# AWS Certified AI Practitioner (AIF-C01) Free Practice Exam and Practice Test

AIF-C01 · FOUNDATIONAL. Page: https://blowstack.com/practice-exams/aws-certified-ai-practitioner

> Prepare for the AWS Certified AI Practitioner exam with our practice exam simulator. Realistic questions on AI, machine learning and generative AI fundamentals, with detailed explanations, help you find your knowledge gaps and assess your readiness.

Exam version AIF-C01: current, beta start Aug 13, 2024, release Oct 22, 2024. All exam dates: https://blowstack.com/tools/cloud-certifications/aws

## Exam at a glance

- Question bank: 928 practice questions
- Questions updated: 2026-10-30
- Free AIF-C01 practice exam (20 questions weighted by domain, in the exam simulator on this page, no sign-up): https://blowstack.com/practice-exams/aws-certified-ai-practitioner
- AIF-C01 exam simulator (full-length or custom exams with explanations and a pass probability indicator): https://app.blowstack.com/practice-exams/configurator/aws-certified-ai-practitioner
- Questions: 65
- Duration: 90 min
- Passing score: 70%
- Domains: 5
- Interactive AIF-C01 course (246 lessons built around exercises, no videos): https://blowstack.com/courses/aws-certified-ai-practitioner
- Official AIF-C01 exam guide (the AWS document listing every domain, task and service in scope): https://docs.aws.amazon.com/aws-certification/latest/ai-practitioner-01/ai-practitioner-01.html

## AIF-C01 example questions

One example question per domain from the AIF-C01 question bank, with the correct answer and its explanation.

### Fundamentals of AI and ML

**Question:**

A telecommunications company wants to use Amazon SageMaker AI to discover previously unknown groups of customers with similar usage patterns. It has a table with one row per customer and columns for customer_id, account_age_months, average_monthly_data_usage, and average_monthly_call_minutes. No customer groups have been defined or assigned. Which approach best matches the available data and the company's goal?

**Correct answer:** Treat the table as structured, unlabeled data and use the account-age and usage columns as inputs to a clustering model.

**Explanation:** The consistent rows and columns make this structured, tabular data. It is unlabeled for customer grouping because no group assignments exist. Clustering in SageMaker AI can discover similarities without predefined group labels.

### Fundamentals of GenAI

**Question:**

A logistics company uses an Amazon Bedrock foundation model to assign shipment exception reports to company-specific routing categories. Providing expert-labeled examples in each prompt improves accuracy, but the examples make requests too long. Sending those examples once does not improve later, independent requests. The company has thousands of reviewed report-and-category pairs, and its selected model supports customization. It wants a reusable adaptation that works without including labeled examples in every request. Which approach best meets these requirements?

**Correct answer:** Fine-tune the model using the reviewed report-and-category pairs, then use the customized model for classification.

**Explanation:** Fine-tuning uses targeted examples to update learned parameters or adaptation parameters. It can teach the report-to-category mapping as reusable behavior without requiring the training examples in every inference request.

### Applications of Foundation Models

**Question:**

A consulting firm uses a retrieval-augmented generation (RAG) assistant to answer questions about market research. Document embeddings are stored in an Amazon OpenSearch Service vector index. The firm switches to a different embedding model for user questions but leaves the existing document embeddings unchanged. Although both embedding models produce vectors of the same length, the assistant now retrieves unrelated passages. Which change should the firm make to restore meaningful retrieval?

**Correct answer:** Regenerate the document embeddings with the new embedding model, rebuild the OpenSearch vector index, and use that model for query embeddings.

**Explanation:** Document and query embeddings must use compatible representations for meaningful vector similarity searches. Regenerating the document embeddings with the new model aligns them with query embeddings. Equal vector lengths alone do not establish compatibility.

### Guidelines for Responsible AI

**Question:**

A healthcare provider is evaluating an AI model that recommends patients for additional screening. The model has high overall accuracy, but the prevalence of the condition differs across demographic groups. Clinical leaders require that patients who actually have the condition have comparable chances of being referred, regardless of demographic group. Which evaluation approach most directly determines whether the model meets this requirement?

**Correct answer:** Among patients with independently confirmed cases, compare referral rates across demographic groups and investigate any material gaps.

**Explanation:** This compares true positive rates across groups, directly testing whether patients who have the condition have comparable chances of referral. Investigating material gaps supports addressing harmful disparities before deployment.

### Security, Compliance, and Governance for AI Solutions

**Question:**

A company is piloting an AI assistant on AWS that summarizes regulatory publications and includes citations to source documents. During user testing, compliance analysts assume that a cited summary has been verified against its sources and can be used without further checking. The company's governance policy requires users to understand the system's limitations before receiving production access. Which change to mandatory onboarding would BEST address the finding?

**Correct answer:** Explain that citations identify supporting material but do not establish correctness, demonstrate how to compare summaries with the original sources, and define when to request expert review.

**Explanation:** This provides understandable information about the assistant's limitations and teaches appropriate use. Source verification and an expert-review path directly address the mistaken assumption that citations prove a summary is correct.

## Free AIF-C01 cheat sheets

- [Domain 1: Fundamentals of AI and ML](https://blowstack.com/cheat-sheets/aws-certified-ai-practitioner/aif-c01-domain-1.pdf) (20% of the exam, PDF)
- [Domain 2: Fundamentals of GenAI](https://blowstack.com/cheat-sheets/aws-certified-ai-practitioner/aif-c01-domain-2.pdf) (24% of the exam, PDF)
- [Domain 3: Applications of Foundation Models](https://blowstack.com/cheat-sheets/aws-certified-ai-practitioner/aif-c01-domain-3.pdf) (28% of the exam, PDF)
- [Domain 4: Guidelines for Responsible AI](https://blowstack.com/cheat-sheets/aws-certified-ai-practitioner/aif-c01-domain-4.pdf) (14% of the exam, PDF)
- [Domain 5: Security, Compliance, and Governance for AI Solutions](https://blowstack.com/cheat-sheets/aws-certified-ai-practitioner/aif-c01-domain-5.pdf) (14% of the exam, PDF)

## Exam domains

The AIF-C01 questions are grouped into 5 domains divided into 14 tasks.

