Skip to main content

AI Glossary

info

This content was developed with AI assistance and is regularly reviewed for accuracy.

Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.

This glossary explains the words you meet when you use AI, build with it, or buy it. It merges the site's earlier beginner and business glossaries and adds the agent-era vocabulary that became standard in 2025-2026: MCP, skills, tool use, computer use, context engineering, effort, connectors, harnesses, evals, and agent loops.

Terms are grouped by topic so related ideas sit together. To jump to one, use your browser's find (Ctrl+F or Cmd+F). Where a term matters for decisions, a short "why it matters" note follows the definition.

Core concepts​

Artificial intelligence (AI) Computer systems that perform tasks that normally need human judgment, such as understanding language, recognizing images, or making recommendations. AI is a capability built into products rather than a single thing you buy, so judge "AI-powered" claims by what the feature actually does.

Machine learning (ML) A kind of AI in which software learns patterns from examples instead of following hand-written rules. A fraud model trained on past transactions is ML. Most ML depends on good data, so data quality usually matters more than the algorithm.

Deep learning Machine learning that uses neural networks with many layers. It powers modern speech recognition, image generation, and language models.

Neural network A system of connected numerical units, loosely inspired by brain neurons, that learns to recognize patterns by adjusting the strength of its connections during training.

Generative AI AI that creates new content (text, images, audio, video, or code) rather than only classifying existing data. ChatGPT, Claude, Gemini, and image generators are generative AI.

Large language model (LLM) A model trained on very large amounts of text to understand and generate language. The assistants most people use (ChatGPT, Claude, Gemini, Copilot) run on LLMs; the current model names are listed in AI Model Comparison.

Foundation model A large model trained on broad data that can be adapted to many tasks. Most LLMs are foundation models.

Frontier model A model at or near the leading edge of capability at a given time. In 2026, labs increasingly gate some frontier capabilities behind safety verification programmes, especially for cybersecurity.

Algorithm A set of step-by-step instructions a computer follows to solve a problem.

Natural language processing (NLP) The field of AI concerned with reading, interpreting, and producing human language, including chatbots, sentiment analysis, and document search.

Computer vision AI that interprets images and video, used for tasks from defect detection on a production line to reading receipts.

How models work​

Token The unit of text a model reads and writes: a word, part of a word, or punctuation. Prices and limits are measured in tokens. On Anthropic's current tokenizer, 1 million tokens is roughly 555,000 English words.

Context window The amount of text, measured in tokens, a model can consider at once: your conversation, attached files, and instructions together. Current Claude models and Gemini 3.1 Pro accept about 1 million tokens. Accuracy still drops on very long inputs, so a bigger window is not a guarantee of better answers.

Training data The text, images, and other material a model learned from. Gaps and imbalances in training data are a common source of bias.

Parameters The internal numbers a model adjusts during training. Parameter count is a rough proxy for size, not a reliable measure of quality.

Transformer The neural network design behind nearly all modern language models, good at relating each part of a sequence to every other part.

Diffusion model A type of generative model, common in image and video generation, that creates output by gradually refining random noise.

Inference Running a trained model to get an output. Every time you send a message, the model performs inference; API prices are inference prices.

Knowledge cutoff The date after which a model's training data ends. In 2026 cutoffs are typically a few months before release, and assistants search the web by default, so "the AI doesn't know recent events" is much less true than it was.

Reasoning (thinking) Work a model does before answering: planning, checking, and trying alternatives. Current frontier models reason by default, and on some (such as Claude Opus 5.5 and Fable 5.1) thinking cannot be switched off. You control how much they reason with effort.

Effort A setting that controls how much reasoning and how many tokens a model spends on a task, typically from low to max. Higher effort helps on hard problems and costs more time and money; low effort suits quick, simple jobs. Effort has replaced the older idea of turning a separate "reasoning mode" on and off.

Hallucination A plausible-sounding answer that is false, such as an invented citation or statistic. Rates vary by task: low on short, grounded summaries and much higher on citations, long documents, and legal or medical specifics. Trackers counted roughly 2,000 court cases involving AI-fabricated citations by September 2026 (reported), which is why verification remains essential.

Multimodal AI AI that works with more than one kind of input or output, such as text, images, audio, and video.

Latency The delay between sending a request and receiving the response. Lower effort and smaller models usually mean lower latency.

Open-weight model A model whose trained parameters are published so you can download and run it yourself. "Open-weight" is more precise than "open source" because licences vary, from permissive (Apache 2.0, MIT) to custom terms with usage limits. See Open-Weight Models and Running One Locally.

Local model A model running on your own computer or server rather than a vendor's cloud, using tools such as Ollama or LM Studio. It keeps data on your machine at the cost of lower peak capability.

Using AI day to day​

Prompt Whatever you send the model: a question, instruction, or request, plus any attached material.

Brief A structured request with six parts: objective, context, resources, constraints, output, and verification. It is this site's recommended way to delegate a task. See the Brief and Prompt Library.

