AWS Certified AI Practitioner (AIF-C01) Course

Machine learning (ML) learns patterns from data instead of requiring a person to write every decision rule. To recognize unwanted email, for example, we can learn from examples of unwanted and legitimate messages. A model represents the learned patterns and uses them to turn inputs into outputs.

Machine learning is an approach within artificial intelligence in which models learn patterns from data rather than relying entirely on hand-written decision rules. It is a broad category, not the name of one model design. A model family is a group of models that share a basic structure, such as decision trees or neural networks.

A neural network is a mathematical model that transforms numerical inputs into outputs. For example, it can take numbers representing image pixels and produce a score for a manufacturing defect. Learning an input-to-output mapping means learning how those input values should influence the result, rather than storing a separate answer for every possible image.

Computer vision is the field of AI that extracts information from images and videos. A computer receives numbers representing pixels, not an understanding of what a picture shows. Computer vision connects those visual patterns to useful descriptions.

Amazon Comprehend is a natural language processing service. Natural language processing (NLP) applies computing techniques to human language. Comprehend uses machine learning to analyze text such as reviews, messages, and documents.

Machine learning (ML) describes learning patterns from data rather than relying only on hand-written rules. It answers a question about how a model gains its capabilities. A model can learn from past emails to recognize unwanted messages, for example.

Machine learning produces models that learn patterns from data. A foundation model is a machine learning model with broad initial training that provides a starting point for multiple applications. Foundation models are therefore part of machine learning, not an alternative to it.

Natural language processing (NLP) is AI concerned with human language. An NLP task might identify the sentiment in a review or produce a summary of it. Both involve language, but they ask for different kinds of results: an analysis of existing text versus newly generated content.

In AI, systematic bias means patterns in data, predictions, or decisions that can disadvantage particular people or groups. A speech recognition model may repeatedly make more errors for speakers with certain accents. That pattern matters because it can make the service less useful or accessible to those speakers. One isolated mistake does not, by itself, establish systematic bias.

Artificial intelligence (AI) is the broad field of building systems that perform tasks associated with intelligence, such as reasoning, perception, planning, and language processing. Producing content is one possible use of AI, not its defining purpose. AI also includes systems that follow explicitly programmed rules.

An AI agent pursues a goal by choosing and taking actions. A goal describes the desired outcome; an action is a step toward it. For example, “help this customer request an eligible return” is a goal. Looking up the order is one possible action, not the whole task.

Generative AI produces content, such as text, images, or code. A system might draft a refund explanation or write a plan for arranging a meeting. The output is content: writing a plan does not mean the system has carried it out.

Batch inference applies a trained model to a collected dataset and saves predictions for later use. For example, a job can score customer records for a campaign. The campaign uses the saved scores after the job finishes.

Online inference means using a deployed model to handle incoming application requests on demand. The application sends input to a model-serving endpoint, an address through which it can request a prediction. The output might be a fraud score, a product ranking, or a generated answer. Running inference normally does not retrain the model.

Inference means using an already-trained model to produce a prediction or generated output. Real-time and batch inference both do this; neither is model training. Their main difference is when the result is needed.

Inference means applying a trained model to input to produce an output, such as a recommendation or a risk score. Batch inference processes a collected dataset as a job, with predictions available for later use. Online inference processes incoming application requests on demand.

Inference means using a trained model to produce an output from an input, such as predicting a category for an image. Serverless inference runs that work without requiring you to provision and maintain individual compute servers.

Structured data organizes information into defined fields with consistent meanings. A schema describes those fields and their data types, such as text, number, or timestamp. This gives applications a shared way to interpret the data.

Structured data organizes information into predefined fields with consistent meanings and types. For example, each product record might have a product identifier, a price, and a category. You can compare prices directly because the price field has the same meaning in every record.

Unstructured data includes the meaningful content of reviews, photographs, and recorded conversations. These examples have different modalities: text, images, and audio. Modality means the form in which information is expressed.

A machine learning task defines what a model should predict. The information given to the model is its input; the answer it is meant to predict is the prediction target. In supervised learning, training examples include known targets so the model can learn relationships between inputs and answers.

Supervised learning learns from examples that pair input information with a known answer, called a label. For example, an email can be paired with its known category: unwanted or legitimate. These answers guide learning so the model can predict answers for new inputs.

Supervised learning learns from examples that pair inputs with known answers. The inputs are described by features. The answer the model should learn to predict is the target, also called a label. A labeled example contains both.

Both approaches learn from data, but they learn different things. Supervised learning uses examples that pair inputs with known outputs. It learns a relationship that can help predict an output for a new input. Unsupervised learning looks for structure in the inputs without being given target outputs to learn from.

Reinforcement learning trains a decision-maker through the consequences of its actions. The agent is the decision-maker, and the environment is the world it interacts with. For a delivery robot, the agent controls the robot and the environment includes the warehouse, obstacles, and delivery locations.

Supervised learning learns from examples that pair an input with a known answer, called a label or target. For example, a support ticket can be paired with its correct department. Training compares the model's predicted department with that supplied answer and adjusts the model to reduce prediction error.

Machine learning applications use learned patterns to turn new inputs into useful outputs. In computer vision, the input is an image or video. Patterns in visual features help a model identify objects, locate damage, or recognize printed text. For example, a package photograph can become a list of detected objects or an extracted shipping address.

A suitable solution starts with the required behavior, not with a model. Rules suit decisions whose logic is already known: calculate a delivery charge from an approved price list, or reject a form with a missing required field. These tasks require following a specification, not predicting an unknown answer.

Human-in-the-Loop means people deliberately participate in an AI workflow rather than treating its output as automatically authoritative. AI can help by sorting work, extracting information, or recommending an action. A person then applies evidence and judgment where the workflow calls for review.

Amazon Textract is a managed machine learning service that extracts text and data from documents, including scanned pages, images, and PDF files. Instead of a person reading each page and retyping its contents, an application can use Textract’s output.

Determinism means that identical inputs, initial state, and execution conditions produce identical results. Think of a calculator applying a fixed tax rate and rounding rule: with everything held constant, it returns the same total.

AWS Lambda runs your code without requiring you to provision or manage servers. An event is data describing something that happened, such as an order being placed. A configured trigger causes Lambda to run your function with that event.

Cost-benefit analysis asks whether an option creates enough value to justify its resources and risks. For AI, the useful question is not just “Does it work?” but “Does it improve on a simpler way of doing the job?”