### Fundamentals of AI and ML (20% of the exam)

- Explain basic AI concepts and terminologies
- Identify practical use cases for AI
- Describe the AI/ML development lifecycle

### Fundamentals of GenAI (24% of the exam)

- Explain the basic concepts of generative AI (GenAI)
- Understand the capabilities and limitations of GenAI for solving business problems
- Describe AWS infrastructure and technologies for building GenAI applications

### Applications of Foundation Models (28% of the exam)

- Describe design considerations for applications that use foundation models (FMs)
- Choose effective prompt engineering techniques
- Describe the training and fine-tuning process for FMs
- Describe methods to evaluate FM performance

### Guidelines for Responsible AI (14% of the exam)

- Explain the development of AI systems that are responsible
- Recognize the importance of transparent and explainable models

### Security, Compliance, and Governance for AI Solutions (14% of the exam)

- Explain methods to secure AI systems
- Recognize governance and compliance regulations for AI systems

## Exam mode and practice mode

| | Exam mode | Practice mode |
|---|---|---|
| Questions count | 65 | 10 - 75 |
| Time limit | 90 minutes | Optional, 1 - 210 minutes |
| Exam scope | 5 domains with the official questions ratio | Selected domains with the official questions ratio |
| Correct answers | After exam submission | After exam submission or after each answer |
| Question types | Single answer, multiple answer, ordering, and matching | Single answer, multiple answer or all types |
| Question hints | Never | Optional |
| Question domain | Revealed after exam submission | Revealed after submission or during the exam |
| Scoring | 15 of 65 questions do not count towards the result | Official AWS method or mathematical mean |

See all simulator features on the AWS Exam Simulator page: https://blowstack.com/practice-exams

