AWS Certified AI Practitioner (AIF-C01)

Correct answer: C. This is known as a hallucination, where the model generates seemingly accurate information that is, in fact, incorrect or fabricated

When an LLM generates plausible but factually incorrect information, it is called a hallucination. Terms like overfitting or data drift describe model training or data distribution issues, not the specific behavior of generating fabricated text.

Correct answer: B. Influences the percentage of most-likely candidates that the model considers for the next token

The Top P parameter defines the percentage of the most likely candidates the model considers for the next token. Temperature controls creativity, while Top K limits the fixed number of candidates considered.

Correct answer: B. With Amazon Bedrock, you can choose the underlying Foundation Model. However, Amazon Q does not allow you to choose the underlying Foundation Model · C. Amazon Q is a generative AI–powered assistant that allows you to create pre-packaged generative AI applications, whereas, Amazon Bedrock provides an environment to build and scale generative AI applications using a Foundation Model (FM)

Amazon Bedrock is a managed service for building generative AI applications using a choice of foundation models. Amazon Q is a pre-packaged assistant without model selection. Remember that Bedrock exposes models via API, while Q provides a ready-to-use assistant.

Correct answer: D. Amazon SageMaker Ground Truth

Amazon SageMaker Ground Truth creates high-quality labeled datasets using human workers. Feature Store manages ML features, while JumpStart deploys pre-trained models. Ground Truth is essential for data labeling.

Correct answer: B. Utilize a Retrieval-Augmented Generation (RAG) system by indexing all product catalog PDFs and configuring the LLM chatbot to reference this system for answering queries

A Retrieval-Augmented Generation system indexes documents to fetch relevant context per query. Fine-tuning is expensive and slow to update, and passing all PDFs exceeds context windows and inflates costs.

Correct answer: A. Amazon Bedrock · B. Amazon SageMaker JumpStart

Amazon Bedrock and SageMaker JumpStart provide managed environments to deploy and customize large language models. AWS Trainium and Inferentia are hardware chips, while Amazon Q is an assistant.

Correct answer: D. The company should use BLEU (Bilingual Evaluation Understudy) score, a metric specifically designed to evaluate the quality of text that has been machine-translated by comparing it with one or more reference translations

The BLEU score evaluates machine translation quality by comparing n-gram overlaps with reference translations. ROUGE is primarily used for summarization, while BLEU remains the standard for translation.

Correct answer: C. Shapley values provide a local explanation by quantifying the contribution of each feature to the prediction for a specific instance, while PDP provides a global explanation by showing the marginal effect of a feature on the model's predictions across the dataset. Use Shapley values to explain individual predictions and PDP to understand the model's behavior at a dataset level

Shapley values offer local explainability by assigning contribution scores to features for a specific prediction. Partial Dependence Plots provide global explainability by showing feature effects across the dataset.

Correct answer: B. Amazon Kendra

Amazon Kendra is an enterprise search service powered by machine learning that connects to various data repositories to provide accurate answers. Textract extracts text from documents, while Comprehend performs natural language processing but neither functions as a unified search endpoint.

Correct answer: A. The company should mask the confidential information from the model responses by leveraging Amazon Bedrock Guardrails

Amazon Bedrock Guardrails can dynamically scan and mask sensitive data or personally identifiable information in model responses. Retraining the model without the data is inefficient and expensive, while encryption only protects data in transit or at rest, not in the generated output.

Correct answer: B. The company should opt for on-demand pricing, which allows it to pay only for the actual usage of resources without any long-term commitments

On-demand pricing is best for companies with unpredictable usage because you only pay for what you use without long-term commitments. Spot and Reserved Instances are EC2 pricing models, while Provisioned Throughput requires a predictable usage commitment.

Correct answer: B. Amazon SageMaker Clarify

Amazon SageMaker Clarify provides feature attribution tools to explain how specific input features contribute to a model's predictions. Feature Store manages data inputs, Model Monitor tracks production drift, and Canvas is a no-code interface.

Correct answer: C. Tokens

Tokens represent the discrete units of text processed by generative AI models. Embeddings and vectors are numerical representations of these tokens, while the context window measures the total token capacity the model can process.

Correct answer: C. Using generative AI to create photorealistic images from textual descriptions

Creating photorealistic images from text descriptions is a classic generative AI capability because it produces entirely new content. The distractors fail because predicting housing prices or classifying images rely on traditional supervised machine learning rather than generative models.

Correct answer: C. AWS is responsible for the security "of" the cloud, including infrastructure, hardware, and software, while the customer is responsible for security "in" the cloud, including data, applications, and access management

AWS operates the security of the cloud, meaning it handles the underlying infrastructure, hardware, and facilities. Customers handle security in the cloud, which includes securing their data, managing access, and configuring applications.

