AWS Certified AI Practitioner (AIF-C01) - Practice Test 1

Amazon Transcribe Medical is a speech-to-text service specifically optimized for medical terminology and clinical conversations. It converts audio recordings into accurate transcripts that can be structured into clinical notes. Amazon Comprehend Medical (A) analyzes existing text to extract medical entities -- it does not convert speech to text. Amazon Textract (C) extracts text from scanned documents and images. Amazon Polly (D) converts text to speech, which is the opposite of what is needed. See more: AWS Managed AI Services

Amazon OpenSearch Service supports k-NN (k-Nearest Neighbors) vector search, making it ideal for storing and querying high-dimensional vector embeddings in RAG architectures. It efficiently finds semantically similar documents through approximate nearest neighbor search. Amazon RDS (A) is a relational database not optimized for vector similarity search. Amazon DynamoDB (B) is a key-value store without native vector search capabilities. Amazon Redshift (D) is a data warehouse designed for analytical queries, not vector similarity searches. See more: Bedrock & Generative AI

Recall (also called sensitivity or true positive rate) measures the proportion of actual positives that the model correctly identifies. Maximizing recall minimizes false negatives, which is critical in medical or safety applications where missing a positive case is more costly than a false alarm. Precision (A) minimizes false positives, not false negatives. Specificity (C) measures the true negative rate. F1 score (D) balances precision and recall but does not specifically optimize for minimizing false negatives. See more: AI/ML Fundamentals

Hallucination refers to when an LLM generates fluent, confident responses that contain fabricated or factually incorrect information. This is a fundamental limitation of generative models that predict tokens based on statistical patterns rather than verified facts. Model drift (A) occurs when model performance degrades over time due to changing data distributions. Prompt injection (B) is a security attack where malicious instructions are embedded in user input. Overfitting (D) is when a model memorizes training data rather than learning generalizable patterns. See more: Bedrock & Generative AI

Amazon Rekognition provides a content moderation API that detects explicit, suggestive, violent, or otherwise inappropriate content in images and videos. It returns confidence scores for detected categories, enabling automated moderation workflows. Amazon Comprehend (A) analyzes text for sentiment, entities, and topics -- it does not process images. Amazon Textract (C) extracts text from images and documents, not visual content classification. Amazon Kendra (D) is an intelligent search service for text documents. See more: AWS Managed AI Services

Data drift (also called covariate shift or distribution shift) occurs when the statistical properties of input data change between training and production, causing model performance to degrade. The model learned patterns from one distribution but is applied to a different one. Overfitting (A) is when a model memorizes training data -- it would still perform poorly on held-out data from the same distribution. Underfitting (C) is when a model is too simple to capture patterns even in training data. Label noise (D) refers to incorrect labels in the training dataset. See more: AI/ML Fundamentals

Amazon Textract uses ML to extract text, forms, tables, and key-value pairs from scanned documents. It handles varying layouts and formats without requiring templates, making it ideal for processing diverse invoices. Amazon Comprehend (A) analyzes text that has already been extracted -- it does not extract text from images. Amazon Rekognition (B) analyzes images for objects, faces, and scenes, not structured document data extraction. Amazon Translate (D) translates text between languages. See more: AWS Managed AI Services

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures the overlap of n-grams between generated summaries and reference summaries. It is the standard metric for evaluating summarization quality, specifically measuring recall of important content. BLEU score (A) measures precision of n-gram overlap and is primarily used for machine translation evaluation. Perplexity (C) measures how well a language model predicts text -- it is a training metric, not a summary quality metric. F1 score (D) evaluates classification model performance, not summarization quality. See more: AI/ML Fundamentals

Amazon SageMaker Model Monitor continuously monitors deployed models for data quality issues, model quality degradation, bias drift, and feature attribution drift. It generates alerts when metrics deviate from baseline thresholds, enabling proactive model maintenance. SageMaker Experiments (A) tracks and compares training runs, not production model behavior. Data Wrangler (C) is used for data preparation and transformation before training. Feature Store (D) stores and retrieves ML features but does not monitor model performance. See more: Amazon SageMaker