## Technologies and concepts

- **Access Control**: Access control turns security requirements into enforceable permissions for people, applications, and cloud services. It limits unauthorized activity while preserving legitimate operations, including the ability to deliver logs, run health checks, and investigate incidents.
- **Adversarial Robustness**: Unlike ordinary reliability testing, adversarial robustness evaluates failures caused by intentional attempts to deceive or misuse a model. It matters because high accuracy on normal data does not establish safety under attack, and defenses must be evaluated alongside usefulness, explainability, latency, and cost.
- **Agentic AI**: An agentic system does more than produce a response: it can coordinate multiple steps, interact with external systems, and use results to decide what to do next. This makes it useful for tasks such as resolving support requests or investigating operational incidents, but it also requires stronger controls over permissions, reliability, cost, and human approval.
- **AI Agents**: In foundation model applications, an AI agent combines a model with instructions, context, and tools to complete tasks that may require several decisions and interactions with external systems. Agents matter because they can move beyond generating answers to performing business work, but their access, reliability, cost, and degree of autonomy require explicit controls.
- **AI Guardrails**: Guardrails help AI applications produce appropriate responses, protect sensitive information, and avoid unauthorized actions. They combine prompt engineering with independently enforced checks, permissions, and operational monitoring because model instructions alone cannot reliably enforce every requirement.
- **AI Transparency**: AI transparency gives users, engineers, reviewers, and auditors appropriate information about how an AI system is developed, operated, and controlled. It supports informed use, accountability, and compliance by documenting system behavior and responsibilities without unnecessarily exposing private data, intellectual property, or security-sensitive details.
- **Algorithmic Bias**: Algorithmic bias can arise from training data, labels, model objectives, prompts, retrieval sources, and the way an AI system is deployed. It matters because a technically accurate system can still discriminate, expose users to harm, create legal risk, and undermine customer trust.
- **Algorithmic Fairness**: Fairness requires choosing context-appropriate criteria for equitable treatment, examining data and model behavior, and addressing harmful disparities throughout the AI lifecycle. It matters because high overall accuracy can conceal systematically worse outcomes for particular populations, especially in consequential decisions such as hiring, lending, and healthcare.
- **APIs**: An application programming interface defines available operations, accepted inputs, responses, and error behavior. In the cloud, APIs enable applications and automation tools to provision resources, configure services, and process data, making them the foundation of programmatic access.
- **Audit Logging**: Audit logging creates an evidence trail showing who or what performed an action, when it occurred, which resource was affected, and what the outcome was. In cloud environments, effective auditing requires selecting complementary log sources and protecting their delivery, retention, and accessibility across accounts, Regions, and workloads.
- **AWS Compliance**: Compliance in AWS combines AWS-provided infrastructure assurances with customer responsibilities for data, applications, configurations, and operating procedures. Engineers must identify applicable requirements, verify service and geographic scope, implement controls, and retain evidence rather than assume that using AWS makes a workload compliant.
- **AWS shared responsibility model**: AWS and its customers have distinct responsibilities for protecting systems, data, and operations, with the boundary changing according to the services used. Understanding that boundary helps engineers implement appropriate controls and collect the right evidence instead of assuming that AWS compliance certifications automatically make a workload compliant.
- **Batch Inference**: Batch inference processes accumulated inputs on a schedule or when data becomes available, then stores the results for later use. It is useful when throughput and cost efficiency matter more than immediate responses, such as nightly customer scoring, document classification, or generating recommendations in advance.
- **Benchmark Datasets**: A benchmark dataset provides a consistent basis for evaluating foundation models using test inputs, reference answers where appropriate, and a defined scoring procedure. It helps engineers compare candidate models, detect regressions, and assess suitability for an application, but its value depends on how well the test cases represent real workloads.
- **BERTScore**: BERTScore compares the meaning represented by individual tokens rather than relying only on matching words or phrases. It is useful for evaluating foundation model outputs that can express an acceptable answer in different ways, but it does not independently establish factual correctness, safety, or instruction compliance.
- **Bias and Fairness**: Bias can enter an AI system through its data, labels, design choices, or operating environment, potentially creating unjustified differences in outcomes. Fairness means evaluating and addressing those differences using criteria appropriate to the application, affected people, and applicable requirements.
- **Bias-Variance Tradeoff**: Models can be inaccurate because they are too simple to capture important patterns or because they fit training-specific details that do not generalize. Managing this balance helps engineers reduce underfitting and overfitting, while responsible AI development also requires checking whether errors disproportionately affect particular demographic groups.
- **BLEU**: Bilingual Evaluation Understudy (BLEU) was developed for machine translation and is also used to evaluate foundation-model outputs when reference wording is a meaningful quality signal. It provides a reproducible measure of lexical overlap, but it does not directly establish factual accuracy, semantic equivalence, or usefulness.
- **Caching**: A cache serves repeated requests without always accessing the original data source or recomputing a result. In cloud architectures, caching can reduce latency, origin load, and data transfer, but it requires deliberate decisions about freshness, capacity, security, and failure behavior.
- **Chain-of-Thought Prompting**: The technique encourages a model to break a multistep problem into smaller reasoning steps, either through instructions or worked examples. It can improve performance on tasks such as arithmetic and constraint-based decisions, but its effectiveness depends on the model and task, and generated explanations do not guarantee correct results.
- **CI/CD**: Continuous integration validates frequently merged code changes through automated builds and tests, while continuous delivery or continuous deployment automates the path toward production. Together, these practices shorten feedback cycles, make releases more repeatable, and reduce the risks associated with large, manually executed changes.
- **Class Imbalance**: Unequal class frequencies can cause a classifier to favor common outcomes while missing rare but important ones, such as fraud or equipment failures. Responsible AI development requires examining class distributions alongside data quality, inclusivity, and subgroup coverage, then selecting training and evaluation methods that reflect real-world risks.
- **Classification Metrics**: Classification metrics describe different aspects of predictive performance, including overall correctness, false alarms, and missed positive cases. Choosing the right metric requires understanding class imbalance, the business cost of errors, and whether improvements produce measurable operational value.
- **Confidence Scoring**: Confidence scores help engineers manage uncertain AI outputs by connecting model predictions and supporting evidence to explicit decision thresholds. For generative AI, they can help identify potentially unsupported responses, but they do not prove factual correctness and should complement grounding, output validation, and human review.
- **Context Window**: The context window limits how much instruction, conversation history, retrieved information, and generated output a model can handle together. Engineers use this limit when selecting foundation models and designing prompts, retrieval pipelines, and conversation management so requests fit without sacrificing relevant information or exceeding cost and latency targets.
- **Context Windows**: A context window bounds the information available to a model when it generates a response, including instructions, conversation history, retrieved documents, and tool results. Context engineering selects and organizes that information so the model receives relevant evidence without exceeding its limits or introducing unnecessary cost, latency, and distraction.
- **Continued Pre-Training**: Continued pre-training builds on a foundation model rather than training a new model from scratch, typically using unlabeled documents from a target domain or a newer corpus. It can improve familiarity with specialized language and concepts, but requires careful data preparation, compute budgeting, and evaluation to avoid degrading existing capabilities.
- **Conversation Memory**: Conversation memory lets a generative AI application carry forward relevant messages, user preferences, decisions, and task progress instead of treating every request as unrelated. It is part of context engineering: selecting and organizing the information a foundation model needs for its next response while controlling cost, privacy, and accuracy.