Prompt engineering The craft of wording prompts to get better results. Many 2023-era tricks (personas, "think step by step") matter little with current models; clear objectives, good context, examples of the format you want, and a way to verify the result still matter a great deal.

Context engineering Deciding what information a model or agent gets at each step, trading off accuracy, speed, and cost. It covers which files to attach, what memory to keep, which tools to expose, and when to summarize or reset. Practitioners describe it as the successor skill to prompt engineering for agents.

Few-shot and zero-shot Zero-shot means asking a model to do a task with no examples; few-shot means including a few examples in the request. Examples are most useful for showing an output format.

Memory An assistant feature that carries facts and preferences across separate conversations. You can usually view, edit, or turn it off in settings, and it is worth checking what it stores.

Project A workspace in an assistant (ChatGPT, Claude) with its own instructions and files that apply to every conversation inside it. Projects are the simplest way to stop repeating context.

Custom instructions Standing guidance you give an assistant about yourself and how you want answers, applied to every conversation.

Connector An authorized link that lets an assistant read or act in another service, such as Gmail, Google Drive, Calendar, Slack, or a CRM. Connectors are often built on MCP. Grant the narrowest access that works, because a connected assistant can act with your permissions.

Web search and grounding An assistant looking up current information and citing sources while it answers. Grounded answers are easier to verify than answers from training data alone.

Deep research A mode in which an assistant spends minutes searching many sources and writes a cited report. Useful for broad questions; still check the citations.

Artifact A standalone piece of output an assistant creates alongside the chat, such as a document, chart, app, or slide deck, that you can edit, download, or share.

Agents and building with AI​

AI agent An AI system that pursues a goal over multiple steps by deciding which actions to take, using tools, and checking results, rather than answering a single message. Agents can do whole tasks (research a topic, fix a bug, triage an inbox) but need clear boundaries and human oversight. See Agents vs Assistants.

Agent loop The cycle an agent repeats: look at the situation, decide on an action, act through a tool, check the result, and either continue or stop. Good loops have an explicit stop condition and a budget. Claude Code's /loop (repeat on an interval) and /goal (keep working until a condition is met) are built-in loop primitives. See Anatomy of an Agent Loop.

Harness The scaffolding around a model that turns it into a working agent: its instructions, tools, orchestration, memory handling, and checks. Claude Code, Codex, and agent frameworks are harnesses. Two agents using the same model can perform very differently because of their harnesses.

Tool use (function calling) A model asking your application to run a defined function, such as "search orders" or "create calendar event", with structured arguments. The application runs it and returns the result, and the model continues. Tool use is how models act on the world.

Computer use An agent operating a computer or browser the way a person does: reading the screen, moving the mouse, typing, and clicking. It reaches software that has no API, but vendors themselves describe it as risky and advise against using it on financial, legal, or medical sites without supervision.

MCP (Model Context Protocol) An open standard for connecting AI applications to tools and data sources through "servers" that expose capabilities in a consistent way. Introduced by Anthropic, it is now governed by the Agentic AI Foundation under the Linux Foundation and supported across major assistants. The specification revision of July 28, 2026 moved to stateless HTTP, deprecated the older SSE transport, and strengthened authorization. Why it matters: one integration can serve many AI tools, which reduces lock-in. See MCP Explained.

Skills (Agent Skills) Packaged instructions, plus optional scripts and reference files, that an agent loads only when a task needs them. A skill is a folder with a SKILL.md file, and Agent Skills is an open standard used by Claude and other agents. Skills turn a brief you reuse into something an agent applies automatically. See Skills, AGENTS.md, and CLAUDE.md.

AGENTS.md and CLAUDE.md Plain-text files at the root of a project that tell coding agents how the project works: commands, conventions, and rules. Agents read them at the start of a session.

Subagent A separate agent instance a main agent hands a sub-task to, with its own fresh context. Subagents keep the main conversation focused and let work run in parallel.

Scheduled task and routine An agent job that runs on a timetable or trigger instead of when you type. ChatGPT and Claude both offer scheduled tasks on paid plans, and Claude Code's Routines run in the cloud on a schedule, an API call, or a GitHub event. Check the output of each run, not just its status.

Human-in-the-loop (approval) A design in which a person reviews or approves an AI's work at key points, such as before an email is sent or a file is deleted. Too many approval prompts cause "approval fatigue", so place them where mistakes are costly or irreversible.

Blast radius Everything an agent could affect if it made a mistake: the files, accounts, and systems its permissions reach. Shrinking blast radius (scoped credentials, read-only access, sandboxes) is the most reliable agent safety measure.

Prompt injection An attack in which instructions hidden in content an AI reads (a web page, email, or document) try to make it do something its user did not ask for. Any agent that reads untrusted content and can act is exposed; limit its permissions and require approval for sensitive actions.

Compaction Automatically summarizing earlier parts of a long conversation or agent session so it fits in the context window. It keeps work going but can lose detail.