Clustering is an unsupervised learning technique that discovers groups of similar records. It uses unlabeled data: examples without supplied target categories for the model to learn. Named columns do not make data labeled; a customer dataset can contain purchase frequency and average order value without containing customer segment labels.

Natural language processing (NLP) applies computational techniques to human language. It helps applications turn messages, reviews, and documents into useful information. Language is not a set of tidy database fields: context, spelling variations, and phrases such as “not bad” affect what words mean.

Amazon Personalize is a managed machine learning service that helps applications recommend products or content relevant to individual users. Instead of developers writing a separate set of recommendation rules for every person, the service learns patterns from data. A shopping app can suggest products, while a video app can suggest something to watch.

A foundation model may write fluent answers without knowing your organization’s current policies or product manuals. Knowledge Bases for Amazon Bedrock connects those sources to model requests so answers can use information beyond the model’s training.

Amazon SageMaker is the broader platform for data, analytics, and AI. SageMaker Unified Studio brings together tools for working with data and building AI applications. Within that platform, Amazon SageMaker AI provides managed capabilities for preparing data, building and training models, and deploying them. SageMaker AI is also available as a standalone service.

A foundation model is a model trained on broad data that can be adapted to different tasks. Starting with a pretrained model means you do not have to build and train that model from scratch. Depending on its capabilities, it can help with tasks such as generating text, summarizing documents, or creating images.

Amazon Bedrock provides managed access to foundation models: models pretrained on broad datasets that can support tasks such as generating text, summarizing documents, or creating images. Supported tasks depend on the model you choose.

Amazon Transcribe is a managed automatic speech recognition service. It turns speech in audio into text, adding punctuation to help make the result readable. You use the service without building or training your own speech recognition model.

Amazon Translate is a managed service that translates text from a source language into a target language. It helps make product descriptions, support messages, websites, and documents accessible to people who use different languages.

Transcription turns speech into written words. Amazon Transcribe performs this speech-to-text task: for example, English speech becomes an English transcript. This makes spoken information available to read rather than listen to.

Amazon Lex is an AWS service for building conversational bots that accept voice or text. It provides speech recognition for spoken input and natural language understanding to help determine what the user wants. A bot can guide a conversation toward a task, rather than simply return the user's words.

Amazon Lex helps you build interactive conversations using voice or text. Choose it when an application needs to understand a user's request and respond as part of a back-and-forth exchange.

Model explainability helps people understand why a model produced an output and what patterns influence its behavior. Before choosing a model, identify who needs the explanation and what they must understand. An engineer investigating errors needs different detail from a person asking why an application was declined.

A foundation model is broadly pretrained, so an application can reuse capabilities such as understanding language, recognizing visual patterns, or producing text. This makes it a useful candidate for summarizing conversations, drafting descriptions, or extracting information from varied documents. Not every foundation model accepts the same inputs or produces the same outputs: check its supported modalities, such as text, images, or audio.

A train-validation-test split divides a dataset into separate collections of examples called partitions. Each partition has a different job. Keeping them separate helps distinguish learning from examples from performing well on examples not used for learning.

A feature is an input variable a model uses. A label is the outcome a supervised model learns to predict. For customer churn prediction, features might describe subscription length and recent usage; the label records whether the customer later canceled.

Hyperparameter tuning compares training configurations to find a promising candidate model. A configuration is a set of choices that controls training or the model's structure, such as a learning rate or a decision tree's maximum depth.

Pre-training is the learning stage that builds a model's foundational capabilities. A model repeatedly processes examples and adjusts its weights: numerical parameters that shape its predictions. The learned weights are a central part of what training produces, not just settings for accessing the original dataset.

A pretrained model has already learned patterns from training data. Reusing it provides a starting capability without training from scratch. Research institutions, community repositories, model hubs, and commercial publishers distribute particular versions as model releases.

Fine-tuning adapts a pretrained model to a more specific task using targeted data. It updates learned parameters, or trains additional adaptation parameters, to produce a reusable adaptation. For example, a support team could adapt a model to turn customer messages into consistently structured issue summaries.

A model’s parameters, often called weights, are the learned values that shape its responses. Pre-training from scratch starts with newly initialized parameters and learns broad capabilities from a large, varied dataset. For a language model, these capabilities include patterns in language, code, and other training text.

Inference means using a trained model to produce an output from new input. The output might be a prediction, a classification, or generated text. An inference endpoint is a network address through which an application requests that work, commonly over HTTPS.

Inference uses a trained model to produce a prediction or generated output. An application sends input to an API endpoint, an address that accepts requests, and receives the result. Receiving a request does not automatically train or update the model.

Amazon Bedrock provides managed access to foundation models: models already trained on broad data that can perform tasks such as generating text, summarizing, creating embeddings, or generating images. Inference means using a model to produce a result from an input; it is different from training the model.

Amazon EC2 provides virtual machines called instances. With self-hosted inference, you run a model-serving application on an instance. Inference means using a trained model to produce an output from new input; it is not training the model.

Model hosting means running a trained model so an application can send input and receive a prediction. This use of a model is called inference. An endpoint is the address an application calls to request that prediction.

Kiro is an AI-assisted development tool. Engineers use it to understand source code, develop features, and prepare changes for review. Its specification-driven development workflow makes the intended behavior explicit before moving into implementation.

AWS Glue helps discover and prepare data. For discovery, a crawler inspects a configured data source, such as files in Amazon S3. It uses classifiers, which recognize supported data formats and help infer their schemas.

AWS Glue and AWS Lake Formation solve different problems in a data lake: a central store of datasets, often held in Amazon S3. Glue helps discover, describe, and prepare data. Lake Formation helps govern who can access it. They are complementary services, not replacements for each other.

A prompt tells a foundation model what to do. If an application copies that text into several places, changing the instructions can leave different parts of the application using different wording. Amazon Bedrock Prompt Management lets you create, organize, test, and version reusable prompts instead of managing scattered copies.

Model evaluation checks a foundation model's responses against defined criteria. Amazon Bedrock supports automated and human evaluations that help compare models and assess their suitability for a task. The results are evidence about observed behavior, not a guarantee about every future response or an explanation of the model's internal workings.

Amazon SageMaker Clarify, a capability of SageMaker AI, helps teams detect potential bias in data and model predictions. It compares groups using bias metrics: numerical measures of differences in representation, outcomes, or errors. You choose the group-defining attribute, called a facet, such as age group, and the groups to compare.

