilities and is a supported option for storing and querying the vector embeddings used in RAG architectures (along with services like Amazon Aurora, Amazon Neptune, and Amazon RDS for PostgreSQL).
Incorrect Options:
Option B, C, and D: SNS, Step Functions, and SQS are messaging/orchestration services and are not vector database options for storing embeddings.
Source: Knowledge bases for Amazon Bedrock
Source: What are Vector Databases?

---

A developer wants a foundation model’s responses to be MORE deterministic and focused, consistently choosing the most likely next token rather than exploring less probable alternatives. Which inference parameter adjustment achieves this?

⬜ A. Increase the temperature value.
✅ B. Decrease the temperature value.
⬜ C. Increase the maximum output token length.
⬜ D. Enable a higher top-p value only.

Explanation:
Lowering the temperature parameter makes the model’s output distribution sharper, so it favors higher-probability tokens and produces more deterministic, focused, and consistent responses.
Incorrect Options:
Option A: Increasing temperature increases randomness and creativity, the opposite of the desired effect.
Option C: Maximum output token length controls response length, not determinism.
Option D: Increasing top-p widens the pool of candidate tokens considered, increasing variability rather than reducing it.
Source: Inference parameters for foundation models

---

A prompt engineer provides a large language model with several example input-output pairs directly within the prompt before asking it to complete a new, similar task. Which prompt engineering technique is being used?

⬜ A. Zero-shot prompting
✅ B. Few-shot prompting
⬜ C. Retrieval Augmented Generation (RAG)
⬜ D. Fine-tuning

Explanation:
Few-shot prompting (also called in-context learning) provides the model with a small number of example input-output pairs within the prompt itself, helping calibrate the model’s response format and behavior for the new task without any additional training.
Incorrect Options:
Option A: Zero-shot prompting gives the model instructions with no examples at all.
Option C: RAG retrieves external documents/context, which is a different technique from providing example pairs directly in the prompt.
Option D: Fine-tuning involves further training the model’s weights on a dataset; it is not a prompt-only technique.
Source: Prompt engineering concepts

---

Which of the following is a well-documented RISK of prompt engineering that a company should guard against when exposing a chatbot to public users?

✅ A. Prompt hijacking, where a malicious user manipulates the model into ignoring its original instructions.
⬜ B. Improved response quality from well-structured prompts.
⬜ C. Faster experimentation and iteration on prompt design.
⬜ D. Clearer, more specific outputs due to well-crafted instructions.

Explanation:
Prompt hijacking (also related to jailbreaking) is a risk where a malicious actor crafts input designed to override or bypass the model’s original system instructions, potentially causing it to produce unintended or harmful output.
Incorrect Options:
Option B, C, and D: These describe benefits of good prompt engineering practice, not risks.
Source: Prompt engineering concepts

---

A company wants to save, version, and reuse a library of tested prompts across multiple generative AI applications and team members on Amazon Bedrock. Which capability should they use?

✅ A. Amazon Bedrock Prompt Management
⬜ B. Amazon Bedrock Guardrails
⬜ C. AWS Config
⬜ D. Amazon Inspector

Explanation:
Amazon Bedrock Prompt Management provides a way to create, test, version, and share prompts, making it easier to maintain a reusable and consistent prompt library across an organization’s generative AI applications.
Incorrect Options:
Option B: Bedrock Guardrails is for implementing content-safety policies, not prompt versioning/reuse.
Option C: AWS Config tracks resource configuration compliance, unrelated to prompt libraries.
Option D: Amazon Inspector performs vulnerability scanning, unrelated to prompt management.
Source: Prompt management in Amazon Bedrock

---

A company wants to adapt a general-purpose foundation model to understand company-specific terminology and respond in a particular tone, by further training the model on a smaller, labeled, domain-specific dataset. Which FM customization approach is this?

✅ A. Fine-tuning
⬜ B. Retrieval Augmented Generation (RAG)
⬜ C. Prompt engineering with zero-shot prompts
⬜ D. Increasing the temperature parameter only

Explanation:
Fine-tuning further trains a pre-trained foundation model on a smaller, labeled, domain-specific dataset, adjusting the model’s weights so it better reflects specific terminology, tone, or task requirements.
Incorrect Options:
Option B: RAG retrieves external context at inference time without modifying the model’s weights.
Option C: Zero-shot prompting relies solely on instructions with no additional training or examples.
Option D: Temperature is an inference-time parameter and does not adapt the model’s underlying knowledge.
Source: Custom models in Amazon Bedrock

---

Which customization approach generally has the LOWEST upfront cost and fastest time-to-value, since it requires no additional model training?

