AWS Certified AI Practitioner (AIF-C01) — All Questions

Why is encrypting data both at rest and in transit important for AI workloads?

• a.It completely removes the need to configure or manage any access permissions on the resources
• b.It permanently increases the model's prediction accuracy
• c.It keeps sensitive training data and model inputs unreadable if they are intercepted✓
• d.It is only ever needed for public data anyone may already read

Encryption at rest and in transit safeguards confidentiality: even if storage or network traffic is accessed by an unauthorized party, the data stays unreadable. It complements, not replaces, access controls.

Which service continuously assesses the configuration of AWS resources against desired policies for compliance?

• a.Amazon Polly, which converts written input text into natural-sounding spoken audio output
• b.Amazon Lex, which builds conversational chatbots using speech and language understanding
• c.AWS Config, which records configurations and evaluates them against defined rules✓
• d.Amazon Rekognition, which detects objects, scenes, and faces within images and in video

AWS Config records resource configurations and evaluates them against rules for compliance and governance, alerting on drift. Lex, Rekognition, and Polly are AI application services, not governance tools.

Which service automatically discovers and helps protect sensitive data such as personal information stored in Amazon S3?

• a.Amazon Macie, which finds and classifies sensitive data like personal information✓
• b.Amazon Translate, which converts written text from one human language into another one
• c.Amazon Transcribe, which converts speech into written text
• d.Amazon Forecast, which predicts future values from time-series datasets

Amazon Macie uses machine learning to discover, classify, and help protect sensitive data like PII in S3, supporting data-privacy and governance goals. The other services translate, transcribe, or forecast.

When using data to train or customize an AI model, why does data governance emphasize provenance and quality?

• a.Because governance rules require deleting all data-source records
• b.Because tracking provenance is what actually lets the finished model run on much cheaper hardware
• c.Because the quality of the data only matters after the model is retired
• d.Because the origin and accuracy of training data directly affect reliability and compliance✓

Knowing where data came from and ensuring it is accurate, permitted, and representative is central to trustworthy models: poor or improperly sourced data can create bias, errors, and legal exposure. Governance documents these sources rather than erasing them.

Which AWS service records API calls and user activity to provide an audit trail for security and governance?

• a.AWS CloudTrail, which logs API calls and account activity as an audit trail✓
• b.Amazon Polly, which converts written input text into natural-sounding spoken audio
• c.Amazon Kendra, which provides intelligent search across enterprise content sources
• d.Amazon Textract, which extracts text, forms, and tables from scanned document files

AWS CloudTrail logs API calls and account activity, creating an audit trail for security analysis, compliance, and investigating who did what and when. The other options are AI application services, not audit tools.

A robot learns to navigate a warehouse by trying actions and receiving rewards for reaching goals and penalties for collisions. Which learning paradigm is this?

• a.Supervised learning, because a fully labeled dataset of the correct navigation moves was provided in advance
• b.Transfer learning, because the robot simply reuses a model trained on a different problem
• c.Unsupervised learning, because the robot groups similar warehouse locations on its own
• d.Reinforcement learning, where an agent learns from rewards and penalties to maximize long-term reward✓

Reinforcement learning trains an agent through trial and error: it takes actions, receives rewards or penalties, and learns a policy that maximizes cumulative reward. No labeled dataset of correct moves is provided, which rules out supervised learning.

A model performs poorly on BOTH its training data and new test data. What does this most likely indicate?

• a.Overfitting, because the model simply memorized the training data far too closely
• b.Convergence, because the model has reached its optimal solution
• c.Data leakage between the training and test sets
• d.Underfitting, because the model is too simple to capture the underlying pattern✓

Underfitting means the model is too simple or undertrained to capture the data's patterns, so it performs poorly even on training data. Overfitting is the opposite: strong on training data but weak on unseen data.

Which statement best describes structured data?

• a.Data organized in a fixed schema of rows and columns, such as a relational database table✓
• b.Free-form text documents, images, and audio recordings that have no predefined fixed format
• c.Data that always requires a deep neural network in order to be processed
• d.Data that has already been split into separate training and test sets

Structured data fits a predefined schema of rows and columns (spreadsheets, SQL tables). Unstructured data such as text, images, and audio has no fixed format and often needs different techniques to process.

In supervised learning, what is 'labeled data'?