Amazon QuickSight is a managed business intelligence service. It turns data into charts, tables, and interactive dashboards so people can explore business questions without writing model-training code.

Experiment tracking preserves evidence from repeated attempts to build and evaluate models. An experiment groups related work; a run is one execution, such as a training job or evaluation. Each run needs a unique identifier so parallel attempts stay separate and teammates can refer to the same result.

Continuous integration (CI) means developers frequently merge small code changes into a shared codebase and receive automated build and test feedback. A pipeline runs a defined sequence of stages instead of relying on someone to remember release commands. In machine learning, those stages must check both the software and the model it will use.

AWS CodePipeline is a managed service that coordinates a release workflow. Instead of remembering a list of manual tasks for every change, a team defines a pipeline and runs changes through that workflow. A configured source-change trigger can start a run automatically.

Amazon CloudWatch monitors resources and applications. A metric is a series of numerical measurements over time, such as response time or CPU usage. Looking at these measurements helps a team notice when a deployed model becomes slow, unreliable, or short of capacity.

Model drift describes changes in a model’s operating environment that can make its learned patterns less useful. The model itself can stay unchanged while customers, products, or behavior change. Good validation results before deployment do not guarantee good predictions forever.

A classification model predicts a category. To measure its performance, compare its predictions with ground-truth labels: the known actual outcomes. For binary classification, first choose the positive class. Positive means the outcome being detected, not necessarily something desirable.

Cost per AI outcome connects spending to useful work, such as a customer served or a task completed successfully. It helps a team judge whether an AI application delivers business value, rather than simply choosing the model with the lowest request price.

AWS Cost Explorer helps you explore historical AWS costs and usage. For an ML project, start with a defined date range, such as a completed month. Daily views reveal changes within that period; monthly views help compare spending over time. These are spending records, not measures of model accuracy or business value.

Return on investment (ROI) compares the financial benefits of an AI initiative with its investment costs. It answers a business question: does the value gained exceed what we spent? A higher accuracy score or faster response is useful evidence, but neither is a financial return by itself.

A language model does not process text directly as words on a page. A tokenizer divides text into units called tokens and maps them to numerical token identifiers in its vocabulary. The model processes the resulting sequence of identifiers.

An embedding is a numerical representation produced by a machine learning model. It commonly takes the form of a fixed-length vector: an ordered list of numbers. Depending on the model, the input can be a passage of text, an image, or another supported type of data.

A language model needs text in a form it can process. Tokenization divides text into units called tokens and maps those units to numerical token identifiers from a vocabulary. This answers the question: “Which text units will the model process?”

A transformer is a neural network architecture that uses attention to build representations of relationships within a sequence. For text, a tokenizer first splits the input into tokens and assigns each one an identifier. A token can be a word, part of a word, or punctuation; it is not necessarily a whole word.

Text generation uses generative AI to produce written output from instructions and context. It can create summaries, translations, code, and conversational replies. A prompt supplies the request and any supporting material. Clear instructions about the audience, purpose, and format help define what a useful result should look like.

A modality is a form of information, such as text, images, audio, or video. Multimodal generation involves more than one modality. Turning a written description into an image and answering a question about a photograph in text are both multimodal interactions.

Semantic search finds existing material by its relationship to the meaning of a request. An employee asking “Can I expense a train ticket?” could find a passage titled “Reimbursement for business travel,” even though the wording is different.

Text generation composes a written response to instructions. Semantic search finds existing material whose meaning is relevant to a query. The main difference is the job being done: composing an output versus selecting material that already exists.

Semantic search finds existing information by meaning. A question such as “Can I expense a train ticket?” can return a passage titled “Reimbursement for business travel,” even though the wording differs. Its output is relevant source material, such as passages or document links—not a newly written answer.

Amazon Q is in maintenance since June 30, 2026. It is closed to new customers; existing customers can keep using it. AWS provides a migration guide at https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/qbusiness-availability-change.html. This lesson explains the service as AWS offered it, rather than recommending it for new workloads.

Model evaluation asks whether a model is suitable for its intended task under real operating conditions. Start with the intended users, supported tasks, and unacceptable failures. Turn these into acceptance criteria: measurable conditions the application must meet. For a support assistant, these could cover correct answers, appropriate escalation, and protection of customer information. Set the criteria before inspecting results, rather than changing them to make a result look successful.

Model selection means choosing a model that suits a particular task. Start with its capabilities: the kinds of input and output it supports, the languages it handles, and whether it can accept the required amount of context. A text-only model cannot meet a requirement to interpret images directly.

Inference means using a trained model to produce a result. For text generation, many Amazon Bedrock models offer token-based, on-demand pricing. A token is a unit of text processed by a model, often a word or part of a word. Tokens are not the same as words, and token counts can differ between models.

Inference latency is the time needed to obtain an output from a deployed model. Always define what you measure: model execution covers processing within the serving system, while application end-to-end latency can also include network travel, authentication, retrieval, queueing, and response delivery.

A context window is the finite capacity a model can use during a request. Language models measure this capacity in tokens: words, word fragments, punctuation, or other encoded content. Limits and counting rules depend on the selected model. Images and audio also consume capacity under model-specific rules.

Conversation memory is interaction history that an application retains and reuses. If a user says, “Compare that with the cheaper option,” earlier messages supply the missing meaning. In a typical invocation, the model uses the context it receives; the application or a managed conversation capability maintains continuity between invocations.

Conversation memory is information an application or managed service retains from interactions. It can include earlier messages, saved preferences, or a conversation summary. Persistent memory keeps information beyond the request that created it, potentially across sessions.

A model's training does not give it reliable access to every company's current policies, manuals, or product details. Retrieval-Augmented Generation (RAG) helps by finding external information relevant to a question and supplying it to the model when that question is answered.

Tool calling lets a language model request an operation outside its text-generation process, such as checking an order's current delivery status. The application supplies a tool definition: a name, a description of its purpose, and an input schema describing the accepted arguments.

Model Context Protocol (MCP) standardizes communication between AI applications and external capabilities. A capability might look up an order, read a document, or provide a reusable prompt template. A common interface reduces the need to invent a different connection method for every application and external system. It does not remove the work of connecting those capabilities to their underlying systems.

Model Context Protocol (MCP) standardizes communication between AI applications and external capabilities. It answers an integration question: how can an application discover and access capabilities through a common interface rather than a separate custom interface for each system?