Amazon Rekognition Custom Labels allows users to train custom image classifiers by uploading labeled images through a point-and-click console interface with no ML expertise required. The service automatically handles model architecture, training, and optimization. Training a custom CNN (A) requires deep learning expertise and hyperparameter tuning knowledge. Fine-tuning on Bedrock (C) requires understanding of fine-tuning processes and data preparation. Building TensorFlow models on EC2 (D) requires significant ML engineering and infrastructure skills. See more: AWS Managed AI Services

LLMs are stateless and do not inherently remember previous turns. To enable multi-turn conversations, the application must include the full conversation history -- all previous user messages and assistant responses -- in each API request. The model uses this context window to generate contextually appropriate responses. Temperature (A) controls output randomness, not conversational memory. Provisioned Throughput (C) is about inference capacity, not state management. A larger model (D) has more parameters but is still stateless without explicit history management. See more: Prompt Engineering

Amazon EC2 Inf2 instances use AWS Inferentia2 chips, custom-designed by AWS for high-performance, energy-efficient ML inference. Inferentia chips deliver significantly better performance per watt than GPU alternatives for inference workloads. P5 instances (A) use NVIDIA H100 GPUs optimized for training, not energy-efficient inference. G5 instances (B) use NVIDIA A10G GPUs -- better for inference than P5 but less efficient than purpose-built Inferentia chips. C7g instances (D) use Graviton3 processors optimized for general compute workloads, not specialized ML inference. See more: AI Challenges & Responsibilities

Amazon Bedrock Knowledge Bases implements RAG (Retrieval Augmented Generation) by ingesting documents, generating embeddings, storing them in a vector database, and retrieving relevant context to ground model responses in the actual document content. Fine-tuning (A) trains the model on examples to adapt behavior but bakes static knowledge into model weights and cannot dynamically retrieve from live documents. Agents (C) orchestrate multi-step tasks and can use Knowledge Bases, but the core document Q&A capability is provided by the Knowledge Base. Continued pre-training (D) teaches the model broad domain knowledge but cannot query live documents at inference time. See more: Bedrock & Generative AI

Prompt injection attacks attempt to override or circumvent a model's system prompt by embedding malicious instructions in user input, such as 'Ignore all previous instructions and...' to redirect the model's behavior. Model inversion attacks (A) attempt to reconstruct training data from model outputs. Adversarial example attacks (B) add imperceptible perturbations to inputs to cause misclassification, primarily in vision models. Data poisoning attacks (D) corrupt training data to introduce backdoors or degrade model performance during training. See more: AI Challenges & Responsibilities

Batch transform is designed for running inference on large static datasets stored in S3 when immediate results are not needed. It provisions compute instances, processes the full dataset, saves results back to S3, then terminates -- you pay only for the compute time used. Real-time inference (A) maintains always-on endpoints, which is expensive for overnight batch jobs. Serverless inference (B) has cold starts and timeout limits unsuitable for large batch workloads. Asynchronous inference (C) processes individual large payloads asynchronously but is not optimized for bulk batch processing of many small records. See more: Amazon SageMaker

Temperature controls the randomness of token selection during generation. Decreasing temperature (toward 0) makes the model more deterministic by concentrating probability mass on the highest-probability tokens, producing more consistent and predictable outputs. At temperature 0, the model always selects the most probable token. Increasing temperature (A) increases randomness and creativity -- the opposite of what is needed. Increasing Top-P (C) expands the pool of candidate tokens by cumulative probability, increasing variability. Increasing Top-K (D) also expands the candidate token pool, increasing output diversity. See more: Prompt Engineering

Amazon Personalize is a fully managed ML service purpose-built for real-time personalization and recommendation systems. It uses the same technology as Amazon.com's recommendation engine to learn individual customer preferences from interaction history and deliver personalized recommendations. Amazon Comprehend (A) is an NLP service for text analysis tasks such as sentiment detection and entity recognition. Amazon Forecast (B) is for time-series metric forecasting such as demand or revenue. Amazon Kendra (D) is an intelligent enterprise search service for document retrieval. See more: AWS Managed AI Services