⬜ A. Pre-training a foundation model from scratch
⬜ B. Fine-tuning with a large labeled dataset
✅ C. In-context learning through prompt engineering (zero-shot/few-shot)
⬜ D. Continuous pre-training on a domain corpus

Explanation:
In-context learning (zero-shot or few-shot prompting) uses only prompt design to guide the model’s behavior, requiring no additional training, which makes it the fastest and least expensive customization approach to implement.
Incorrect Options:
Option A: Pre-training a model from scratch is by far the most expensive and time-consuming approach.
Option B and D: Fine-tuning and continuous pre-training both require additional training compute, curated datasets, and more time than prompting alone.
Source: Custom models in Amazon Bedrock

---

A company wants to improve a foundation model’s responses using human feedback about which of several generated responses is preferred, and use that feedback signal to further align the model’s behavior. Which technique does this describe?

✅ A. Reinforcement learning from human feedback (RLHF)
⬜ B. Zero-shot prompting
⬜ C. Vector embedding generation
⬜ D. Model distillation

Explanation:
RLHF uses human evaluators to rank or rate model outputs, and that feedback is used as a reward signal to further train and align the model’s behavior with human preferences.
Incorrect Options:
Option B: Zero-shot prompting is a prompting technique, not a training/feedback-based alignment method.
Option C: Vector embedding generation converts data into numeric representations; it doesn’t involve human preference feedback.
Option D: Model distillation transfers knowledge from a larger model to a smaller one; it does not inherently involve human feedback ranking.
Source: What is Reinforcement Learning from Human Feedback (RLHF)?

---

A company wants a smaller, faster, and cheaper model that mimics the behavior of a larger, more capable foundation model for a narrower production use case. Which technique BEST achieves this?

✅ A. Model distillation
⬜ B. Retrieval Augmented Generation (RAG)
⬜ C. Prompt hijacking
⬜ D. Increasing the context window size

Explanation:
Model distillation trains a smaller ‘student’ model to replicate the behavior of a larger ’teacher’ model, producing a more compact, faster, and cheaper model suitable for specific production use cases.
Incorrect Options:
Option B: RAG improves grounding with external data; it does not create a smaller model.
Option C: Prompt hijacking is a security risk, not a customization technique.
Option D: Increasing context window size affects how much input a model can process, not model size/cost.
Source: Custom models in Amazon Bedrock

---

A team wants to evaluate the quality of a foundation model’s generated text summaries against human-written reference summaries using an automated, well-established metric. Which metric is commonly used for this purpose?

✅ A. ROUGE
⬜ B. F1 score
⬜ C. Precision and recall only
⬜ D. Mean squared error (MSE)

Explanation:
ROUGE (Recall-Oriented Understudy for Gisting Evaluation) is a widely used automated metric for evaluating the quality of generated text, such as summaries, against human-written reference text.
Incorrect Options:
Option B, C, and D: F1 score, precision/recall, and MSE are classification and regression metrics; they are not the standard automated metrics for evaluating generated text quality against references.
Source: Model evaluation in Amazon Bedrock

---

A company wants to run a structured evaluation job on Amazon Bedrock that compares multiple foundation models against built-in metrics and, optionally, human review. Which Amazon Bedrock capability supports this?

✅ A. Amazon Bedrock Model Evaluation
⬜ B. Amazon Bedrock Guardrails
⬜ C. AWS Trusted Advisor
⬜ D. Amazon Macie

Explanation:
Amazon Bedrock Model Evaluation lets a company run evaluation jobs that assess and compare foundation models using built-in automatic metrics as well as human-based evaluation workflows.
Incorrect Options:
Option B: Guardrails enforces content-safety policies, not model performance evaluation.
Option C: Trusted Advisor gives account best-practice recommendations, unrelated to model evaluation.
Option D: Macie is for discovering sensitive data in S3, unrelated to model evaluation.
Source: Model evaluation in Amazon Bedrock

---

A company deployed a RAG-based question-answering agent and wants to measure whether end users actually complete their intended task and are satisfied with the answers, beyond just automated text-quality metrics. Which type of metric BEST addresses this?

