the ultimate
AI UX Glossary

Systems that perform tasks requiring human-like intelligence, such as reasoning, learning, or decision-making.

Maps apps use AI to predict traffic and reroute drivers.

An autonomous system that can act on behalf of a user to complete tasks across tools or platforms.

A recruiting agent screens resumes and schedules interviews automatically.

Ensuring AI systems behave in ways that match human goals, values, and safety expectations.

A customer-support AI is tuned to calm angry users instead of escalating.

A reasoning style where AI shows its intermediate steps before the answer.

AI reveals its step-by-step math before the final total.

A signal of how certain an AI is about its output, shown numerically or visually.

A model shows 85% confidence that an image contains a cat.

The practice of shaping what information goes into a model’s context window to improve output quality.

Including FAQs and chat history so the model answers support queries more accurately.

A design where AI supports a human but doesn’t act fully autonomously.

GitHub Copilot suggests code while the developer accepts or rejects it.

Distinguishing between AI that assists (copilot) and AI that acts autonomously (agent).

Copilot suggests copy; an agent drafts, sends, and tracks it automatically.

Vector representations of data that capture semantic meaning for search and clustering.

Searching 'reset password' retrieves the right help doc even if wording differs.

A clear user option to exit an AI-driven process when it’s not working well.

A 'Talk to a person' button in a chatbot.

How understandable an AI’s decision or output is to humans.

A loan model lists the top 3 factors behind an approval.

Collecting user input or corrections to improve AI behavior over time.

Thumbs up/down buttons that inform future chatbot responses.

Improving AI responses by giving it a few examples to learn from.

Providing three labeled emails so the model classifies new ones correctly.

A command-line tool for searching text, often used to debug logs or outputs in AI workflows.

Using `grep 'error'` to find failed inference calls in server logs.

Verified, correct data used as a benchmark for training and evaluation.

Labeling 1,000 customer emails as 'billing' or 'support' to train a classifier.

Linking AI outputs to trusted sources or data to ensure accuracy.

An internal policy bot cites the exact page from the company handbook.

Safeguards that prevent harmful or inappropriate AI outputs.

Blocking violent or hateful language from appearing in chatbot responses.

When an AI generates false or fabricated information that sounds plausible.

A chatbot invents a non-existent feature in a product description.

When an AI agent passes control to a human or another system after completing its part.

A support bot escalates a case to a human agent after trying its best.

Design rules and constraints that keep AI outputs safe, relevant, and on-brand.

A support bot avoids providing medical advice it’s not qualified for.

The process of running a trained model to produce outputs or predictions.

A recommendation model suggests your next video to watch.

A structured network of entities and relationships that improves AI reasoning and retrieval.

A customer-support AI understands that 'login issue' relates to 'password reset'.

A model trained on vast text data to understand and generate human-like language.

An LLM digests long policy documents and drafts summaries.

The time between a user’s input and the AI’s response.

A chatbot that replies in 1–2 seconds feels conversational.

AI’s ability to remember context or details across interactions or sessions.

A shopping bot remembers your shoe size and past purchases.

A framework for securely connecting AI systems to private data and workflows.

Using MCP so an assistant queries internal Confluence pages safely.

Models that work across multiple input/output types like text, images, and audio.

An assistant analyzes an uploaded chart and answers a spoken question about it.

Monitoring how AI behaves in production using metrics, logs, and traces.

Tracking response times and hallucination rates of a chatbot.

A model generalizing from a single example to new cases.

Showing one invoice format so the model processes others like it.

Directing tasks between multiple agents, APIs, or models to achieve a goal.

AI decides to search documents before generating an answer.

When a model performs well on training data but fails to generalize to new data.

A churn model works perfectly on past data but fails on new customers.

Adjusting AI behavior to user preferences, history, or signals for relevance.

A writing assistant adopts your preferred tone after repeated use.

The instruction or input given to an AI model to guide its response.

Typing 'Summarize this article in 3 bullets' into a chat box.

A technique where outputs from one prompt feed into another to complete multi-step tasks.

AI extracts meeting action items, then generates a follow-up email draft.

Usage caps set by APIs to prevent overload or abuse.

An AI API allows only 60 requests per minute per user.

Testing AI systems adversarially to find vulnerabilities or risky behaviors.

A team tries to trick a bot into giving disallowed instructions.

A training method where models learn from feedback in the form of rewards or penalties.

An agent learns to maximize click-through rate by receiving reward signals.

Combining search with generation by retrieving data before producing answers.

A support bot fetches KB articles and includes them in its reply.

Reverting to a previous model or version when a new AI update causes problems.

Switching back to the old summarizer after the new one confuses users.

A safe test environment for experimenting with AI features without affecting production.

Designers trial prompt variations in a sandbox before release.

A group of AI agents working together, often loosely coordinated, to solve parts of a bigger task.

Multiple agents research competitors and combine results into one report.

Artificially generated data used to augment training or protect privacy.

Creating synthetic patient data to train models without real records.

When a user reclaims control from an AI during or after it starts acting.

A designer cancels an AI-generated layout mid-process to edit manually.

The units of text AI models process and bill against (≈¾ of a word per token).

A 1,000-word draft consumes ~1,300 tokens.

A neural network architecture based on attention mechanisms that powers most modern LLMs.

GPT and other LLMs use transformer layers to process long sequences effectively.

Letting users reverse or roll back an AI’s action.

Clicking 'Undo AI Rewrite' in a text editor.

Collaborating with AI in a conversational way to co-create product flows, designs, or code.

A team chats with AI to sketch and refine an onboarding flow in real time.

AI that communicates through real-time speech, often conversationally.

A language tutor AI speaks and listens to practice dialogue.

When a model performs a task it wasn’t trained on, using only instructions.

A model summarizes a contract without being trained specifically on legal data.