• a.Data that has been encrypted before it is used for training
• b.Data that has had all personally identifiable information removed from it
• c.Data in which each example includes the correct target answer the model should learn to predict✓
• d.Data that has no known outputs and consists only of raw inputs

Labeled data pairs each input with its known correct output (the label), which supervised models learn from. Data with only inputs and no known outputs is unlabeled and is used in unsupervised learning.

Which sequence best represents a typical machine learning workflow?

• a.Deployment, then training, then data collection, then evaluation
• b.Model training, then data cleaning, then problem definition
• c.Data collection and preparation, then model training, then evaluation, then deployment and monitoring✓
• d.Inference first, then labeling, then feature selection

A standard ML pipeline flows from collecting and preparing data, to training a model, to evaluating it, and finally deploying and monitoring it in production. The other orderings scramble these dependent stages.

What is a hyperparameter in machine learning?

• a.A value the model learns automatically from the data during training
• b.A configuration value, such as learning rate, that is set before training and is not learned by the model itself✓
• c.The final prediction the model outputs at inference time
• d.A single column of input data fed to the model

Hyperparameters (learning rate, number of layers, etc.) are set before training and control how learning happens. Parameters such as weights are what the model actually learns from data during training.

Which tool summarizes how many predictions a classifier got right and wrong for each class?

• a.A learning-rate schedule
• b.An embedding vector
• c.A loss function
• d.A confusion matrix✓

A confusion matrix lays out true positives, false positives, true negatives, and false negatives per class, making classification performance easy to inspect. A loss function guides training but does not summarize per-class results this way.

In classification, what does 'recall' measure?

• a.The fraction of the model's positive predictions that turned out to be correct
• b.Of all the actual positive cases, the fraction the model correctly identified✓
• c.The total time the model takes to make a single prediction
• d.The total number of parameters contained in the model

Recall (sensitivity) is true positives divided by all actual positives, i.e., how many real positives the model caught. The fraction of positive predictions that are correct describes precision, a different metric.

Predicting whether a tumor is 'malignant' or 'benign' is an example of what?

• a.Regression, predicting a continuous numeric value
• b.Binary classification, choosing between two discrete classes✓
• c.Clustering, grouping unlabeled data into natural sets
• d.Dimensionality reduction, compressing many features into fewer

Choosing between exactly two categories is binary classification. Regression predicts continuous numbers, and clustering groups unlabeled data without predefined categories.

Automatically spotting defective products on an assembly line from camera images is an example of which AI capability?

• a.Natural language processing
• b.Speech recognition
• c.Computer vision✓
• d.Time-series forecasting

Analyzing images or video to detect objects, defects, or scenes is computer vision. NLP handles text, speech recognition converts audio to text, and forecasting predicts future values from time-series data.

Which field of AI focuses on enabling machines to understand and generate human language?

• a.Computer vision
• b.Reinforcement learning
• c.Natural language processing (NLP)✓
• d.Robotic process automation software

Natural language processing (NLP) covers understanding and generating human language, for example translation, sentiment analysis, and chatbots. Computer vision instead deals with images and video.

What is the primary purpose of a validation set that is kept separate from the training and test sets?

• a.To tune hyperparameters and compare models during development without touching the final test set✓
• b.To serve production traffic after the model is deployed
• c.To store the model's encrypted backups
• d.To train the final model weights

The validation set guides model selection and hyperparameter tuning during development, keeping the test set untouched for an unbiased final estimate. Training happens on the training set, not the validation set.

A deployed model's accuracy slowly declines over months as customer behavior changes. What is the best response?

• a.Monitor for this drift and retrain the model on more recent, representative data✓
• b.Delete all logs so the decline is never recorded anywhere
• c.Increase the model's temperature parameter to compensate
• d.Assume the deployed model is fine and stop measuring its accuracy entirely from now on

Real-world data shifts over time (model or data drift), degrading performance. The responsible fix is ongoing monitoring plus retraining on fresh data. Ignoring it or hiding the evidence increases risk.

What is feature engineering?

• a.Encrypting the entire dataset before any training begins
• b.Deploying the finished trained model into production
• c.Transforming or creating input variables from raw data to improve model performance✓
• d.Choosing which AWS Region the training job runs in

Feature engineering shapes raw data into informative input variables (features), for example combining columns or extracting date parts, which can significantly improve model performance. It is a data-preparation activity, not deployment.

Which Amazon SageMaker capability lets business analysts build ML models through a no-code visual interface?