✅ A. Business objective alignment metrics, such as task completion rate and user satisfaction.
⬜ B. BLEU score only
⬜ C. ROUGE score only
⬜ D. Token count per request

Explanation:
Business objective alignment metrics — such as task completion rate, user satisfaction, and cost per interaction — measure whether an AI application built on foundation models (including RAG and agents) is actually delivering value in the real world, beyond automated text-similarity scores.
Incorrect Options:
Option B and C: BLEU and ROUGE measure text similarity to reference outputs, not real-world task completion or user satisfaction.
Option D: Token count relates to cost/usage, not whether the application is meeting business objectives.
Source: Model evaluation in Amazon Bedrock

---

A company wants to detect and filter harmful, biased, or off-topic content generated by its foundation model before it reaches end users. Which Amazon Bedrock capability is purpose-built for this?

✅ A. Amazon Bedrock Guardrails
⬜ B. Amazon Bedrock Prompt Management
⬜ C. Amazon SageMaker JumpStart
⬜ D. AWS Config

Explanation:
Amazon Bedrock Guardrails lets a company implement safeguards, such as content filters and denied topics, to help detect and block harmful, biased, or otherwise undesired model outputs before they reach end users.
Incorrect Options:
Option B: Prompt Management is for organizing and versioning prompts, not filtering harmful content.
Option C: SageMaker JumpStart provides pre-built models and templates; it doesn’t filter model output for harmful content.
Option D: AWS Config tracks resource configuration compliance and is unrelated to content filtering.
Source: Amazon Bedrock Guardrails

---

A company training a hiring-recommendation model discovers that its training dataset contains far more examples from one demographic group than others. What is the PRIMARY risk this creates?

⬜ A. The model will train faster because the dataset is smaller.
✅ B. The model may produce biased predictions that unfairly disadvantage underrepresented groups.
⬜ C. The model will automatically balance the dataset during training with no risk.
⬜ D. The model will be immune to overfitting.

Explanation:
An imbalanced or non-representative training dataset is a primary source of bias, which can cause a model to make unfair or inaccurate predictions for underrepresented demographic groups — a key responsible-AI concern to detect and mitigate before deployment.
Incorrect Options:
Option A, C, and D: Dataset size affecting training speed, automatic self-balancing, and immunity to overfitting are not accurate characterizations of what an imbalanced dataset causes.
Source: Responsible AI

---

Which AWS capability helps document a machine learning model’s intended use, training details, risk ratings, and evaluation results in a single, structured, shareable record to support transparency and governance?

✅ A. Amazon SageMaker Model Cards
⬜ B. AWS Direct Connect
⬜ C. Amazon Kinesis Data Streams
⬜ D. AWS Snowmobile

Explanation:
Amazon SageMaker Model Cards provide a structured way to document essential model details — including intended uses, risk ratings, training data, and evaluation results — supporting transparency, governance, and responsible AI practices.
Incorrect Options:
Option B, C, and D: Direct Connect, Kinesis Data Streams, and Snowmobile are networking, streaming, and data-transfer services, respectively, with no relation to model documentation.
Source: Amazon SageMaker Model Cards

---

A company is evaluating the legal risks of deploying a generative AI application that produces marketing copy and images. Which of the following is a recognized legal risk category associated with generative AI?

⬜ A. Guaranteed accuracy of all generated content
✅ B. Intellectual property (IP) infringement claims from content that resembles copyrighted material
⬜ C. Elimination of all end-user risk
⬜ D. Complete removal of the need for human review

Explanation:
Generative AI models can produce content that closely resembles existing copyrighted works, creating a real risk of intellectual property infringement claims — one of the key legal risks companies must evaluate and mitigate (for example, through licensing, guardrails, and human review) before deployment.
Incorrect Options:
Option A, C, and D: These describe the opposite of actual risk — GenAI does not guarantee accuracy, does not eliminate end-user risk, and does not remove the need for human oversight.
Source: Responsible AI

---

Which AWS capability can help a data science team analyze training data and model predictions for potential bias across different demographic groups, both before and after training?