Multi-agent orchestration organizes specialist AI agents around a shared task. Each agent has a role, instructions, and permitted capabilities. Specialists can use the same underlying model; different models are not required.

Multi-agent orchestration organizes several AI agents around a shared task. It assigns work, manages handoffs, and decides when the application can move forward. Specialist agents might research options, check rules, or prepare a response.

AWS Step Functions coordinates work through a state machine: a definition of steps and the rules for moving between them. Each step is a state. A Task state calls a service to perform work, such as running an agent or invoking a tool. Other states control which work happens next.

An agent workflow is the wider process around an AI agent's work, including calls to tools and processing their results. Within that process, code execution does a task; orchestration coordinates how tasks fit together.

An AI agent uses a model and tools to work toward a goal. Moving it into production requires more than a working prompt: it needs somewhere to run, controlled access to tools, continuity between interactions, and a way to investigate failures. Amazon Bedrock AgentCore provides managed components for these needs.

Amazon Bedrock provides managed access to foundation models. An application sends input to a selected model and receives output, such as a summary or generated text. AWS operates the underlying inference infrastructure. A model request does not, by itself, require an agent.

Conversational AI lets people interact with a system through natural-language dialogue. A person can ask a question, add details, and ask again. The system responds through text or speech.

In-context learning adapts a model’s response by changing the input it receives, not its weights, the learned parameters inside the model. Context can include instructions, examples, reference information, and conversation history. The same model can write a friendly reply or a formal summary when given different guidance.

Amazon Q Business has been in maintenance since June 30, 2026. It is closed to new customers; existing customers can continue using it. This lesson describes the service as AWS offered it, not as a recommendation for new workloads. AWS's migration guide: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/qbusiness-availability-change.html.

A hallucination is generated information that is false, fabricated, or unsupported by the evidence the system should use. Examples include an invented API parameter, a nonexistent legal case, or a summary that attributes a statement to a document that never contained it. An unsupported claim is not necessarily false, but it has not been established by the available evidence.

Nondeterminism means that repeating an operation with the same apparent inputs does not necessarily produce the same result. A generative model might use different wording, organize an answer differently, or make a different recommendation. This variation can help with brainstorming, but makes stable application behavior harder to achieve.

Hallucination concerns what an answer claims. A generative AI system hallucinates when it presents false, fabricated, or unsupported information. For example, a summary might attribute a statement to a document that never contained it. Fluent language does not establish that a claim is supported.

Model interpretability helps people understand how inputs relate to a model's predictions. Knowing that a model predicts “high risk” is different from understanding which inputs led to that result. This understanding helps people investigate errors and spot reliance on unexpected information.

Model transparency concerns visibility into a model and how it is developed and used. Here, we focus on transparency of its decision process: can someone inspect the rules or calculations that produce an output? A small decision tree can expose that process directly. Access to a complex model’s code or weights permits inspection, but does not necessarily make its decisions easy to follow.

Model interpretability concerns whether people can understand how a model's inputs relate to its behavior and outputs. Post-hoc explainability is a way to support that understanding: an explanation method is applied to a model after training.

Data residency describes where data resides geographically. Requirements can come from laws, contracts, industry rules, or internal policy. Turn each requirement into an explicit rule: which data classes it covers and which locations are approved. Customer records might have a strict boundary while public website images do not.

Total Cost of Ownership (TCO) is the complete cost of using a solution over an agreed period. For generative AI, it includes more than the price of a model call: it also includes preparing the model for the task, supporting infrastructure, and the people who keep it working.

A key performance indicator (KPI) is a selected measurement that shows progress toward an important objective. Start with the change the business wants, then choose a measure of that change. For a generative AI support assistant, “help customers resolve problems” points to successful resolution, not the number of answers generated.

Amazon Bedrock Guardrails provides configurable checks that help an application reduce responsible-AI risks. These checks examine content against policies you choose, such as restrictions on harmful content or sensitive information. They do not retrain the model or make it inherently trustworthy.

Amazon Q is in maintenance since 2026-06-30. It is closed to new customers; existing customers can keep using it. AWS provides this migration guide: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/qbusiness-availability-change.html. This lesson explains Q as AWS offered it, not as a recommendation for new workloads.

Amazon Q Business has been in maintenance since June 30, 2026. It is closed to new customers; existing customers can keep using it. This lesson explains the service as AWS offered it, not as a recommendation for new workloads. AWS's migration guide: https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/qbusiness-availability-change.html.

The AWS shared responsibility model divides security work between AWS and its customers. AWS provides security of the cloud: protecting the facilities, physical hardware, underlying networking, and virtualization layer that run AWS services. Customers provide security in the cloud: protecting their data, identities, permissions, applications, and customer-controlled settings.

Generative AI applications handle valuable information: prompts, retrieved documents, model outputs, and sometimes training data. Security protects that information from unauthorized access or changes and helps keep the application available.

AWS Identity and Access Management (IAM) controls access to AWS resources. An IAM role is an identity with permissions that an authorized person, application, or AWS service can use. A role does not have its own permanent password or long-term access keys.

Amazon Virtual Private Cloud (Amazon VPC) gives you a logically isolated network in AWS. You choose its IP address ranges and divide it into subnets, which are smaller network sections. For a generative AI application, this lets you separate application servers from resources intended for public access.

AWS Key Management Service (AWS KMS) manages cryptographic keys and controls their use. Integrated AWS services use these keys to protect data at rest: data saved in storage. AI training datasets, saved prompts, and generated outputs need this protection just like other application data.

Compliance means meeting applicable requirements and demonstrating that the required controls work. Requirements can come from laws, industry standards, contracts, and internal policies. A control is a safeguard or practice, such as restricting access or reviewing permissions.

AWS compliance programs provide assurance about AWS controls against particular standards or requirements. For a generative AI workload, this evidence can support assessment of the infrastructure underneath the application, such as AWS-operated facilities and infrastructure operations. It reduces the need to investigate those controls from scratch.

AWS Artifact provides access to AWS security and compliance documents, including SOC reports, ISO certificates, and PCI compliance reports. These documents help you assess AWS and supply evidence for reviews or audits. A control is a safeguard or practice, such as restricting physical access to a data center.

A modality is a type of information, such as text, images, audio, or video. Multimodality means working with multiple information types. Start model selection by writing down what the application must receive and what it must produce—not by looking for a model labeled “multimodal.”