• a.Amazon SageMaker Clarify
• b.Amazon SageMaker Model Monitor
• c.Amazon SageMaker Ground Truth
• d.Amazon SageMaker Canvas✓

Amazon SageMaker Canvas provides a no-code, point-and-click interface so business users can build and use ML models without writing code. Clarify detects bias, Ground Truth labels data, and Model Monitor watches deployed models.

Which Amazon SageMaker capability offers pre-trained models and prebuilt solution templates to help you get started quickly?

• a.Amazon SageMaker Data Wrangler
• b.Amazon SageMaker Model Monitor
• c.Amazon SageMaker JumpStart✓
• d.Amazon SageMaker Clarify

Amazon SageMaker JumpStart provides pre-trained models, foundation models, and prebuilt solution templates for a fast start. Data Wrangler prepares data, Clarify checks bias, and Model Monitor tracks deployed models.

Why is the phrase 'garbage in, garbage out' emphasized in machine learning?

• a.A large enough model automatically makes the quality of its training data completely irrelevant
• b.Models automatically clean any bad data they are given
• c.Data quality only starts to matter after a model has been retired
• d.Poor-quality or biased training data produces unreliable, poor-quality model predictions✓

A model can only learn from the data it is given; flawed, biased, or noisy data leads to flawed predictions no matter how advanced the model. Data quality is foundational, not something the model fixes on its own.

Suggesting products a customer is likely to buy based on their past behavior is an example of which ML application?

• a.Anomaly detection
• b.Optical character recognition
• c.A recommendation system✓
• d.Speech synthesis

Predicting items a user will like from past behavior is a recommendation system. Anomaly detection flags unusual data points, OCR reads text from images, and speech synthesis generates audio.

A bank wants to flag transactions that deviate sharply from a customer's normal pattern, without any predefined fraud labels. Which technique fits best?

• a.Supervised regression on labeled fraud dollar amounts
• b.Image classification of the customer's photos
• c.Text summarization of the transaction descriptions
• d.Anomaly detection, an unsupervised approach that flags unusual data points✓

With no labeled fraud examples, anomaly detection (typically unsupervised) identifies data points that differ markedly from normal patterns. Supervised regression would require labeled target values that are unavailable here.

What does a large language model's 'context window' refer to?

• a.The maximum amount of text, measured in tokens, the model can consider at once, including prompt and response✓
• b.The physical AWS Region where the model happens to be hosted
• c.The set of safety filters that are applied to the model's output
• d.The graphical console window used to chat with the model

The context window is the token budget the model can attend to in a single request (input plus output). Exceeding it forces truncation. It is about capacity, not hosting location or safety filtering.

The Top P (nucleus sampling) inference parameter controls what?

• a.The maximum number of tokens allowed in the response
• b.The AWS Region that is used to run inference
• c.Sampling from the smallest set of tokens whose cumulative probability reaches a threshold, affecting output diversity✓
• d.Whether the model's weights are updated during the request

Top P restricts sampling to the most probable tokens that together reach a cumulative probability p, balancing diversity and focus. It never updates weights and is separate from the max-tokens length limit.

A team wants to generate original images from written text descriptions. Which type of foundation model do they need?

• a.A text-embeddings model
• b.A speech-to-text model
• c.A tabular regression model
• d.A text-to-image generation model, such as Amazon Titan Image Generator or Stable Diffusion on Amazon Bedrock✓

Creating images from text prompts requires a text-to-image generative model (for example Titan Image Generator or Stable Diffusion). Embeddings models output vectors and speech-to-text handles audio, not image creation.

What is a multimodal foundation model?

• a.A model that can process only plain text
• b.A model that is jointly owned by several companies
• c.A model that is hosted in multiple AWS Regions simultaneously
• d.A model that can process and/or generate more than one type of data, such as text and images together✓

Multimodal models work across multiple data types (for example accepting an image plus a text question). A text-only model is unimodal. 'Multimodal' describes the data types handled, not hosting or ownership.

Which AWS service is a generative AI assistant that answers questions and generates content grounded in a company's own enterprise data and systems?

• a.Amazon Q Business✓
• b.Amazon Rekognition
• c.Amazon Q Developer
• d.Amazon Comprehend

Amazon Q Business is an enterprise generative AI assistant that connects to company data sources to answer questions and generate content. Amazon Q Developer targets coding and AWS development tasks instead.