- **Conversational AI**: Conversational AI combines language processing, dialogue management, and sometimes generative models to help people obtain information or complete tasks through conversation. It matters because users can express goals in natural language, clarify requests, and receive context-sensitive responses instead of navigating fixed forms or menus.
- **Cost management**: In the cloud, cost management connects resource consumption and pricing choices to business outcomes. It helps engineers compare deployment options, eliminate waste, and balance spending against performance, availability, security, and operational effort.
- **Cost-Benefit Analysis**: For cloud and AI/ML projects, cost-benefit analysis evaluates whether measurable improvements justify development, infrastructure, operational, and error-related costs compared with simpler alternatives. It helps engineers recognize when predictions provide useful business value and when deterministic software, established rules, or human judgment are more appropriate.
- **Data Cataloging**: A data catalog helps engineers discover datasets and determine whether they are appropriate for analytics, model training, evaluation, or retrieval. By documenting provenance, classifications, and usage restrictions, it supports trustworthy AI development, security reviews, and compliance without replacing access controls or data validation.
- **Data Curation**: For foundation model fine-tuning, data curation turns raw source material into governed, representative examples that teach the desired behavior. It matters because incorrect labels, sensitive information, duplicated examples, and coverage gaps can undermine model quality even when the training process is technically correct.
- **Data governance**: Data governance makes data use accountable by defining ownership, permitted purposes, quality expectations, access rules, and retention requirements. For foundation model fine-tuning, it helps ensure that training examples are authorized, relevant, representative, consistently labeled, and traceable to their sources.
- **Data Integrity**: Data integrity combines controls that prevent improper changes with checks that detect corruption, invalid transformations, and tampering. In cloud and AI systems, it helps ensure that training data, inference inputs, and outputs remain trustworthy and traceable rather than merely accessible or confidential.
- **Data Labeling**: Labels connect input data to the outcomes or interpretations a model should learn, such as a transaction's fraud status or the location of an object in an image. Reliable labeling helps define the learning task, improves training and evaluation quality, and makes model errors easier to investigate.
- **Data Loss Prevention**: Data Loss Prevention combines data classification, inspection, access restrictions, and enforcement at the points where information is stored, processed, or transmitted. In cloud and generative AI applications, it helps prevent confidential records, credentials, and proprietary content from escaping through prompts, retrieved context, model responses, logs, or external tools.
- **Data Poisoning**: Attackers can poison training datasets, fine-tuning examples, feedback records, or external knowledge sources consumed by AI applications. Preventing poisoning requires controls over data provenance, modification permissions, validation, and evaluation because well-written prompts alone cannot establish that the underlying information is trustworthy.
- **Data Provenance**: Data provenance connects a dataset to its sources, collection conditions, permissions, processing steps, and downstream uses. In AI systems, it helps engineers assess whether training and evaluation data are appropriate, investigate quality or bias issues, reproduce results, and explain the evidence behind model development.
- **Data Quality**: Data quality determines whether engineers can rely on data for analytics, application behavior, and AI training or inference. In secure data engineering, quality assessments work alongside privacy protections, access controls, and integrity checks to reduce unreliable outputs, unauthorized changes, and inappropriate use of sensitive information.
- **Data Representativeness**: Representative data provides appropriate coverage of the populations and situations an AI system will encounter, including relevant minority groups and uncommon but important cases. It supports reliable evaluation and responsible development by reducing coverage gaps that can hide poor performance or unequal outcomes.
- **Data Residency**: Data residency connects the physical location of data to legal, contractual, and organizational requirements. Cloud architectures must account for primary databases, replicas, backups, logs, and downstream systems because choosing a deployment Region alone does not establish compliance.
- **Data Retention**: Retention policies balance business needs, recovery requirements, legal obligations, privacy, and storage cost. In the cloud, engineers implement them through lifecycle rules, backup schedules, retention periods, deletion controls, and exceptions such as legal holds.
- **Determinism**: Determinism concerns repeatability, not whether an answer comes from explicit rules or a learned model, and not whether that answer is correct. When choosing between conventional software and AI/ML, distinguish repeatable execution from guaranteed compliance with a specification: a model can repeatedly produce the same prediction without satisfying a requirement for an exact calculation or mandatory business rule.
- **Embeddings**: An embedding model maps inputs such as text or images into a vector space where related inputs are intended to have similar representations. Embeddings support semantic search, recommendations, clustering, and retrieval-augmented generation by helping applications find relevant information beyond exact keyword matches.
- **Encryption**: Encryption helps prevent unauthorized parties from understanding data stored in cloud systems or transmitted over networks. Its effectiveness depends on protecting cryptographic keys, controlling decryption permissions, and applying encryption wherever sensitive data travels or persists.
- **Encryption at Rest**: Encryption at rest protects data stored in objects, disks, databases, snapshots, and backups from disclosure when the underlying storage is accessed without the required keys. In cloud environments, it combines storage encryption with key management and access controls, but it does not replace authorization, backups, or encryption in transit.
- **Encryption in Transit**: Encryption in transit prevents observers on a network path from reading protected data and, with correctly configured secure protocols, helps detect tampering and authenticate communicating endpoints. In cloud architectures, it applies to internet traffic, internal service communication, and hybrid or multi-cloud connections, independently of encryption at rest.
- **End-to-End Testing**: End-to-end testing checks whether an application achieves the intended user outcome when its components work together under realistic conditions. For foundation model applications, it evaluates the combined behavior of prompts, retrieval, models, tools, safeguards, and workflows rather than relying on model benchmarks alone.
- **Environmental Sustainability**: In cloud computing and AI, environmental sustainability means considering energy consumption, greenhouse gas emissions, water use, and hardware resource demands when designing and operating workloads. Responsible model selection weighs these impacts alongside accuracy, safety, fairness, latency, privacy, and cost rather than automatically choosing the largest or most capable model.
- **Experiment Tracking**: Experiment tracking connects each model result to the data, code, parameters, and execution environment that produced it. It supports repeatable development, informed model selection, and traceability as models move toward production and are later retrained.
- **Feature Engineering**: Features are the measurable attributes a model uses to make predictions, such as transaction amount, account age, or recent purchase frequency. Feature engineering converts available data into suitable model inputs, improving predictive usefulness while controlling leakage, complexity, and production processing costs.
- **Feedback Loops**: In responsible AI, feedback loops connect model decisions and explanations with the people affected by them, allowing users to question results, report problems, and request review. They support human-centered design by turning those signals into evaluated changes to models, explanations, interfaces, or operating policies.
- **Fine-Tuning**: Fine-tuning reuses the capabilities of an existing model rather than learning everything from scratch. It can improve task accuracy, response style, and output consistency, but introduces training, evaluation, deployment, and maintenance costs that must be compared with prompting and retrieval-based approaches.
- **Foundation Models**: Foundation models provide reusable capabilities for working with language, images, audio, and other data through prompting, retrieval, or additional training. Choosing between a foundation model and traditional task-specific machine learning depends on the workload, available data, explainability requirements, regulatory obligations, and operational constraints.
- **Generative AI**: Generative AI creates outputs from learned patterns and supplied context rather than only assigning labels or predicting numerical values. In cloud applications, it supports capabilities such as document summarization, conversational assistance, and code generation, but requires evaluation and controls because plausible outputs are not necessarily correct.