✅ A. Amazon SageMaker Clarify
⬜ B. AWS Snowball Edge
⬜ C. Amazon CloudFront
⬜ D. AWS Direct Connect

Explanation:
Amazon SageMaker Clarify provides tools to help detect potential bias in datasets and trained models across different groups, as well as to help explain model predictions, supporting fairness and explainability goals.
Incorrect Options:
Option B, C, and D: Snowball Edge, CloudFront, and Direct Connect are data-transfer, content-delivery, and networking services, and are unrelated to bias detection.
Source: Fairness and Explainability with SageMaker Clarify

---

A company wants end users to understand, in general terms, why an AI system produced a particular recommendation, and wants to incorporate a feedback mechanism so users can flag incorrect or concerning outputs. Which responsible-AI principle does this BEST reflect?

✅ A. Human-centered design for explainable AI
⬜ B. Maximizing model throughput
⬜ C. Minimizing infrastructure cost
⬜ D. Reducing the number of foundation model parameters

Explanation:
Human-centered design for explainable AI focuses on giving users appropriate visibility into how and why an AI system reached a decision, along with mechanisms like user feedback, to build trust and support responsible use.
Incorrect Options:
Option B, C, and D: Throughput, infrastructure cost, and model size are technical/operational concerns, not the responsible-AI principle of explainability and user-centered feedback described here.
Source: Responsible AI

---

A company must choose between a highly transparent, simpler model and a more complex, higher-performing model that is harder to interpret. Which concept describes this decision?

✅ A. The tradeoff between model interpretability and performance.
⬜ B. The tradeoff between AWS Regions and Availability Zones.
⬜ C. The tradeoff between synchronous and asynchronous networking.
⬜ D. The tradeoff between block storage and object storage.

Explanation:
This scenario describes the classic tradeoff between model interpretability (how easily humans can understand how the model works) and performance (predictive accuracy or capability) — more complex models are often more capable but harder to explain, and vice versa.
Incorrect Options:
Option B, C, and D: These describe infrastructure and storage tradeoffs, unrelated to the model transparency vs. performance decision described.
Source: Responsible AI

---

Under the AWS Shared Responsibility Model, who is responsible for securing an AI/ML workload’s data and configuring appropriate access controls (such as IAM policies) for resources like Amazon Bedrock or Amazon SageMaker AI?

⬜ A. AWS is solely responsible for all security, with no customer responsibility.
✅ B. The customer is responsible for security ‘in’ the cloud, including data protection and access management, while AWS is responsible for security ‘of’ the cloud infrastructure.
⬜ C. Security responsibility is randomly assigned per project.
⬜ D. Only AWS Support is responsible for customer data security.

Explanation:
Under the AWS Shared Responsibility Model, AWS is responsible for the security ‘of’ the cloud (the underlying infrastructure), while the customer is responsible for security ‘in’ the cloud — including their data, identity and access management, and configuration of the AI/ML services they use.
Incorrect Options:
Option A, C, and D: These misstate the shared responsibility model; both AWS and the customer have defined, non-random responsibilities, and customer data protection is the customer’s responsibility, not AWS Support’s.
Source: AWS Shared Responsibility Model

---

A company wants to grant fine-grained permissions so that only specific users can invoke certain Amazon Bedrock foundation models, while denying access to others. Which AWS service should be used to define and enforce these permissions?

✅ A. AWS Identity and Access Management (IAM)
⬜ B. Amazon Macie
⬜ C. Amazon Inspector
⬜ D. AWS Artifact

Explanation:
AWS IAM lets a company define fine-grained policies and permissions to control which users or roles can access specific AWS resources and actions, including invoking specific foundation models on Amazon Bedrock.
Incorrect Options:
Option B: Macie discovers and protects sensitive data (like PII) in S3; it does not manage access permissions.
Option C: Inspector performs automated vulnerability scanning; it does not manage fine-grained access permissions.
Option D: AWS Artifact provides access to AWS compliance reports and agreements; it does not manage IAM permissions.
Source: What is IAM?

---

A company storing training data in Amazon S3 wants to automatically discover and alert on personally identifiable information (PII) present in its data before it is used to train or fine-tune a model. Which AWS service is purpose-built for this?