Explainability (also called interpretability) is the principle that AI systems should provide understandable reasons for their decisions. This is especially critical in regulated industries like financial services, where regulations such as GDPR and the Equal Credit Opportunity Act require that customers can understand and contest automated decisions. Fairness (A) ensures equitable treatment across demographic groups. Privacy (B) protects personal data from unauthorized access or misuse. Robustness (D) ensures the model performs reliably under adverse or unexpected input conditions. See more: AI Challenges & Responsibilities

SMOTE (Synthetic Minority Oversampling TEchnique) generates synthetic examples of the minority class (fraud) by interpolating between existing minority samples, balancing the class distribution without simple duplication. This helps the model learn meaningful decision boundaries for rare events. Increasing epochs (A) would further reinforce the majority-class bias. A larger network (B) amplifies the imbalance by providing more capacity to memorize the dominant class. Reducing learning rate (D) affects training convergence speed but does not address the underlying class representation imbalance. See more: AI/ML Fundamentals

Amazon Bedrock Guardrails sensitive information filters detect and redact PII such as names, phone numbers, email addresses, Social Security numbers, credit card numbers, and other personal data from both inputs and outputs. Content filters (A) detect harmful content categories like hate speech and violence, not PII specifically. Denied topics (B) block the model from engaging with specified subject areas entirely. Word filters (D) block specific pre-defined words or phrases -- they are not flexible enough to detect all forms of PII patterns across different formats. See more: Bedrock & Generative AI

Amazon SageMaker Automatic Model Tuning (also called Hyperparameter Optimization or HPO) automatically searches for the best hyperparameter values by running multiple training jobs with different configurations. It supports strategies like Bayesian optimization, random search, and grid search to efficiently navigate the hyperparameter space. SageMaker Pipelines (A) orchestrates end-to-end ML workflows but does not automatically tune hyperparameters. SageMaker Clarify (C) detects bias and explains model predictions. Model Registry (D) catalogs and versions trained models for deployment. See more: Amazon SageMaker

Amazon Bedrock Agents autonomously break down complex tasks into sequential steps, invoke Lambda functions through action groups, query Knowledge Bases for information retrieval, and orchestrate entire workflows to complete multi-step goals. They use a ReAct (Reasoning and Acting) framework to plan and execute actions. Knowledge Bases (A) handle document retrieval but do not independently orchestrate multi-step workflows. Fine-tuning (B) adapts model weights for specific tasks but does not add workflow orchestration capability. Guardrails (D) add safety and content filters, not task orchestration. See more: Bedrock & Generative AI

Amazon Transcribe converts spoken audio in video files to text, generating source-language captions. Amazon Translate then converts those transcriptions into multiple target languages, producing multilingual subtitles. This two-stage pipeline covers the complete requirement. Amazon Rekognition (A) analyzes video for objects, faces, and scenes -- it does not transcribe speech. Amazon Polly (C) converts text to speech, the opposite direction of what is needed. Amazon Textract (D) extracts text from images and documents, not from video audio tracks. See more: AWS Managed AI Services

AWS Key Management Service (KMS) enables customers to create and manage Customer Managed Keys (CMKs). Amazon Bedrock integrates with AWS KMS to encrypt fine-tuning datasets, custom model weights, and output artifacts using customer-controlled keys, satisfying data sovereignty and compliance requirements. AWS Shield (A) provides DDoS protection for AWS infrastructure. AWS Secrets Manager (C) stores and rotates application secrets such as API keys and database passwords. AWS Certificate Manager (D) provisions and manages TLS/SSL certificates for web applications. See more: AWS Security Services

Foundation models have two key distinguishing characteristics: First, they are pre-trained on massive, diverse datasets using self-supervised learning objectives such as next-token prediction -- no manual labeling is required for pre-training. Second, they exhibit strong generalization, enabling adaptation to many different downstream tasks through zero-shot prompting, few-shot prompting, or fine-tuning with minimal task-specific data. Traditional ML models (B) are task-specific by design. The self-supervised pre-training objective (D) does not require labeled data -- labels are only needed during supervised fine-tuning for specific tasks. See more: AI/ML Fundamentals