A model's context window is the information it can work with during one inference request. For text models, this capacity is measured in tokens: words, parts of words, punctuation, or other text fragments. Tokens are not the same as words, and the same text can have different token counts with different models.

A foundation model processes the input in a prompt before generating an answer. Prompt caching saves reusable processing of an input prefix: the beginning of the prompt. A later request with a matching prefix can reuse that work instead of processing the whole beginning again.

A language model generates a response one token at a time. A token can be a word, part of a word, or punctuation. For each position, the model scores possible next tokens using the prompt and the text generated so far.

A language model generates text one token at a time. A token can be a word, part of a word, or punctuation. At each step, the model assigns probabilities to possible next tokens. Nucleus sampling, also called top-p sampling, uses those probabilities to build a shortlist of eligible choices.

A language model generates text by repeatedly choosing a next token, which can be a word, part of a word, or punctuation. Temperature and nucleus sampling affect that choice in different ways: temperature reshapes the probabilities, while nucleus sampling cuts down the candidate set.

A text embedding is an array of numbers, called a vector, produced by a machine learning model. It represents learned patterns in the meaning of a word, sentence, or passage.

Amazon OpenSearch Service is a managed search and analytics service that can also store and search embeddings. An embedding is a list of numbers, called a vector, produced by a machine learning model to represent content such as text.

An embedding is a list of numbers produced by a model to represent features of content, such as the meaning of a paragraph. Vector search compares a query embedding with stored embeddings to find nearby representations. A search for “recover account access” can therefore find “Reset your password” without matching the exact words.

An embedding is a list of numbers produced by a machine learning model to represent content, such as a document. Embeddings let an application compare content by meaning rather than only matching exact words.

An embedding is a list of numbers that represents features of content, such as the meaning of a text passage. An embedding model creates this list. Comparing embeddings can help an application find passages with similar meanings, even when they use different words.

Amazon RDS is a managed relational database service that offers several database engines. Amazon RDS for PostgreSQL runs the PostgreSQL engine. Amazon Aurora PostgreSQL-Compatible Edition runs an AWS-built engine designed to work with PostgreSQL applications and tools.

Neptune Analytics is an analytics database engine in Amazon Neptune that supports graph analysis and vector search. It is separate from Neptune Database. A graph represents things as vertices and their relationships as edges: for example, an article vertex can connect to a product vertex through an ABOUT relationship.

Fine-tuning and in-context learning can both help a pretrained model follow a task’s conventions. The key difference is where the adaptation lives: in learned model parameters, or in the context supplied for a request.

Model distillation transfers behavior from a teacher model to a student model. In the approach covered here, a larger teacher generates answers to representative application inputs. Those examples are reviewed and used to train a typically smaller student; the teacher's parameters are not simply copied.

Knowledge distillation identifies learning from a teacher model. A student model is trained to approximate useful behavior from that teacher, using signals such as teacher-generated answers or output probabilities. The defining question is: What provides the guidance for learning? Distillation transfers behavior through training; it does not mean directly copying the teacher’s parameters.

Model distillation describes a student model learning selected behavior from a teacher model. The teacher provides the learning signal: information that guides what the student learns. For example, the teacher can generate answers that become training targets for the student.

Fine-tuning adapts a pretrained model using targeted examples. Its value might be more consistent classifications, a required response format, or better task performance. The training bill alone does not tell you whether that improvement is worth buying.

An AI agent is a software system that selects actions to achieve a goal, observes their results, and decides what to do next. The goal describes the desired outcome, such as resolving a delayed order. It does not have to specify every step in advance.

Tool use lets a model request capabilities outside its learned knowledge. A tool can retrieve current inventory, calculate a total, or reserve an item. This gives business applications access to live information and actions without retraining the model.

Tool use lets a model-based application use capabilities outside the model’s learned knowledge. A tool can retrieve an order’s current status, calculate a total, or create a support ticket. The model requests an operation; application code executes it and returns the result.

Task planning connects a business goal to the actions needed to achieve it. An AI agent can interpret the request, identify missing information, and break the work into smaller tasks. Planning determines what work is needed and how actions depend on one another; scheduling determines when work runs.

Instruction hierarchy determines which instruction takes precedence when directions conflict. Authority comes from the instruction roles or fields supported by the model and API, not from how important the text sounds. A role is part of the structured input supplied to the model; typing SYSTEM: inside an ordinary user message does not create a system message.

Prompt context is background information a model receives for a particular task. It can include reference documents, purchase details, earlier conversation turns, or tool results. A support assistant answering a refund question needs the applicable policy and relevant purchase facts, not just the question. This helps it answer for the situation rather than rely only on patterns learned during training.

Negative prompting tells a generative AI model what its output should avoid. The exclusion might concern content, such as customer names; behavior, such as guessing an unknown cause; or presentation, such as Markdown code fences. Here, negative means excluding something, not using an unfriendly tone.

Chain-of-thought prompting guides a foundation model toward intermediate reasoning before a final answer. An intermediate step produces something needed later, such as a subtotal used in a comparison. The goal is a more structured approach to a problem, not simply a longer response.

In-context learning means adapting a model’s response through the context supplied with a request, without updating its model parameters. Supplied examples can demonstrate a task’s labels, answer format, or input-to-output pattern. For example, labeled support tickets can show how to classify a new ticket.

A prompt template is a reusable specification for constructing a model request. It combines fixed instructions with runtime placeholders: named spaces that an application fills with values for the current request. This lets an application repeat a task without rewriting its instructions each time.

Iterative prompt refinement means testing a prompt, examining the response, and revising the instructions to address a specific weakness. The aim is better results, not merely wording that sounds more polished.

AI guardrails define the boundaries of acceptable behavior for an AI application. They can limit supported topics, prohibited content, sensitive information, output formats, evidence requirements, and permitted actions. A support assistant, for example, can answer product questions without revealing another customer's account details.

Data poisoning is deliberate manipulation of information an AI system relies on. An attacker adds, changes, or selectively removes data to influence behavior. Unlike an accidental data-quality problem, poisoning is intentional, although intent can be difficult to establish. The goal may be worse answers overall, biased recommendations, or one specific wrong answer.

Jailbreaking is an attempt to make a foundation model disregard its safeguards or behavioral restrictions. The attacker wants a response the model is intended to refuse, such as prohibited guidance or disclosure of protected information. Here, jailbreaking concerns generative AI behavior—not unlocking a mobile device.