✅ A. Amazon Macie
⬜ B. AWS Config
⬜ C. Amazon Inspector
⬜ D. AWS Trusted Advisor

Explanation:
Amazon Macie uses machine learning and pattern matching to automatically discover, classify, and alert on sensitive data such as PII stored in Amazon S3, which is important for data governance before that data is used in AI/ML training.
Incorrect Options:
Option B: AWS Config tracks and evaluates resource configuration compliance, not PII discovery in data.
Option C: Amazon Inspector scans for software vulnerabilities and unintended network exposure, not PII in datasets.
Option D: AWS Trusted Advisor provides account best-practice checks (cost, security, performance), not dedicated PII discovery.
Source: What is Amazon Macie?

---

A company wants to invoke Amazon Bedrock APIs from within its Amazon VPC without traversing the public internet, to reduce exposure of sensitive AI workload traffic. Which AWS capability should be used?

✅ A. AWS PrivateLink
⬜ B. Amazon Route 53 public hosted zones
⬜ C. An internet gateway
⬜ D. AWS Direct Connect only, with a NAT gateway

Explanation:
AWS PrivateLink allows private connectivity between a VPC and supported AWS services (including Amazon Bedrock) without the traffic traversing the public internet, reducing exposure of sensitive AI workload data.
Incorrect Options:
Option B: Public hosted zones are for public DNS resolution and do not provide private connectivity.
Option C: An internet gateway is what enables (rather than avoids) internet-routed traffic.
Option D: Direct Connect provides a dedicated network link to AWS but is a different, heavier-weight solution than PrivateLink for private service access, and isn’t required for this use case.
Source: What is AWS PrivateLink?

---

A company must track which datasets and configurations were used to produce a particular version of a deployed model, to satisfy an audit request. Which SageMaker capability BEST supports this requirement?

✅ A. SageMaker Lineage Tracking
⬜ B. SageMaker JumpStart
⬜ C. SageMaker Ground Truth
⬜ D. SageMaker Studio Notebooks only

Explanation:
Amazon SageMaker Lineage Tracking captures information about the steps of an ML workflow, including datasets, code, and configurations used, making it possible to trace and audit how a specific model version was produced — supporting data governance and source citation requirements.
Incorrect Options:
Option B: JumpStart provides pre-built models/templates; it doesn’t track workflow lineage.
Option C: Ground Truth is a data-labeling service, not a lineage-tracking capability.
Option D: Studio Notebooks are a development environment; on their own they don’t provide structured lineage tracking.
Source: Lineage Tracking in Amazon SageMaker

---

A company needs to continuously evaluate whether its AWS resources supporting an AI workload comply with internal security and configuration policies, and be alerted when a resource drifts out of compliance. Which AWS service is designed for this?

✅ A. AWS Config
⬜ B. Amazon Polly
⬜ C. Amazon Lex
⬜ D. Amazon Comprehend

Explanation:
AWS Config continuously monitors and records AWS resource configurations and can evaluate them against desired configuration rules, alerting when resources drift out of compliance — a key governance capability for AI workloads.
Incorrect Options:
Option B, C, and D: Polly (text-to-speech), Lex (conversational interfaces), and Comprehend (text analysis) are AI services unrelated to configuration compliance monitoring.
Source: What Is AWS Config?

---

A company wants a complete, immutable audit trail of every API call made against its Amazon Bedrock and Amazon SageMaker AI resources, including who made the call and when, to support a compliance review. Which AWS service provides this?

✅ A. AWS CloudTrail
⬜ B. AWS Trusted Advisor
⬜ C. AWS Artifact
⬜ D. Amazon Inspector

Explanation:
AWS CloudTrail records detailed API call activity across AWS services, including who made a call, when, and from where, providing the audit trail needed for governance, security investigation, and compliance reviews of AI workloads.
Incorrect Options:
Option B: Trusted Advisor provides best-practice recommendations, not a detailed API call audit trail.
Option C: AWS Artifact provides access to AWS’s own compliance reports and agreements, not a customer account’s API activity log.
Option D: Amazon Inspector performs vulnerability assessments, not API call auditing.
Source: AWS CloudTrail User Guide