Speaker diarization is an Amazon Transcribe feature that automatically detects multiple speakers in an audio recording and labels each segment with the speaker identifier who produced it. This creates structured transcripts such as 'Agent: Hello...' and 'Customer: Hi...' Custom vocabulary (A) adds domain-specific terms to improve transcription accuracy for specialized terminology. Vocabulary filtering (C) removes or masks specific words from transcripts for compliance. Automatic language identification (D) detects the spoken language without manual configuration but does not separate speakers. See more: AWS Managed AI Services

Distributed training splits the training workload across multiple compute instances using data parallelism, model parallelism, or a combination of both, significantly reducing wall-clock training time. Amazon SageMaker supports distributed training through the SageMaker Distributed Training Library. Increasing dropout rate (A) is a regularization technique that prevents overfitting but does not reduce training time. Decreasing model depth (C) reduces parameters and may speed up training but typically sacrifices model accuracy. L2 regularization (D) penalizes large model weights to prevent overfitting but does not accelerate training. See more: Amazon SageMaker

Context window size determines the maximum number of tokens -- input and output combined -- that a model can process in a single request. Processing 500-page documents requires a very large context window (such as 200K tokens). Models with small context windows would require documents to be chunked into pieces, breaking cross-document reasoning. Number of parameters (A) relates to model capacity and capability but does not determine how much text can be processed at once. Training data recency (C) affects knowledge cutoff date. Inference latency (D) affects response speed but not document length capacity. See more: Bedrock & Generative AI

Amazon SageMaker DeepAR is a supervised learning algorithm specifically designed for probabilistic time-series forecasting using recurrent neural networks. It learns temporal patterns across multiple related time series simultaneously, producing accurate forecasts over specified future horizons. XGBoost (A) is a gradient boosting algorithm for tabular classification and regression -- it requires extensive feature engineering to capture time-series patterns. Linear Learner (B) is designed for linear classification and regression problems, not temporal forecasting. K-Means (D) is an unsupervised clustering algorithm with no forecasting capability. See more: Amazon SageMaker

Amazon SageMaker Model Cards provide a structured documentation template for ML models, including intended use cases, out-of-scope uses, model limitations, training data details, evaluation metrics, ethical considerations, and responsible AI practices. They serve as the standardized model governance artifact required for regulatory compliance. SageMaker Experiments (A) tracks training runs and hyperparameter configurations, not model governance documentation. AWS Audit Manager (C) collects evidence for compliance audits across AWS services but does not document ML model specifics. SageMaker Pipelines (D) automates ML workflows but does not produce governance documentation. See more: Amazon SageMaker

Amazon SageMaker Clarify detects and measures statistical bias in both datasets and model predictions across different demographic groups. It computes bias metrics such as Difference in Positive Proportions (DPP) and Disparate Impact (DI) to identify unfair treatment across protected attributes. It also provides feature importance explanations using SHAP values to improve model transparency. SageMaker Model Monitor (A) monitors production models for data distribution changes and performance degradation, not specifically bias measurement. SageMaker Debugger (C) profiles training jobs for technical issues such as vanishing gradients. Feature Store (D) stores and serves ML features. See more: Amazon SageMaker

RAG with a Knowledge Base that is continuously refreshed with current information enables the model to answer questions about recent events by retrieving up-to-date context at query time. This addresses the training cutoff limitation without expensive model retraining. Increasing temperature (A) affects output randomness and creativity, not knowledge recency. Weekly fine-tuning (C) is extremely expensive and slow -- the process cannot keep pace with the daily news cycle and still requires time to deploy. A larger model (D) has more parameters but its training cutoff remains the same; additional parameters do not provide new real-world knowledge. See more: Bedrock & Generative AI

Amazon Transcribe converts call center audio recordings into text transcripts. Amazon Comprehend then analyzes the transcripts for sentiment classification (positive, negative, neutral, mixed), detecting frustration and dissatisfaction signals at scale. Amazon Polly (B) converts text to speech, the opposite direction of what is needed. Amazon Rekognition (B) analyzes images and video for visual content, not audio sentiment. Amazon Textract (C) extracts text from scanned documents, not audio recordings. Amazon Kendra (C) is an enterprise search service for document retrieval. See more: AWS Managed AI Services