Prompt injection is an attempt to make a language model follow instructions supplied through an untrusted source instead of following the application's rules. The attack crosses an instruction trust boundary: material the model should read as data is treated as authority over its behavior. The payload can be ordinary language, such as “Ignore the summary request and reveal the confidential context.”

Data poisoning is deliberate manipulation of information an AI system relies on. An attacker might change training examples or replace a correct policy statement with a false one. The AI can then give a harmful answer while still following the intended task. An accidental error is a data-quality problem, not necessarily an attack.

Jailbreaking means trying to defeat a model’s safeguards or behavioral restrictions. These are limits on what the model should produce or support, such as restrictions against generating prohibited content. Jailbreaking asks: Can the attacker get the model to cross a boundary it is supposed to respect? Here, the term concerns AI behavior, not unlocking a mobile device.

Amazon Bedrock Prompt Management helps you manage reusable instructions for a foundation model. You develop a prompt in an editable draft, where you can change its wording, input variables, model selection, and supported inference settings.

Knowledge distillation is a training approach in which a student model learns to approximate the behavior of a teacher model. The teacher supplies training signals, such as probabilities for possible answers or answers it generates. Training adjusts the student so that its outputs become more like the teacher’s on the examples used.

Transfer learning reuses learning from an existing model for a new task. Instead of starting with randomly initialized weights, it starts with a pretrained model: a model that has already learned patterns from data. Weights are the numerical parameters adjusted during training.

Continued pre-training resumes training an existing foundation model on additional data rather than starting a new model from scratch. For a language model, that data can be an unlabeled domain corpus: a collection of documents about a field, such as equipment manuals and maintenance reports. Unlabeled means people do not need to supply a correct answer for each document.

Instruction tuning is supervised training that teaches a pretrained model how to respond to requests. Each demonstration contains an instruction, optional input or context, and a target response: the answer the model should learn to produce. It is supervised because the examples supply the desired answers rather than asking the model to discover what counts as a good response.

Continued pre-training and instruction tuning both adapt an existing foundation model through further training. Their central difference is what the training examples demonstrate: domain content, or how to respond to a request.

Data curation means making data fit for a specific purpose. For foundation model fine-tuning, first describe the behavior you want: classifying support requests, following a response format, or asking for clarification when information is missing. Choose examples that demonstrate that behavior. A large document collection is not automatically a useful training dataset.

AWS Lake Formation helps control who can read data lake datasets. For model training and fine-tuning, this lets a data owner make approved data available without giving every training workflow unrestricted access to the lake.

Amazon SageMaker Ground Truth helps organize data labeling: people receive tasks, follow instructions, and submit judgments that become training data. It supports human labeling workflows, including custom tasks. Creating this data is separate from training or fine-tuning a model.

Reinforcement Learning from Human Feedback (RLHF) uses human judgments to help shape a model's behavior. In a common workflow, reviewers compare candidate answers, a reward model learns to predict those preferences, and reinforcement learning updates the language model using that reward. This lesson focuses on preparing the comparisons, before reward training begins.

A benchmark dataset supplies a common set of test cases for comparing models. A case might contain a question and correct answer, a document and reference summary, or a prompt and scoring guide called a rubric. Shared cases make evaluation more repeatable than trying a few handpicked prompts.

Human-in-the-loop evaluation means people assess model outputs using explicit criteria. It is especially useful for open-ended tasks: several answers may be acceptable, yet differ in usefulness, completeness, tone, or suitability for an audience. Matching one reference answer word for word would miss those differences.

An n-gram is a contiguous sequence of n tokens. Tokens are pieces of text; here, treat each space-separated word as one token. A unigram has one token, a bigram has two, and a trigram has three. Overlap compares these sequences between a model's candidate output and a reference answer.

BLEU, short for Bilingual Evaluation Understudy, compares generated text, called the candidate, with one or more approved references. It measures lexical overlap: matching tokens and sequences of tokens, not whether the text means the same thing. A token is a unit produced by splitting text; for these examples, treat each word as one token.

ROUGE is a family of automatic text-evaluation metrics originally developed for summarization. It compares generated text, called the candidate, with a reference: an example of an acceptable output. It is useful for repeatable comparisons when a task has meaningful reference answers.

Both BLEU and ROUGE compare generated text, called the candidate, with acceptable text, called the reference. They measure overlap in wording, rather than directly judging whether the candidate means the right thing.

BERTScore compares a generated response, called the candidate, with a reference answer. The reference is a human-written or otherwise trusted example of the desired answer. BERTScore measures semantic similarity to that reference, not quality in isolation.

When evaluating generated text, a candidate is the model’s answer and a reference is an answer used for comparison. Two candidates can express the same idea using different words. Evaluation metrics differ in how much credit they give that change in wording.

LLM-as-a-Judge means using a large language model to assess a response produced by a model, including another output from the same model. Instead of answering the original question, the evaluator judges how well the candidate answer meets explicit criteria. Its result might be a score, a pass/fail decision, or a preference between two answers.

End-to-end testing checks an application as one connected system. A test begins where a user would begin, such as a chat interface or API, and follows the journey through to its result. It checks the response, any changed records, and any external actions—not just whether the model returned text.

A fluent answer can still be wrong. In Retrieval-Augmented Generation (RAG), evaluate two things separately: whether retrieval supplied useful evidence, and whether generation used that evidence correctly. This separation helps explain what failed instead of merely marking an answer as bad.

Trajectory evaluation checks the observable steps an application takes to reach an outcome. A trace can record an agent's tool calls, arguments, returned results, retries, and stopping decisions. For retrieval-augmented generation, or RAG, it can record retrieval and use of evidence; for workflows, it can record routing and stage transitions.

Task completion rate measures how often an application achieves a defined user goal. Calculate it as successfully completed tasks ÷ total eligible attempted tasks × 100. If an assistant completes 850 of 1,000 eligible tasks, its rate is 85%.

User satisfaction measures how users perceive an application, not just the quality of its underlying AI model. Clear language alone does not guarantee a good experience: slow responses, missing context, or confusing instructions can frustrate users. Begin by naming the intended users, their tasks, and the business objective, such as providing accurate support with less customer effort.

Task completion rate asks whether users achieved their intended goals. It measures the share of eligible attempted tasks that were successfully completed. For an appointment assistant, this means the correct booking was actually made—not just that the assistant said it was.