In Amazon Bedrock, what is 'Amazon Titan'?

• a.Amazon's own family of foundation models available through Amazon Bedrock✓
• b.A managed vector database service
• c.A no-code model-building tool for business analysts
• d.A hardware chip used for training large models

Amazon Titan is Amazon's family of first-party foundation models (for text, embeddings, and images) offered via Bedrock. Amazon's training and inference chips are Trainium and Inferentia, which are separate products.

Why might a large language model be unable to correctly answer a question about an event that happened last week?

• a.The model automatically refuses any question that mentions a date
• b.The temperature parameter was set too low for recent events
• c.The prompt was encrypted before it reached the model
• d.Its knowledge is limited to its training-data cutoff, and it has no inherent access to real-time information✓

An LLM only knows what was in its training data up to a cutoff and cannot see live events unless connected to a tool or retrieval system. This is why techniques like RAG are used to supply current information.

Condensing a long report into a short paragraph of its key points is which generative AI task?

• a.Classification
• b.Clustering
• c.Text summarization✓
• d.Named-entity extraction

Producing a shorter version that preserves the key points is text summarization, a common generative AI use case. Classification assigns labels and clustering groups data, neither of which rewrites the content into a summary.

A team needs to turn text into numeric vectors for a semantic search feature. Which type of model should they use?

• a.An embeddings model, such as Amazon Titan Text Embeddings✓
• b.A text-to-image generation model
• c.A speech synthesis (text-to-speech) voice generation model
• d.A chat model tuned for long conversations

Semantic search relies on embeddings that place similar meanings near each other in vector space, so an embeddings model is the right tool. Chat or image models generate content rather than produce search vectors.

Modern large language models are built primarily on which neural network architecture?

• a.Convolutional neural networks
• b.The transformer architecture✓
• c.Decision trees
• d.Linear regression

Transformers, which use self-attention, are the foundation of modern LLMs and most generative models. CNNs are common in classic computer vision, while decision trees and linear regression are traditional ML methods.

Which offering lets people experiment with and build generative AI apps in a no-code playground based on Amazon Bedrock?

• a.Amazon SageMaker Ground Truth data labeling
• b.Amazon Macie
• c.AWS CloudTrail
• d.PartyRock, an Amazon Bedrock Playground✓

PartyRock is a no-code, hands-on playground built on Amazon Bedrock for experimenting with and sharing generative AI apps. Ground Truth labels data, Macie finds sensitive data, and CloudTrail logs API activity.

A user hides malicious instructions inside input text to trick an LLM into ignoring its original instructions. What is this security risk called?

• a.Overfitting
• b.Data drift
• c.Prompt injection✓
• d.Model hallucination

Prompt injection manipulates a model by smuggling instructions into its input to override intended behavior. Hallucination is fabricated output, and drift or overfitting are training-related issues, not input attacks.

Which is a genuine limitation of generative AI that practitioners must plan for?

• a.It can produce plausible but incorrect (hallucinated) output, and its responses can be non-deterministic✓
• b.It is fundamentally incapable of generating any text
• c.It can never be given instructions of any kind
• d.It always returns identical output for every possible prompt

Generative models can hallucinate and vary their answers across runs, so outputs should be reviewed and, where accuracy matters, grounded in trusted sources. They clearly can generate text and follow instructions.

Building a customer-service virtual assistant that holds natural, open-ended conversations is a common use case for what?

• a.Generative AI large language models✓
• b.A tabular regression model
• c.Optical character recognition
• d.A classical time-series forecasting model

Open-ended conversational assistants are a hallmark generative AI and LLM use case. Regression and forecasting predict numbers, and OCR extracts text from images, none of which conduct free-form dialogue.

A marketing team wants to create original concept artwork from written descriptions. Which generative AI capability fits?

• a.Sentiment analysis
• b.Text-to-image generation✓
• c.Speech-to-text transcription
• d.Anomaly detection

Generating novel images from text prompts is text-to-image generation. Sentiment analysis and anomaly detection analyze existing data, and speech-to-text transcribes audio rather than creating images.

Generating a working function from a natural-language description of what it should do is an example of which generative AI use case?

• a.Code generation✓
• b.Data encryption
• c.Language translation
• d.Image classification

Turning a plain-language request into source code is code generation, a strength of coding-focused models and assistants such as Amazon Q Developer. The other options do not produce code.