Correct answer: D. Predict product demand to accurately vary inventory and pricing at different store locations

Amazon Forecast uses time-series data to predict future metrics, making it ideal for retail demand planning and inventory optimization. The other options map to different AI services like Personalize for recommendations or Lex for conversational interfaces.

Correct answer: A. Amazon Polly is used to deploy high-quality, natural-sounding human voices in dozens of languages · B. Amazon Comprehend service uses machine learning to find insights and relationships in the text

Amazon Polly converts text into lifelike speech, and Amazon Comprehend uses natural language processing to extract insights from text. The distractors mismatch the services, such as pairing Transcribe with conversational interfaces instead of Lex.

Correct answer: B. FMs use self-supervised learning to create labels from input data, however, fine-tuning an FM is a supervised learning process

Foundation models rely on self-supervised learning to generate implicit labels from vast amounts of raw data. Fine-tuning customizes the model for specific tasks using labeled examples, which makes it a supervised learning process.

Correct answer: C. OpenSearch Serverless vector store

Knowledge Bases for Amazon Bedrock natively defaults to OpenSearch Serverless for vector storage. Although it supports Pinecone and Aurora, the default store is automatically created if you do not provide one.

Correct answer: A. Amazon Transcribe and Amazon Comprehend

Amazon Transcribe converts audio calls into text. Amazon Comprehend then runs sentiment analysis on that text. Avoid Rekognition for audio, as it handles computer vision tasks.

Correct answer: C. The company should use hyperparameters for model tuning, which involves adjusting parameters such as regularization, learning rates, and dropout rates to enhance the model's ability to generalize well to new data

Hyperparameter tuning adjusts regularization and dropout rates to prevent overfitting, improving model generalization. While adding data helps, hyperparameter tuning specifically targets the gap between training and production.

Correct answer: D. Augment the data by generating new instances of data for underrepresented groups

Data augmentation directly balances underrepresented groups by creating new synthetic training examples, which mitigates bias. Manual correction is not scalable, regularization only prevents overfitting, and switching models ignores the root cause of bad data.

Correct answer: A. Amazon Rekognition

Amazon Rekognition offers pre-trained computer vision capabilities that you can customize for specific business needs. Textract is strictly for text extraction, while SageMaker is a broader service requiring you to build models from scratch.

Correct answer: C. Negative prompting

Negative prompting tells a generative AI model exactly what concepts or terms to exclude from its output. This is ideal for blocking competitor names or sensitive topics, whereas few-shot or chain-of-thought prompt techniques are used to guide desired content.

Correct answer: A. Foundation Models serve as a broad base for various AI applications by providing generalized capabilities, whereas Large Language Models are specialized for understanding and generating human language

Foundation Models are broad, general-purpose systems trained on massive datasets for various modalities like text or images. Large Language Models are a specialized subset of this category focused exclusively on understanding and generating human language.

Correct answer: D. Instructions, Context, Input data, Output Indicator

A well-crafted prompt consists of instructions, context, input data, and an output indicator. Remember that hyperparameters and parameters are configured at the model level, not manually embedded directly within the prompt text itself.

Correct answer: C. The data should include user-input along with the correct user intent, providing examples of user queries and the corresponding intent

Few-shot prompting teaches the model by pairing user inputs with their corresponding correct intent. Providing matching input and response pairs might show the final answer, but it fails to explicitly demonstrate the intermediate intent classification step.

Correct answer: A. The company should enable model invocation logging, which allows for detailed logging of all requests and responses during model invocations in Amazon Bedrock

Enabling model invocation logging in Amazon Bedrock captures the actual prompt inputs and generated outputs for auditing. CloudTrail only tracks control plane API calls for account security and does not record the payload data of the model responses.

Correct answer: B. Confusion matrix

A confusion matrix evaluates classification models by showing true positives, true negatives, false positives, and false negatives. Mean Absolute Error and Root Mean Squared Error are regression metrics designed for continuous numerical predictions instead of categorical ones.

Correct answer: A. Reinforcement learning involves an agent interacting with an environment by taking actions and receiving rewards or penalties, learning a policy to maximize cumulative rewards over time

Reinforcement learning uses an agent that interacts with an environment, receiving rewards or penalties to maximize cumulative success. It does not rely on labeled datasets like supervised learning or group unlabeled data like unsupervised learning.