Algorithmic fairness concerns whether an automated system treats people and groups equitably in its intended context. It requires asking who benefits, who is harmed, and whether those differences are acceptable—not just whether predictions are usually correct.

Model robustness is the ability to maintain acceptable behavior when inputs or operating conditions differ from those used during development. Real inputs contain spelling mistakes, missing values, background noise, and unfamiliar patterns. Some inputs are deliberately manipulated to cause a failure.

Amazon SageMaker Model Cards, a capability of SageMaker AI, provide a structured record of important information about a machine learning model. They bring together its purpose, training details, evaluation results, and limitations so people can understand the model without having to inspect its code.

Output validation checks a model response before another component relies on it. Fluent text and valid-looking JSON are still untrusted input: they can contain malformed data, prohibited content, private information, or an action the user cannot perform.

Output validation checks a generated response before someone relies on it. For evidence checking, the important unit is a claim: a statement that can be compared with reference material. A fluent answer can contain both supported claims and invented details.

Responsible model selection means choosing an approach that meets the application's needs with justified environmental impact. First define acceptable quality, safety, fairness, privacy, and response time. Compare options that meet those requirements; lower resource use does not excuse failing them. Some tasks need only rules or a conventional statistical method, not a generative model.

Model compression reduces the resources a model needs while preserving enough quality for its intended use. Targets include smaller model files, less runtime memory, or fewer computations. These are different outcomes: a smaller file does not guarantee faster predictions. Inference means using a trained model to produce predictions or responses.

Intellectual property rights protect certain creations and information. Copyright protects original expression, such as writing, images, and code; trademarks protect source-identifying names and symbols. Patents protect qualifying inventions, and trade secret law protects qualifying confidential information. For generative AI, an early question is whether you have permission to use the material you supply as input.

Algorithmic bias is systematic unfairness in an automated system’s behavior or outcomes. In generative AI, it can appear as repeated stereotyping, less useful answers for a language community, or different descriptions of equally qualified people based on their gender. The issue is an unfair pattern, not simply an answer being wrong.

Data provenance is the evidence trail behind a dataset. It records where data came from, the circumstances of collection, the permissions that apply, and what changed before the data reached a model. Keep the record tied to an identifiable dataset version, so it describes particular data rather than a name whose contents can change.

Data cataloging organizes descriptions of datasets so people can find and understand them. These descriptions are called metadata: information about the data, such as its name, location, columns, business meaning, and owner. A catalog helps answer, “Which dataset contains the information I need?” The data itself can stay in its original storage system.

Source citation attributes information to a source. It answers, “Which source is being referenced?” A citation might identify a report, its publisher, and the release or section used, so a reader can locate and inspect the material.

Data representativeness describes how well a dataset reflects the people and operating conditions relevant to its intended use. It is not a property a dataset has for every purpose. Records from urban stores may represent urban demand well but leave rural demand poorly covered, even when the dataset is large.

A label records the outcome attached to an example, such as legitimate or fraudulent for a payment. Each outcome is a class. The label distribution describes how the examples are divided among those classes.

A target class is a category a model learns to predict, such as billing or technical support. Class balance concerns how examples are distributed across those categories. Data representativeness concerns how well the data reflects the people and conditions the system is intended to serve.

AWS Data Exchange helps you discover and subscribe to third-party data products. A product description is a starting point for assessing suitability, not proof that the data represents everyone your application will serve. Marketplace availability does not guarantee balanced or inclusive data.

AWS Glue DataBrew is a visual data preparation service. Before using data to train an AI model, you can examine individual records and summarize patterns across records. These views provide different kinds of evidence about data quality.

A model learns from training data, the examples used to fit it. Its real goal is to make useful predictions on unseen data, examples it did not learn from. This ability is called generalization.

Underfitting happens when a model fails to learn meaningful relationships in its data. Imagine predicting electricity demand with the same average value every hour. That rule cannot follow demand rising during working hours or changing with temperature.

Generalization means making useful predictions on new examples, not just examples used for learning. Overfitting happens when a model learns training-specific details that do not reliably hold for the people or situations where it will be used.

A model should learn patterns that help it predict new examples, not just reproduce its training examples. Statistical bias is the systematic difference between a model’s average prediction across possible training datasets and the true relationship. High bias commonly causes underfitting: the model’s assumptions are too restrictive to capture important patterns.

A label is an answer attached to a data item, such as a category for a support ticket or a factuality rating for a generated response. Label quality means that the answer fits the intended task, follows clear rules, and has appropriate evidence. Calling a dataset ground truth does not make its labels correct.

Subgroup analysis measures a system’s behavior separately for meaningful groups, such as languages, age ranges, device types, or accessibility needs. An aggregate metric combines everyone into one score, so a large group can hide poor performance for a smaller group. In the example below, 96% overall accuracy conceals 78% accuracy for Language B.

Amazon Augmented AI (Amazon A2I) helps you add human review to machine learning applications. A human review workflow defines the review task, the workforce that performs it, and where results are stored. A review of an individual item is called a human loop.

Model transparency gives people visibility into a model and how it was developed and is used. Documentation can describe its training data, testing, intended purpose, known limitations, and owner. These details help stakeholders—such as developers, reviewers, and users—judge whether it is suitable for a task.

Post-hoc explainability applies explanation methods to an already-trained model to help people investigate its behavior. It is useful when the model is opaque: its internal calculations are difficult for people to follow directly. An explanation can identify patterns worth investigating without exposing a complete decision process.

Model transparency concerns what people can know about a model and how it was developed and operated. This lesson focuses on transparency of its internal decision process: whether a person can directly follow how the model turns inputs into a prediction. A small decision tree can provide this visibility because its actual conditions and branches can be followed.

License entitlements are the rights an agreement grants you. For AI assets, those rights can cover inspecting materials, using a model or dataset, modifying it, and redistributing copies. These are separate permissions: permission to run a model does not automatically include permission to publish its files.

The interpretability-performance tradeoff is the choice between how well a model works and how easily people can understand its behavior. A small decision tree can expose readable decision rules. A large ensemble combines many models and is usually harder to inspect. But complexity does not guarantee better predictions: an understandable model can match or outperform a complex one.

Adversarial robustness means continuing to meet defined accuracy and safety requirements when someone deliberately tries to cause failure. A threat model describes what the attacker knows, what they can change, and which failures matter. An explanation available to an attacker becomes part of those conditions.