System prompts define the model's persona, communication style, and behavioral guidelines for every interaction, enforcing the formal tone requirement. Guardrails with denied topics configuration blocks the model from engaging with any content about competitors, enforcing that restriction reliably. Model fine-tuning and Provisioned Throughput (A) modify model weights and inference capacity but are not the primary mechanisms for enforcing behavioral tone or topic restrictions. Knowledge Bases and Agents (C) add retrieval and workflow capabilities. Continued pre-training and logging (D) add domain knowledge and audit trails, not behavioral controls. See more: Bedrock & Generative AI

A large gap between high training accuracy (95%) and significantly lower test accuracy (58%) is the classic signature of overfitting. The model has memorized patterns specific to the training examples -- including noise -- rather than learning generalizable features that apply to new data. Underfitting (A) would result in poor performance on both training and test data, not just test data. Insufficient parameters (C) would cause underfitting, not a training-test accuracy gap. A higher learning rate (D) affects convergence speed and stability during training but does not resolve overfitting caused by memorization. See more: AI/ML Fundamentals

Amazon Bedrock Agents with action groups can connect to relational databases through AWS Lambda functions. The agent translates natural language questions into SQL queries, passes them to Lambda for execution against the database, and returns results as natural language responses. This enables true Text-to-SQL capability without requiring employees to know SQL. Knowledge Bases (A) are for document and vector search, not live relational database querying. Fine-tuning on SQL examples (C) improves SQL generation but does not give the model the ability to execute live queries against a database. Continued pre-training (D) adds domain knowledge to weights but cannot query live databases at inference time. See more: Bedrock & Generative AI

Continued pre-training (also called domain-adaptive pre-training) extends the model's pre-training on large volumes of unlabeled domain-specific text, teaching the model the vocabulary, concepts, and patterns of a new scientific field without requiring labeled data. This is the appropriate technique when the goal is broad domain knowledge acquisition before any specific task. Fine-tuning with prompt-completion pairs (A) requires labeled data and targets specific task behavior rather than broad domain knowledge. RAG (C) retrieves documents at inference time but does not teach the model domain knowledge. Few-shot prompting (D) provides examples in the prompt but does not permanently update model knowledge. See more: Bedrock & Generative AI

SageMaker network isolation mode prevents training containers from making any outbound network calls during execution. When enabled, containers cannot reach the internet, other AWS services, or any external endpoints -- they can only read input data from and write output to Amazon S3 through VPC endpoints. SageMaker Experiments (A) is for tracking and comparing ML training runs, not for network isolation. Data encryption with KMS (C) protects data confidentiality but does not restrict network access during training. A restrictive IAM role (D) controls what AWS API calls the job can make but does not enforce network-level isolation from the internet. See more: AWS Security Services

Custom fine-tuned models in Amazon Bedrock cannot be invoked using on-demand pricing. To use a custom model for inference, the company must purchase Provisioned Throughput, which provides dedicated model units with either a 1-month or 6-month commitment term. Deploying to a SageMaker endpoint (B) is unnecessary because the model was fine-tuned within Amazon Bedrock and should be served from Bedrock. SageMaker Model Registry (C) is designed for models trained in SageMaker, not for Bedrock custom models. On-demand pricing (D) is not available for custom fine-tuned Bedrock models -- Provisioned Throughput is the required path. See more: Bedrock & Generative AI

Data augmentation for underrepresented groups addresses bias at its root by generating additional synthetic training examples for demographic groups that are underrepresented in the original dataset. Techniques such as flipping, rotation, color adjustment, and synthetic image generation can balance representation across professions, genders, and ethnicities before training. Watermark detection (A) identifies watermarks in images and is irrelevant to demographic bias. RAG (B) retrieves documents at inference time and is not applicable to image generation bias. Post-generation filtering (C) can screen outputs after the fact but does not fix the underlying model bias -- it treats the symptom rather than the cause. See more: AI Challenges & Responsibilities