Correct answer: C. The company should use Domain Adaptation Fine-Tuning, which involves fine-tuning the model on domain-specific data to adapt its knowledge to that particular domain · D. The company should use Continued Pre-Training, which involves further training the model on a large corpus of domain-specific data, enhancing its ability to understand domain-specific terms, jargon, and context

Domain Adaptation Fine-Tuning and Continued Pre-Training are the correct approaches because they use large amounts of unlabeled or labeled domain-specific data to teach the model specialized terminology and patterns. On the exam, remember that reinforcement learning relies on human feedback rather than raw domain knowledge.

Correct answer: D. Exploratory Data Analysis (EDA)

Exploratory Data Analysis is correct because calculating statistical summaries and generating visualizations to uncover trends is the core focus of EDA. Data preparation is the strongest distractor, but it involves cleaning and transforming data rather than exploring it for initial insights.

Correct answer: A. The company should use Amazon OpenSearch Service, which is designed to provide fast search capabilities and supports full-text search, indexing, and similarity scoring

Amazon OpenSearch Service is the correct choice because it provides native support for vector searches, fast index lookups, and similarity scoring required for retrieval-augmented generation workloads. Relational databases like Aurora lack these specialized semantic search capabilities.

Correct answer: A. Top K

Top K is the correct parameter because it directly limits the token pool to a specific number of the most likely next candidates. Temperature controls randomness globally, while Top P uses a probability threshold instead of a fixed count.

Correct answer: D. Amazon SageMaker Model Dashboard

The SageMaker Model Dashboard provides a centralized portal to track deployed models, monitor endpoints, and troubleshoot real-time inference issues. JumpStart is for deploying pre-trained models, while Clarify focuses solely on detecting bias in your data.

Correct answer: A. Sampling bias

Sampling bias is correct because the skewed outcomes stem directly from certain demographic groups being overrepresented or underrepresented in the training data. Measurement bias involves faulty sensors, while confirmation bias applies to human expectations.

Correct answer: C. Enable multilingual user experiences in your applications

Enabling multilingual experiences is the correct answer because translation tasks belong to Amazon Translate, not Rekognition. Rekognition is strictly for computer vision tasks like facial recognition, celebrity detection, and searching image libraries.

Correct answer: B. The company should use Provisioned Throughput mode, which allows the company to reserve a specific amount of capacity in advance

Provisioned Throughput is correct because Amazon Bedrock requires you to purchase this reserved capacity to test and deploy any models customized via fine-tuning. On-Demand mode only applies to base foundation models and cannot access your customizations.

Correct answer: B. Use higher Temperature to get more creative responses for the same prompt on Amazon Bedrock

A higher Temperature value is the correct setting because it increases the randomness of token selection, yielding more creative and diverse responses. Lower temperature values force the model to produce deterministic and predictable outputs.

Correct answer: D. Generative AI encompasses models and algorithms capable of creating new content such as text, images, and audio based on patterns learned from existing data

Generative AI creates new content like text, images, and audio by learning patterns from existing data. Options focusing on data retrieval, predefined tasks, or just analyzing data without creating content are incorrect because they fail to produce original output.

Correct answer: C. The company should reduce the number of tokens in the input

Generative AI model costs on Amazon Bedrock are calculated based on the number of input and output tokens processed. Adjusting temperature or top-P affects the randomness and diversity of the output but does not lower the usage price.

Correct answer: D. You can evaluate, compare, and select Foundation Models quickly based on pre-defined quality and responsibility metrics · E. Pre-trained models are fully customizable for your use case with your data

SageMaker JumpStart allows users to quickly evaluate foundation models and deploy customizable pre-trained models. SageMaker Canvas is the AWS no-code solution, while training data is never used to update the base models surfaced in JumpStart.

Correct answer: B. Model evaluation is the process of evaluating and comparing model outputs to determine the model that is best suited for a use case, whereas, model inference is the process of a model generating an output (response) from a given input (prompt)

Model inference is the act of a model generating an output from a given prompt, whereas model evaluation is the process of comparing these outputs to find the best model. The other options incorrectly swap or merge these two distinct concepts.

Correct answer: A. Manage machine learning experiments

MLflow integrates with SageMaker to help data scientists track, organize, and compare iterative machine learning experiments. Automatic tuning is handled by SageMaker AMT, while Canvas provides no-code tools and Ground Truth handles data labeling.

Correct answer: D. Transformer models use a self-attention mechanism and implement contextual embeddings

Transformer models use self-attention mechanisms to weigh the importance of words in a sequence alongside contextual embeddings. The other options describe generative adversarial networks, variational autoencoders, and diffusion models instead of transformers.