Why does token count matter when using a large language model?

• a.Tokens determine how much text fits in the context window and typically drive usage-based pricing✓
• b.Tokens are the encryption keys that secure the model
• c.Tokens are the model's internal learned weights
• d.Token count has no effect on either cost or capacity

Text is processed as tokens; the number of tokens affects both how much content fits in the context window and how much you pay under per-token pricing. Tokens are units of text, not keys or weights.

How are foundation models typically pre-trained?

• a.On a small, fully hand-labeled dataset built for just one narrow task only
• b.On vast amounts of largely unlabeled data using self-supervised learning✓
• c.Without using any data at all
• d.Only on the individual customer's private documents

Foundation models are pre-trained on enormous, broad, mostly unlabeled corpora using self-supervised objectives, which gives them general capabilities. They are later adapted to specific tasks via prompting or fine-tuning.

In semantic search, two pieces of text are judged similar in meaning when what is true of their embeddings?

• a.They contain the exact same characters
• b.They were stored in the same S3 bucket
• c.Their vectors are close together in the embedding space✓
• d.They happen to contain an identical total number of tokens

Embeddings place semantically similar text near each other, so closeness in vector space signals similar meaning even when the wording differs. Matching characters or token counts is not what semantic similarity measures.

A key benefit of Amazon Bedrock is that it is serverless. What does that mean for the customer?

• a.The customer must manage a fleet of GPU servers themselves
• b.The models run only on the customer's on-premises hardware
• c.They can use foundation models through an API without provisioning or managing any underlying infrastructure✓
• d.The service can only be used from inside a physical data center

Serverless means AWS handles the infrastructure; customers simply call foundation models through an API and pay for what they use. There are no servers to provision, patch, or scale on the customer's side.

Which inference parameter limits how long a model's generated response can be?

• a.The maximum tokens (maximum length) setting✓
• b.The Top K setting
• c.The temperature setting
• d.The stop-sequence termination configuration setting

Maximum tokens caps the length of the generated output. Temperature and Top K shape randomness and diversity, and stop sequences end generation at specific text but do not by themselves set an overall length limit.

A company wants a foundation model to reflect its niche industry vocabulary more deeply than prompting can achieve. Which is a valid customization path on Amazon Bedrock?

• a.Increasing the temperature parameter
• b.Switching to a larger context window only
• c.Fine-tuning or continued pre-training the model on the company's own data✓
• d.Adding more stop sequences to each request

To embed specialized knowledge or style beyond what prompts provide, Bedrock supports customization via fine-tuning (labeled data) or continued pre-training (unlabeled domain data). Inference settings like temperature do not add knowledge.

What is a key advantage of generative AI over traditional rule-based automation for content tasks?

• a.It can produce novel, human-like content and adapt to varied inputs without hand-written rules for every case✓
• b.It guarantees perfectly accurate output every single time
• c.It never requires any human review of its output
• d.It eliminates all compute costs entirely

Generative AI flexibly creates new, human-like content and handles many input variations that rigid rules cannot anticipate. It does not guarantee accuracy, remove the need for review, or make compute free.

A support team needs answers grounded in internal policy documents that change frequently. Which approach avoids retraining the model each time the documents change?

• a.Fine-tuning the model again after every document edit
• b.Continued pre-training the model on the whole public internet
• c.Raising the temperature parameter on each request
• d.Retrieval Augmented Generation (RAG) using a knowledge base✓

RAG retrieves the latest relevant passages at query time and adds them to the prompt, so updated documents are reflected immediately without retraining. Fine-tuning would require re-running training whenever content changes.

Asking a model to 'think step by step' and show its reasoning before giving a final answer is which prompting technique?

• a.Retrieval augmentation with external documents
• b.Zero-shot prompting
• c.Negative prompting
• d.Chain-of-thought prompting✓

Chain-of-thought prompting encourages the model to reason through intermediate steps, which can improve performance on multi-step problems. Zero-shot gives no examples, and retrieval augmentation supplies external documents.

Giving a model an instruction with no worked examples and expecting it to perform the task is called what?

• a.Few-shot prompting
• b.Fine-tuning
• c.Zero-shot prompting✓
• d.Chain-of-thought prompting

Zero-shot prompting provides only the instruction, with no examples. Few-shot includes a handful of demonstrations, while fine-tuning changes the model's weights rather than just the prompt.