- **Hallucinations**: Generative AI can produce convincing answers without establishing that their claims are true or supported by the supplied evidence. Engineers must account for this limitation because fabricated facts, citations, instructions, or commitments can cause operational failures, legal exposure, and loss of customer trust.
- **Human-Centered Design**: For AI systems, human-centered design makes predictions and decisions understandable, useful, and open to appropriate human challenge or correction. It combines user research, accessible explanations, feedback mechanisms, and meaningful oversight so that technical performance supports real human needs rather than becoming the only measure of success.
- **Human-in-the-Loop**: Human-in-the-Loop combines AI’s ability to process information at scale with people’s ability to interpret context, resolve ambiguity, and exercise judgment. It supports assisted decision-making and responsible automation by giving people meaningful opportunities to correct outputs, challenge decisions, and improve system behavior.
- **Human-in-the-Loop Evaluation**: Human-in-the-loop evaluation combines structured human judgment with repeatable testing to measure how well foundation models meet application requirements. It complements automated metrics by assessing qualities such as usefulness, factual correctness, safety, and appropriateness that a single numerical benchmark may not adequately capture.
- **Hyperparameter Tuning**: Hyperparameters control how a machine learning model is structured or trained, unlike model parameters such as weights that are learned from data. Tuning compares training runs with different settings to find a suitable configuration while balancing predictive quality, training cost, and operational requirements.
- **In-Context Learning**: A model uses the current prompt and conversation context to infer the task, recognize patterns, and generate an appropriate response. This enables rapid customization without a training job, but quality depends on the context supplied, and longer prompts can increase inference cost and latency.
- **Inference Latency**: Inference latency determines how quickly an AI-powered application responds and whether it meets user expectations or processing deadlines. For generative AI, input length, generated output length, model choice, and available serving capacity affect latency, while token-based pricing links much of that work to inference cost.
- **Instruction Hierarchy**: Foundation model applications distinguish higher-priority application rules from lower-priority user requests and external content. This distinction helps preserve intended behavior when prompts conflict, but its effectiveness depends on the model, API, and application design rather than on wording alone.
- **Instruction Tuning**: Instruction tuning adapts a foundation model’s behavior using examples of tasks and acceptable answers, often including supporting context or conversation history. It helps engineers improve instruction following, response structure, and task consistency without training a model from scratch.
- **Intellectual Property Rights**: In cloud and generative AI systems, intellectual property rights affect the use of training data, prompts, retrieved documents, models, software, and generated outputs. Engineers must combine licensing reviews, access controls, provenance records, and output review to reduce infringement claims and protect proprietary assets without assuming that a managed service provides legal clearance.
- **Interpretability-Performance Tradeoff**: Models that are easy to understand may sacrifice predictive capability on some tasks, while more complex models can require additional methods to explain their behavior. This tradeoff is not inevitable: engineers must measure model quality, explanation usefulness, safety, latency, and cost rather than assume that complexity always improves results.
- **Iterative Prompt Refinement**: A prompt is treated as a testable application component rather than a one-time question. Engineers examine outputs, identify failure patterns, and make targeted revisions to improve accuracy, relevance, consistency, and safety without necessarily changing model weights.
- **Jailbreaking**: Jailbreaking attempts use adversarial inputs to make a model produce responses or take actions that its safeguards are intended to prevent. For cloud applications, it is a prompt-engineering risk that requires layered controls because carefully written instructions alone cannot guarantee safe behavior.
- **Key Performance Indicators**: A key performance indicator connects a business objective to a defined measurement, target, and review period. For generative AI applications, KPIs help engineers distinguish technical improvements from business value by measuring outcomes such as efficiency, conversion, revenue, accuracy, and return on investment.
- **Knowledge Distillation**: Knowledge distillation transfers learned behavior from a teacher model to a student through training signals such as generated answers or output probabilities. It can reduce inference latency, memory requirements, and serving costs, but the student must be evaluated because it may lose capabilities or inherit the teacher’s errors.
- **Label Quality**: Labels define expected outcomes for supervised learning, human-preference training, and many AI evaluations. Incorrect, ambiguous, outdated, or systematically biased labels can teach models the wrong behavior and make evaluation results misleading, even when the underlying data and training pipeline are technically sound.
- **Least privilege**: Least privilege reduces the damage that compromised credentials, software defects, and operational mistakes can cause. In cloud environments, it applies to identity permissions, workload roles, administrative access, network connectivity, and organizational guardrails, and requires ongoing review as requirements change.
- **License Entitlements**: A license entitlement describes what an organization is authorized to use, including limits such as users, processor cores, features, duration, and permitted deployment environments. In the cloud, entitlement management connects software purchases to deployment and consumption so teams can control costs, enforce contractual limits, and demonstrate compliance.
- **LLM-as-a-Judge**: An evaluator LLM receives a task, a candidate response, and a scoring rubric, optionally with reference answers or supporting evidence, and returns a score, label, or preference. This approach helps engineers evaluate qualities such as relevance, groundedness, and instruction following at scale, but its judgments require calibration and human oversight.
- **Machine learning**: Machine learning turns examples and experience into models that can handle tasks such as detecting suspicious transactions, classifying images, and understanding text. Cloud platforms provide scalable infrastructure and managed APIs for applying these models, but useful results still depend on appropriate data, evaluation, security, and ongoing monitoring.
- **Management and governance**: Management handles activities such as provisioning, monitoring, configuration, and remediation, while governance defines who may act, which standards apply, and how compliance is demonstrated. Together, they help organizations give engineers useful autonomy without losing accountability, cost visibility, or control over risk.
- **Model Compression**: Model compression makes trained models more practical to deploy by reducing their resource requirements through techniques such as quantization, pruning, and knowledge distillation. It supports responsible model selection by helping engineers balance accuracy, safety, latency, cost, and environmental impact rather than choosing the largest available model by default.
- **Model Context Protocol**: MCP provides a common interface between an AI application's integration layer and external systems, reducing the need for custom connectors for every application–tool combination. In agentic applications, it helps agents discover available capabilities and access relevant information or request actions, while leaving reasoning, orchestration, and authorization enforcement to the surrounding system.
- **Model Distillation**: A student model learns to reproduce a teacher model’s behavior using generated examples, output probabilities, or other training signals. For foundation model applications, distillation can exchange upfront data-generation and training costs for lower ongoing serving costs, provided the student meets the application’s quality requirements.
- **Model Drift**: A model learns patterns from historical data, but those patterns may no longer represent the environment in which it operates. Monitoring for drift helps teams identify changing conditions, investigate performance problems, and decide whether to fix data pipelines, adjust the model, or retrain it.
- **Model Evaluation**: Model evaluation uses representative data, defined metrics, and human judgment to determine whether a model meets technical and business requirements. It supports model selection and release decisions throughout the foundation model lifecycle, helping engineers balance output quality, safety, latency, cost, and environmental impact.
- **Model Explainability**: Explainability helps engineers, reviewers, and affected users understand which inputs and model behaviors contributed to an outcome. It matters when choosing between traditional machine learning and foundation models because regulatory obligations, decision risk, and operational constraints can require explanations that a highly complex model cannot reliably provide.
- **Model Interpretability**: Interpretability helps engineers, reviewers, and affected users understand model behavior rather than treating outputs as unexplained decisions. It supports debugging, accountability, and risk assessment, but an understandable explanation does not prove that a model is accurate, fair, or causally correct.