Correct answer: C. Bidirectional Encoder Representations from Transformers (BERT)

BERT uses bidirectional context to generate dynamic word embeddings, capturing nuanced meanings based on surrounding text. Older models like Word2Vec rely on static embeddings, while SVD and PCA are dimensionality reduction techniques that lack language context.

Correct answer: A. Knowledge Bases for Amazon Bedrock

Knowledge Bases for Amazon Bedrock provides a fully managed end-to-end Retrieval Augmented Generation workflow by ingesting documents into a vector store. Guardrails filter harmful content, while continued pretraining customizes model weights without retrieving external data.

Correct answer: B. Agents for Amazon Bedrock

Agents for Amazon Bedrock autonomously orchestrate multi-step tasks by integrating foundation models with proprietary data and APIs. Knowledge Bases only retrieve information, whereas Guardrails and watermark detection focus on safety and image provenance respectively.

Correct answer: D. Continued pre-training uses unlabeled data to pre-train a model, whereas, fine-tuning uses labeled data to train a model

Continued pre-training uses unlabeled data to adapt a foundation model to new domains, while fine-tuning uses labeled data to optimize it for specific tasks. Expect exam questions testing the distinction between unsupervised domain adaptation and supervised task-specific training.

Correct answer: B. Define the use case of the application narrowly

Defining the use case narrowly gives the research team specific, bounded requirements, which directly drives correct model selection. Broadly defining an audience or listing data sources are preliminary discovery steps, but they do not by themselves ensure the best model is chosen.

Correct answer: C. Decision Trees

Decision trees provide high interpretability through clear, visual rules, making them the best fit when you must explain how features drive a classification. Neural networks act as black boxes, and while powerful, they lack the inherent transparency required here.

Correct answer: A. Elevate the customer experience with ML-powered personalization

Amazon Personalize uses machine learning to generate real-time, customized product and content recommendations. The other options describe different AWS AI services: Comprehend for text insights, Transcribe for speech-to-text, and Polly for text-to-speech.

Correct answer: D. Plagiarism

Detecting unauthorized AI-generated essays addresses plagiarism by ensuring applicants submit original work. Hallucination refers to an AI inventing false facts, while bias involves unfair discrimination, neither of which is the primary concern here.

Correct answer: A. Generative AI can automate the creation of new data based on existing patterns, enhancing productivity and innovation

Automating the creation of new data based on existing patterns is the primary advantage of generative artificial intelligence, directly enhancing productivity and driving innovation. The other options are incorrect because no AI can fully replace humans or guarantee absolute security.

Correct answer: B. Large Language Model (LLM) · C. Retrieval-Augmented Generation (RAG)

Amazon Q Business uses Large Language Models to generate text and Retrieval-Augmented Generation to fetch context from connected enterprise data. Diffusion models, GANs, and VAEs primarily generate images or synthetic tabular data, not enterprise search answers.

Correct answer: C. Amazon Comprehend · D. Amazon Bedrock

Amazon Comprehend directly analyzes text sentiment, while Amazon Bedrock allows foundation models to perform custom sentiment analysis. Rekognition handles images, Textract extracts document text, and Personalize builds recommendation systems, so none of those process written sentiment.

Correct answer: C. Asynchronous inference

Asynchronous inference is specifically designed for payloads under one gigabyte where immediate real-time responses are unnecessary. Batch transform is wrong because it targets massive datasets rather than these smaller daily payloads processed individually.

Correct answer: A. The company should leverage reinforcement learning (RL), where rewards are generated from positive customer feedback to train the chatbot in optimizing its responses

Reinforcement learning uses reward signals from positive feedback to dynamically optimize chatbot responses over time. Incremental training could technically achieve ongoing adaptation, but it lacks the explicit reward optimization focus that reinforcement learning provides.

Correct answer: A. This concept is referred to as a context window, which determines the amount of text or information the model can consider at once while generating a response, typically measured in tokens rather than characters

The context window dictates the maximum amount of text, measured in tokens, that a model can process in a single interaction. Tokens themselves are merely the individual subword units, while embeddings represent text as numerical vectors for semantic understanding.

Correct answer: D. AWS Trusted Advisor

AWS Trusted Advisor provides real-time guidance to help you follow AWS best practices across cost, performance, security, and fault tolerance. Distractors like AWS Config focus on resource configuration tracking and compliance auditing rather than holistic optimization recommendations.

Correct answer: D. The company should use transfer learning, a method where a model pre-trained on one task is adapted to improve performance on a different but related task by leveraging knowledge from the original task

Transfer learning applies knowledge from a pre-trained model to a related task. While federated learning is a better fit for the scenario of distributed models sharing insights, transfer learning is the intended answer among the provided choices.