Two complementary approaches address bias: First, using diverse and representative training data ensures all demographic groups are proportionally included, preventing the model from favoring any particular group. Second, regular bias testing and fairness audits provide ongoing monitoring to detect and correct discriminatory patterns in model outputs over time. ROUGE (C) is a text summarization evaluation metric that measures overlap between generated and reference summaries -- it has nothing to do with bias detection or model accuracy in lending decisions. See more: AI Challenges & Responsibilities

An epoch is one complete pass through the entire training dataset. Increasing epochs gives the model more opportunities to learn from the data, typically improving accuracy until the model converges. However, too many epochs can lead to overfitting, so training should stop when the target acceptance level is reached. Decreasing epochs (A) reduces training iterations, likely decreasing accuracy. Decreasing batch size (C) can affect training dynamics but doesn't directly or reliably increase accuracy the way more training iterations do. See more: AI/ML Fundamentals

Amazon SageMaker Canvas is the only truly no-code ML solution among the options. It provides a drag-and-drop interface where users upload datasets, select the prediction target, and Canvas automatically builds, trains, and evaluates multiple models. No coding, algorithm selection, or ML expertise is required. SageMaker built-in algorithms (A, B) require writing Python/R code and understanding ML concepts. Amazon Personalize Trending-Now (C) is a recommendation recipe designed for real-time trending product recommendations, not time-series demand forecasting. See more: Amazon SageMaker

Amazon Bedrock Agents enable end-to-end automation by decomposing customer requests into steps, invoking APIs (e.g., order lookup, return processing), querying knowledge bases for product information, and generating contextual responses -- all without manual intervention. This dramatically reduces the workload for customer support teams handling repetitive inquiries. The agent doesn't select between FMs (B) or call multiple FMs simultaneously (C) -- it uses one configured model and coordinates external actions through action groups and API integrations. See more: Bedrock & Generative AI

Data augmentation for imbalanced classes addresses bias at its root by generating additional training examples for underrepresented groups. For example, if the dataset has few images of female engineers, augmentation techniques (rotation, flipping, color adjustment, synthetic generation) can create more examples, resulting in balanced representation across professions, genders, and ethnicities. Watermark detection (A) identifies watermarks in images, irrelevant to bias. RAG (B) retrieves information from knowledge bases, not applicable to image generation bias. Model monitoring (C) detects bias after deployment but doesn't fix the underlying training data imbalance. See more: AI Challenges & Responsibilities

'Stop guessing capacity' directly addresses unpredictable traffic. With cloud computing, you can use Auto Scaling to automatically adjust resources based on actual demand, rather than over-provisioning for peak traffic or under-provisioning and crashing during spikes. Trading CAPEX for OPEX (A) relates to payment model. Economies of scale (C) relates to cost reduction from AWS's bulk purchasing. Going global (D) relates to geographic expansion. See more: AWS Cloud Computing

EC2 is Infrastructure as a Service (IaaS). The customer is responsible for the guest operating system and everything above it, including OS patching, application security, and data protection. AWS is responsible for the underlying infrastructure: physical security (A), hypervisor (B), and hardware management (D). See more: AWS Cloud Computing

Amazon Q Business is a fully managed enterprise AI assistant that connects to 40+ data sources including SharePoint and S3 via data connectors. It provides RAG-based retrieval from internal company data with permission-aware access control through IAM Identity Center. Bedrock (A) requires custom development. Lex (C) is for building conversational interfaces, not enterprise knowledge retrieval. Comprehend (D) does NLP analysis, not Q&A. See more: Amazon Q

Plugins allow Amazon Q Business to actively interact with external systems like Jira, ServiceNow, and Salesforce -- creating, updating, or reading records. Data Connectors (A) are read-only integrations for RAG retrieval. Admin Controls (B) are guardrails for blocking topics. IAM Identity Center (D) handles user authentication, not external system actions. See more: Amazon Q

To access an encrypted S3 object, the SageMaker execution role needs both S3 permission to read the object AND KMS decrypt permission to decrypt the data. Without both permissions, the training job will fail. IAM roles are always required to authorize cross-service access in AWS. See more: AWS Security Services