- **Model Robustness**: A robust model behaves reliably beyond clean training examples, including when production inputs contain errors, unfamiliar variations, or adversarial manipulation. Robustness is a feature of responsible AI because unexpected model failures can create harmful decisions, unreliable automation, and loss of user trust.
- **Model Transparency**: Model transparency makes an AI system understandable enough for stakeholders to evaluate its suitability and hold its operators accountable. It includes documentation and visibility into how the model works, while explainability specifically addresses why it produces particular outputs. These qualities support debugging, informed deployment decisions, and meaningful review of consequential predictions.
- **Multi-Agent Orchestration**: Multiple specialized agents can collaborate on work that requires different responsibilities, data sources, or decision steps. Orchestration defines how tasks are assigned, how agents communicate, and how the application controls state, failures, permissions, and completion.
- **Multimodal Generation**: A multimodal generative system can use information in one or more modalities to produce content in another modality or a combination of modalities. This enables applications such as image-grounded assistants, narrated videos, and visual content creation, but requires engineers to match model capabilities, quality controls, and infrastructure to each data type.
- **Multimodality**: Multimodal applications combine information types to support tasks such as answering questions about photographs, summarizing recorded meetings, or generating images from text. When selecting a foundation model, engineers must verify its specific input and output capabilities and balance them against accuracy, cost, latency, security, and operational requirements.
- **N-Gram Overlap**: N-gram overlap is a lexical comparison technique used to evaluate generated text against expected outputs. It underpins metrics such as BLEU and ROUGE-N, providing inexpensive, repeatable signals for translation and summarization while requiring complementary checks for meaning, factual accuracy, and usefulness.
- **Natural Language Processing**: Natural language processing (NLP) turns language in documents, messages, and conversations into information that software can use, and can also produce language for users. It supports applications such as ticket routing, sentiment analysis, translation, document summarization, and conversational assistants, reducing manual processing of unstructured information.
- **Negative Prompting**: Negative prompts express exclusions, such as avoiding unsupported claims, personal information, unwanted visual elements, or a particular writing style. They help define acceptable outputs, but their effectiveness depends on the model and wording, so they should complement positive instructions, validation, and safety controls.
- **Network Isolation**: Network isolation creates controlled communication boundaries between workloads, environments, and external networks. During incident response, it helps contain compromised resources, interrupt attacker communications, and protect unaffected systems while engineers investigate, eradicate threats, and restore trusted workloads.
- **Neural Networks**: A neural network transforms inputs into outputs using learned weights, biases, and mathematical operations. Neural networks support applications such as image recognition, speech transcription, prediction, and generative AI, making them an important foundation for understanding modern cloud AI workloads.
- **Nondeterminism**: In generative AI, repeated requests can produce different answers because of sampling during generation and variability in the execution environment. This flexibility supports diverse outputs, but it complicates testing, reproducibility, auditing, and business processes that require consistent results.
- **Nucleus Sampling**: Nucleus sampling dynamically limits the next-token choices available to a generative model, balancing response diversity against the exclusion of unlikely continuations. Engineers configure it during inference to influence output variability, usually alongside controls such as temperature and maximum output tokens.
- **Online Inference**: Online inference connects a trained model to a live application so that new inputs can produce timely outputs, such as fraud scores, recommendations, or generated text. It matters when decisions depend on current requests rather than predictions computed in advance, requiring careful management of latency, availability, capacity, and cost.
- **Open-Source Models**: Open-source models let engineers build on existing model implementations and pretrained capabilities instead of developing every component from scratch. They can improve inspectability, customization, and deployment control, but teams must verify which artifacts are available, what the licenses permit, and whether the model meets their quality and responsible AI requirements.
- **Output Validation**: Generative AI responses must be treated as untrusted data, even when they come from a trusted model endpoint. Output validation helps prevent unsafe content, sensitive-data disclosure, injection attacks, and unauthorized actions by checking responses at the boundaries where applications consume them.
- **Overfitting**: An overfit model can achieve excellent training results without learning relationships that generalize reliably. Preventing overfitting improves prediction accuracy and supports responsible AI, especially when overall performance hides unreliable predictions for particular demographic groups.
- **Post-Hoc Explainability**: Complex models can produce useful predictions while making it difficult for people to understand how inputs influence outputs. Post-hoc explainability adds explanations after training to support debugging, review, and communication, but those explanations must be evaluated rather than treated as complete descriptions of the model’s reasoning.
- **Pre-Training**: For foundation models, pre-training establishes general capabilities such as language prediction, visual representation, or relationships across modalities. It matters because the training data, learning objective, and compute budget strongly influence what downstream applications can achieve, even after fine-tuning.
- **Privacy-Enhancing Technologies**: Privacy-enhancing technologies protect individuals and organizations by limiting what sensitive information is collected, revealed, or inferred during data processing. In cloud-based AI systems, they complement access controls, encryption, data-quality checks, and integrity protections to reduce privacy risks throughout the data lifecycle.
- **Prompt Context**: Context gives a model task-specific facts, conversation history, reference material, or situational details that may not be available from its learned parameters. Selecting relevant, trustworthy context improves response quality, but requires attention to token limits, cost, privacy, and the distinction between supporting information and instructions.
- **Prompt Injection**: Language models process instructions and reference material together, creating opportunities for attackers to disguise malicious commands as ordinary content. In cloud applications, successful injection can manipulate responses, expose sensitive context, or trigger unauthorized tool actions, so defenses must combine model-level guardrails with independently enforced security controls.
- **Prompt Templates**: A prompt template defines the repeatable parts of a model request, such as the task, context boundaries, examples, and expected response format, while leaving placeholders for request-specific data. Templates help engineers standardize application behavior, evaluate changes systematically, and reduce duplicated prompt-building logic without retraining the model.
- **Real-Time Inference**: Real-time inference turns a trained model into a prediction capability available during an application workflow, such as evaluating a payment or generating a chat response. Engineers typically expose it through a managed model API or a self-hosted API, balancing response time, availability, cost, and operational control.
- **Reinforcement Learning**: An agent observes its environment, takes actions, and uses reward feedback to improve its decision-making policy. Reinforcement learning matters when decisions affect future outcomes, such as robot navigation, resource allocation, or inventory replenishment, rather than simply predicting a label.
- **Reinforcement Learning from Human Feedback**: RLHF turns human evaluations of model outputs into a learning signal, commonly by training a reward model and then optimizing a foundation model against that reward. It helps improve instruction following, usefulness, and safety, but its effectiveness depends on representative feedback, clear labeling criteria, and careful evaluation.
- **Retrieval-Augmented Generation**: RAG connects a foundation model to information outside its learned parameters, such as enterprise documents, product catalogs, or operational records. It helps applications answer questions using organization-specific or recently updated information without retraining the model, although accuracy still depends on retrieval quality and how the model uses the evidence.
- **Return on Investment**: ROI helps engineers and business stakeholders determine whether an AI, ML, or cloud initiative produces enough measurable value to justify its costs. Unlike model-quality metrics, ROI connects technical performance to financial outcomes such as reduced operating expenses, increased contribution profit, or avoided losses.