Selective disclosure means choosing what information to reveal, to whom, at what level of detail, and for what purpose. In responsible AI, it supports meaningful explanations and oversight without unnecessarily exposing personal data, confidential information, or details that enable misuse.

Human-centered design starts with people and their tasks, not with a model. Work with intended users, domain experts, and people affected by decisions to understand their goals and the consequences of errors. An explanation succeeds when people can understand a result and take an appropriate next step.

A feedback loop connects an AI result to evidence about what needs improving, responsible review, and a measured response. A rating button collects a signal; it does not complete that loop. Users first need enough information to recognize a problem and explain what is wrong.

A feedback loop connects what an AI system does to evidence used to improve its future behavior. The cycle includes observing results, considering a change, evaluating it, and checking what happens after an accepted change is used. Its central question is: Does evidence about results feed back into evaluated change?

AWS Config records the configurations of supported AWS resources so you can inspect their settings and changes over time. You enable recording and choose which supported resource types to record. Recording provides the evidence that configuration checks can use.

AWS Trusted Advisor evaluates aspects of your AWS environment against AWS best practices. A check looks for a particular condition, such as a security group that allows unrestricted access to certain ports. Its results identify potential risks or opportunities and recommend improvements.

AWS Trusted Advisor asks, “Where could this AWS environment improve?” It uses AWS best-practice checks to identify risks and recommend improvements, such as reducing unnecessary costs or addressing a security concern. Its main purpose is guidance on better use of AWS resources.

A retention policy defines which data to keep, why it is needed, how long to keep it, and what happens afterward. Its scope identifies the data it covers, such as financial records or support tickets. It should also name a data owner responsible for the policy and its exceptions.

AI governance sets the rules for how an organization develops and uses AI. A governance framework organizes those rules into responsibilities, reviews, evidence requirements, and training. A policy states an expectation, such as protecting confidential information; a protocol explains who must do what to meet it.

AI transparency means giving people relevant, understandable information about an AI system throughout its lifecycle. It covers AI use, intended purpose, limitations, development records, and accountability. It is broader than explainability, which concerns understanding how model outputs arise.

Least privilege means giving an AI workload only the access needed for its intended job. A document-reading assistant needs to read approved documents, not delete them or change who can access them. This limits the damage from mistakes, compromised code, or misleading instructions received by an agent.

Encryption converts readable data, called plaintext, into unreadable ciphertext. Decryption recovers the plaintext using the appropriate cryptographic key. The main benefit is confidentiality: obtaining encrypted data alone should not reveal its contents.

Network isolation means deliberately restricting which resources can communicate. An AI application may need to send prompts to a model service while being prevented from contacting unrelated systems. Isolation preserves justified connections rather than necessarily disconnecting everything.

AWS PrivateLink lets applications privately reach supported services without sending their requests over the public internet. An interface endpoint provides an entrance to a particular service through network interfaces with private IP addresses in your VPC, your isolated AWS network.

Amazon Macie is a managed data security service that discovers sensitive data in Amazon S3. For an AI system, that could mean finding personal or financial information in documents intended for training or retrieval.

Data cataloging organizes descriptions of datasets. A catalog entry helps people discover a dataset, understand what it contains, and find its location. The actual data can remain in a database, object storage, or another source system.

Source citation connects an AI artifact to evidence that an authorized person can inspect. An artifact can be an AI answer, a dataset, or model documentation. For an answer, the citation should identify the document and passage supporting a claim. For a dataset, source references can identify the original collection and the version used.

Data quality means fitness for a specific purpose. A daily inventory snapshot may be useful for demand forecasting but too old for deciding whether an order can ship now. Begin with the consuming application's needs and the consequences of wrong, missing, duplicated, or late data—not with a universal definition of “clean.”

Access control determines whether a requester may perform an action on a resource. The requester can be a person or a workload, such as an application, data-processing job, or model-training job. The resource might be a dataset, table, or stored file.

Data integrity means keeping changes to data authorized, valid, and traceable. AI datasets need to change as new examples arrive or labels are corrected. The goal is not to prevent every change, but to prevent improper changes, including accidental damage and deliberate manipulation.

Data integrity concerns whether data keeps its intended meaning and state without improper changes. It covers accidental damage and unauthorized manipulation. Integrity does not mean that data must never change: an approved, valid, traceable correction can preserve integrity.

Privacy-enhancing technologies (PETs) help process data while limiting sensitive information revealed to others. Privacy matters even when access is authorized: an analyst or model user may learn more than they need. Start by identifying the sensitive information, the intended computation, and who could see inputs, intermediate results, or outputs.

Amazon Inspector automatically assesses supported AWS workloads for security vulnerabilities. A vulnerability is a weakness, such as a flaw in a software package, that an attacker could exploit. Inspector helps you find these weaknesses before they cause harm; a finding does not mean an attack has occurred.

Encryption at rest protects data while it is stored. For an AI system, that can include training files, saved prompts, generated reports, model artifacts, and backups. Sensitive information can exist in any of these copies, not just the main dataset.

Transport Layer Security (TLS) protects data moving between two endpoints. It provides confidentiality, so someone watching the network cannot read the protected content, and integrity, so changes to that content are detected. For an AI application, that content can include prompts, retrieved text, credentials, and model responses.

The same data can move between several states. An AI application's report might travel to a storage service, remain saved there, and later travel back to a reader. Encryption at rest and encryption in transit protect different stages of that journey.

Data Loss Prevention (DLP) protects sensitive information from leaving its permitted boundaries. A boundary might separate tenants, organizations, approved recipients, or geographic regions. A disclosure policy states which data is protected, who may receive it, which destinations are allowed, and what to do when a transfer violates the rules.

AWS CloudTrail records AWS account activity as events. These records help answer who did what, when, and from where. Activity can originate from the AWS Management Console, command line tools, applications, or AWS services.

AWS CloudTrail and AWS Config both support governance: checking that AWS resources and their use follow an organization's expectations. They provide different evidence, so begin with the question you need to answer.

An audit record describes security-relevant activity: who acted, what they did, which resource was involved, when it happened, and the outcome. Preserving these records lets reviewers reconstruct events and check accountability. A record is useful as evidence only if it remains trustworthy and available when needed.

A confidence score expresses how strongly a system supports a prediction or response according to a particular method. It can help identify answers that need attention, especially when fluent language hides unsupported claims. Its meaning depends on how it was produced: always check the score's definition, scale, and direction.