- **ROUGE**: ROUGE stands for Recall-Oriented Understudy for Gisting Evaluation and is commonly used to evaluate summaries and other generated text against reference answers. It provides inexpensive, repeatable measurements for comparing foundation models, prompts, or fine-tuned versions, but does not independently establish factual correctness or semantic equivalence.
- **Security**: Security combines technical controls, operational processes, and clearly assigned responsibilities to preserve confidentiality, integrity, and availability. In the cloud, providers secure the underlying infrastructure while customers configure and operate their workloads securely, using capabilities such as identity controls, encryption, network protection, and threat detection.
- **Selective Disclosure**: In responsible AI, selective disclosure balances meaningful transparency with the risks of exposing sensitive data, proprietary information, or details that enable misuse. It helps engineers provide useful explanations and accountability without assuming that transparency requires publishing every model artifact or operational control.
- **Semantic Search**: Semantic search helps users find relevant information even when their queries use different wording from the source material. It commonly uses machine-learning embeddings and relevance models, making it useful for enterprise search, customer support, and retrieving evidence for generative AI applications.
- **Serverless**: Serverless shifts infrastructure operations to the cloud provider so engineers can focus on application behavior, data, and configuration. It can reduce operational work and align costs with usage, but scaling behavior, billing, execution limits, and customer responsibilities vary by service.
- **Source Citation**: Source citation connects information to identifiable evidence, such as a dataset version, document passage, or model-development record. In cloud AI systems, it supports verification, reproducibility, and governance by helping engineers establish where information came from, how it changed, and whether its use is appropriate.
- **Structured Data**: Structured data commonly appears as records and tables containing values such as customer identifiers, transaction amounts, and timestamps. Its predictable organization makes it useful for querying, validation, analytics, and machine learning, but it does not automatically make the data accurate, labeled, or ready for model training.
- **Subgroup Analysis**: An AI system can perform well on average while producing less accurate, less reliable, or more harmful results for particular populations or operating conditions. Subgroup analysis compares relevant groups using appropriate quality and fairness metrics, helping engineers detect potential bias, investigate label quality, and monitor whether improvements benefit all affected users.
- **Supervised Learning**: A supervised model learns the relationship between input features and known target values, such as transaction details paired with fraud labels or property attributes paired with sale prices. It supports classification and regression, making it useful when historical outcomes are available and engineers need predictions for new cases.
- **Task Completion Rate**: Task completion rate measures whether an AI application helps users accomplish their intended goals, rather than merely producing plausible responses. It connects foundation model performance to business objectives and should be evaluated alongside user satisfaction, cost per interaction, latency, and safety.
- **Task Planning**: In foundation-model applications, task planning helps an AI agent determine what actions are needed to achieve a business goal rather than merely generate a response. It enables multistep automation, but reliable implementations must constrain tool access, validate results, and require approval for consequential actions.
- **Temperature Sampling**: Lower temperatures favor more probable tokens, while higher temperatures give less probable tokens a greater chance of being selected. This helps engineers balance consistency and variety, but it does not guarantee factual accuracy, enforce an output format, or directly control response length.
- **Text Embeddings**: An embedding model maps text into a vector space where related passages tend to have similar representations. These vectors support semantic search and retrieval augmented generation (RAG), helping applications find relevant external information even when a question and its answer use different wording.
- **Text Generation**: Text generation uses learned language patterns to produce outputs such as summaries, translations, conversational responses, and code. In cloud applications, it enables engineers to build assistants and content-processing workflows without writing explicit rules for every possible response, but its outputs require controls for accuracy, security, and suitability.
- **Tokenization**: Tokens can represent whole words, word fragments, punctuation, whitespace, or other character sequences, depending on the model’s tokenizer. Tokenization determines how much content fits into a model request and directly affects token-metered inference costs, processing requirements, and response latency.
- **Tool Calling**: Tool calling connects a generative AI model to capabilities such as database queries, calculations, business APIs, and infrastructure operations. It enables agents to use current information and perform actions, while application code remains responsible for execution, authorization, validation, and error handling.
- **Tool Use**: A foundation model can select a tool and generate structured arguments, while the application or managed agent runtime validates and executes the request. This connects language-based reasoning to current data and business systems, enabling AI agents to complete tasks rather than only generate text.
- **Total Cost of Ownership**: Total Cost of Ownership (TCO) includes direct expenses, such as hardware or cloud usage, and indirect expenses, such as administration, facilities, and migration effort. It helps engineers compare deployment options fairly and determine whether moving a workload to the cloud will reduce overall costs rather than simply change how those costs are billed.
- **Train-Validation-Test Split**: Training data teaches a model, validation data guides development decisions, and test data provides an independent final evaluation. Keeping these roles separate helps engineers detect overfitting, prevent data leakage, and estimate how well a model will generalize to production inputs.
- **Trajectory Evaluation**: A trajectory includes actions such as retrieving documents, selecting tools, supplying arguments, handling results, and deciding when to stop. Evaluating this sequence helps engineers detect unsafe actions, incorrect dependencies, unnecessary work, and intermediate failures that a plausible final response can conceal.
- **Transfer Learning**: Instead of training a model from scratch, engineers start with a pretrained model and reuse its learned representations, adapt its parameters, or both. This can reduce training time and data requirements, making it practical to specialize foundation models for business tasks, although success depends on model suitability and training-data quality.
- **Transformer Architecture**: Transformers build context-dependent representations of tokens by learning which other tokens are relevant to their interpretation. They underpin many modern large language models and foundation models, making their capabilities, context limits, and computational costs important considerations when selecting and deploying generative AI in the cloud.
- **Underfitting**: An underfit model typically performs poorly on both training data and unseen data because its assumptions, features, capacity, or training process are inadequate for the task. Recognizing underfitting helps engineers improve predictive accuracy while checking whether errors disproportionately affect particular demographic groups.
- **Unstructured Data**: Unstructured data contains information that typically requires interpretation or transformation before conventional analytics or machine learning can use it effectively. It matters because documents, conversations, photographs, and recordings provide signals that fixed database fields often do not capture.
- **Unsupervised Learning**: Unsupervised learning uses similarities, differences, and statistical relationships within data to identify useful structure rather than learn from labeled examples. It supports tasks such as customer segmentation, exploratory analysis, and anomaly detection when engineers do not have known outcomes for training.
- **Usage-Based Pricing**: Cloud costs are calculated from billable quantities such as instance runtime, stored data, requests, and data transfer rather than solely from an upfront infrastructure purchase. Understanding these meters helps engineers compare purchasing options, estimate total workload costs, and avoid paying for idle resources or unsuitable commitments.
- **User Satisfaction**: For AI applications, user satisfaction measures whether people find the experience useful, trustworthy, understandable, and responsive. It connects foundation model evaluation to business objectives, but must be assessed alongside task completion, factual correctness, safety, and operating cost.
- **Vector Search**: Embedding models represent text, images, audio, or other data as vectors whose relative positions can reflect meaningful relationships. Vector search uses these representations to find related content even when exact words differ, supporting semantic retrieval, recommendations, and retrieval-augmented generation.
- **Version Control**: Version control creates a traceable history for artifacts such as source code, infrastructure definitions, configuration, and AI prompts. It helps engineers collaborate safely, connect deployed behavior to specific revisions, and recover from changes that introduce errors or regressions.

## Services on the exam

The 42 AWS services the AIF-C01 exam covers.

[Amazon Augmented AI (Amazon A2I)](https://blowstack.com/tools/cloud-services-hub/amazon-augmented-ai-(amazon-a2i)), [Amazon Aurora](https://blowstack.com/tools/cloud-services-hub/amazon-aurora), [Amazon Bedrock](https://blowstack.com/tools/cloud-services-hub/amazon-bedrock), [Amazon Bedrock AgentCore](https://blowstack.com/tools/cloud-services-hub/amazon-bedrock-agentcore), [Amazon CloudWatch](https://blowstack.com/tools/cloud-services-hub/amazon-cloudwatch), [Amazon Comprehend](https://blowstack.com/tools/cloud-services-hub/amazon-comprehend), [Amazon EC2](https://blowstack.com/tools/cloud-services-hub/amazon-ec2), [Amazon Inspector](https://blowstack.com/tools/cloud-services-hub/amazon-inspector), [Amazon Lex](https://blowstack.com/tools/cloud-services-hub/amazon-lex), [Amazon Macie](https://blowstack.com/tools/cloud-services-hub/amazon-macie), [Amazon Neptune](https://blowstack.com/tools/cloud-services-hub/amazon-neptune), [Amazon OpenSearch Service](https://blowstack.com/tools/cloud-services-hub/amazon-opensearch-service), [Amazon Personalize](https://blowstack.com/tools/cloud-services-hub/amazon-personalize), [Amazon Q](https://blowstack.com/tools/cloud-services-hub/amazon-q), [Amazon Q Business](https://blowstack.com/tools/cloud-services-hub/amazon-q-business), [Amazon QuickSight](https://blowstack.com/tools/cloud-services-hub/amazon-quicksight), [Amazon RDS](https://blowstack.com/tools/cloud-services-hub/amazon-rds), [Amazon Rekognition](https://blowstack.com/tools/cloud-services-hub/amazon-rekognition), [Amazon S3](https://blowstack.com/tools/cloud-services-hub/amazon-s3), [Amazon SageMaker](https://blowstack.com/tools/cloud-services-hub/amazon-sagemaker), [Amazon SageMaker Ground Truth](https://blowstack.com/tools/cloud-services-hub/amazon-sagemaker-ground-truth), [Amazon Textract](https://blowstack.com/tools/cloud-services-hub/amazon-textract), [Amazon Transcribe](https://blowstack.com/tools/cloud-services-hub/amazon-transcribe), [Amazon Translate](https://blowstack.com/tools/cloud-services-hub/amazon-translate), [Amazon VPC](https://blowstack.com/tools/cloud-services-hub/amazon-vpc), [AWS Artifact](https://blowstack.com/tools/cloud-services-hub/aws-artifact), [AWS CloudTrail](https://blowstack.com/tools/cloud-services-hub/aws-cloudtrail), [AWS CodePipeline](https://blowstack.com/tools/cloud-services-hub/aws-codepipeline), [AWS Config](https://blowstack.com/tools/cloud-services-hub/aws-config), [AWS Cost Explorer](https://blowstack.com/tools/cloud-services-hub/aws-cost-explorer), [AWS Data Exchange](https://blowstack.com/tools/cloud-services-hub/aws-data-exchange), [AWS Glue](https://blowstack.com/tools/cloud-services-hub/aws-glue), [AWS Glue DataBrew](https://blowstack.com/tools/cloud-services-hub/aws-glue-databrew), [AWS Identity and Access Management (IAM)](https://blowstack.com/tools/cloud-services-hub/aws-identity-and-access-management-(iam)), [AWS Key Management Service (AWS KMS)](https://blowstack.com/tools/cloud-services-hub/aws-key-management-service-(aws-kms)), [AWS Lake Formation](https://blowstack.com/tools/cloud-services-hub/aws-lake-formation), [AWS Lambda](https://blowstack.com/tools/cloud-services-hub/aws-lambda), [AWS PrivateLink](https://blowstack.com/tools/cloud-services-hub/aws-privatelink), [AWS Step Functions](https://blowstack.com/tools/cloud-services-hub/aws-step-functions), [AWS Trusted Advisor](https://blowstack.com/tools/cloud-services-hub/aws-trusted-advisor), [AWS Well-Architected Tool](https://blowstack.com/tools/cloud-services-hub/aws-well-architected-tool), [Kiro](https://blowstack.com/tools/cloud-services-hub/kiro)

## Concepts on the exam

The 140 cloud concepts the AIF-C01 exam covers, each explained with the AWS services for it: https://blowstack.com/tools/cloud-concepts-hub?certification=aws-certified-ai-practitioner

## Other practice exams

- [SAP-C03 AWS Certified Solutions Architect - Professional](https://blowstack.com/practice-exams/aws-certified-solutions-architect-professional)
- [SAA-C03 AWS Certified Solutions Architect - Associate](https://blowstack.com/practice-exams/aws-certified-solutions-architect-associate)
- [CLF-C02 AWS Certified Cloud Practitioner](https://blowstack.com/practice-exams/aws-certified-cloud-practitioner)
- [DOP-C02 AWS Certified DevOps Engineer - Professional](https://blowstack.com/practice-exams/aws-certified-devops-engineer-professional)
- [DEA-C01 AWS Certified Data Engineer - Associate](https://blowstack.com/practice-exams/aws-certified-data-engineer-associate)
- [AIP-C01 AWS Certified Generative AI Developer - Professional](https://blowstack.com/practice-exams/aws-certified-generative-ai-developer-professional)
- [MLA-C02 (beta) AWS Certified Machine Learning Engineer - Associate](https://blowstack.com/practice-exams/aws-certified-machine-learning-engineer-associate)
- [SOA-C03 AWS Certified CloudOps Engineer - Associate](https://blowstack.com/practice-exams/aws-certified-cloudops-engineer-associate)
- [SCS-C03 AWS Certified Security - Specialty](https://blowstack.com/practice-exams/aws-certified-security-specialty)
- [AIB-C01 (beta) AWS Certified AI Business Strategist](https://blowstack.com/practice-exams/aws-certified-ai-business-strategist)
- [ANS-C01 AWS Certified Advanced Networking - Specialty](https://blowstack.com/practice-exams/aws-certified-advanced-networking-specialty)
- [DVA-C02 AWS Certified Developer - Associate](https://blowstack.com/practice-exams/aws-certified-developer-associate)