Eval (evaluation) A repeatable test of whether an AI system does its job: a set of tasks, run one or more times (trials), each scored by a grader (code, a model, or a person). Teams typically start with 20-50 tasks drawn from real failures. Two common metrics: pass@k (succeeds at least once in k tries) and pass^k (succeeds every time in k tries), which matters for agents users rely on. See Evals for Agents.

API (application programming interface) A defined way for programs to talk to each other. AI providers sell model access through APIs priced per token. See Development & APIs.

SDK (software development kit) A library that makes an API easier to use from a programming language, such as the official Anthropic, OpenAI, or Google SDKs.

RAG (retrieval-augmented generation) Searching a collection of your documents for the passages relevant to a question and giving only those to the model, so answers are grounded in your data. See RAG: How It Works & When to Use It.

Embedding A list of numbers representing the meaning of a piece of text, so that similar meanings have similar numbers. Embeddings power semantic search and RAG.

Vector database A database built to store embeddings and find the most similar ones quickly. Examples include Pinecone, Weaviate, Qdrant, and pgvector.

Fine-tuning Further training a model on your own examples to change its behavior, style, or specialist vocabulary. It does not reliably teach new facts; RAG or attached context usually does that better.

Synthetic data Artificially generated data that mimics real data, used when real data is scarce, sensitive, or expensive to label.

Business and strategy​

AI literacy The ability to understand what AI can and cannot do, use it effectively, judge its output, and recognize its risks, without needing to code. The EU AI Act's Article 4 asks organizations to take measures to promote it among staff. See Build an AI Literacy Program.

AI readiness How prepared an organization is to adopt AI: data quality, infrastructure, skills, leadership support, and governance.

AI transformation Integrating AI across operations, strategy, and culture rather than deploying a single tool, so that how work gets done changes.

Proof of concept (PoC) and pilot A small, time-limited test of whether an AI solution works for a specific use case before a wider rollout. Define success criteria before you start.

AI champion An employee who helps colleagues adopt AI tools and relays their needs to the teams running the rollout. Champions are one of the most effective adoption levers.

Shadow AI Employees using AI tools the organization has not approved, often on personal accounts. It is a data-leak risk and usually a sign that approved tools do not meet people's needs.

Change management (AI context) The structured work of moving people from current workflows to AI-assisted ones: communication, training, support, and feedback. See Drive AI Adoption on Your Team.

Workflow automation Using software to run repeatable processes automatically. Classic workflow tools (Zapier, Make, n8n) follow fixed steps; adding AI lets them handle unstructured inputs such as emails. A plain workflow is often more reliable and auditable than an agent. See Workflows vs Agents.

Workspace agents and custom assistants Assistants configured with instructions, knowledge, and tools for a specific job, built without code. In ChatGPT, Workspace Agents replaced individual creation of Custom GPTs in August 2026; Claude uses Projects and skills, and Gemini uses Gems.

SaaS (software as a service) Software delivered over the internet by subscription. Most business AI tools are SaaS; compare total cost over two to three years, not just the monthly price.

Seat and usage pricing Seat pricing charges per user per month; usage pricing charges per token, credit, or task. Many 2026 plans mix both, so set budget alerts when usage-based features are switched on.

Free tier A no-cost version of a product with limits on usage, features, or model quality. Free tiers usually run a vendor's smaller models and may have weaker data protections, so they suit evaluation more than confidential business work.

Enterprise plan A business agreement with administrative controls, security features, data-retention terms, and support. Negotiate data handling, retention, and exit terms, not just price.

ROI (return on investment) The value an AI initiative returns relative to its cost. Measure time saved and outcomes against a baseline, and judge cost per completed task rather than per request. See Evaluate and Measure.

Risk, safety, and compliance​

Data privacy (AI context) How an AI service collects, stores, uses, and shares your data, including whether inputs are used to train models and what memory and connectors retain. Check before entering anything sensitive. See the AI Data Privacy and Safety Checklist.

Bias Systematic errors that unfairly favor or disadvantage groups, usually inherited from training data. Test any AI used in hiring, lending, pricing, or service decisions before and after deployment.

Alignment How well an AI system's goals and behavior match what its developers and users intend, including being helpful, honest, and harmless. See The AI Alignment Problem.

Guardrails Controls that keep an AI system within acceptable behavior: input and output filters, permission limits, approval steps, and budget caps.

Red teaming Deliberately trying to make an AI system fail, misbehave, or be misused, so problems are found before attackers or users find them.

Adversarial example An input crafted to fool a model into a wrong answer. Prompt injection is the language-model version that matters most for everyday users.

Explainability (transparency) How well an AI system's decisions can be understood and explained to the people they affect and to regulators. It is increasingly required for decisions such as credit, insurance, and hiring.

AI governance The policies, roles, and processes an organization uses to manage AI: which tools are approved, what data may be used, who signs off on risky uses, and who is accountable. See Govern: Frameworks, Policy, and Compliance.

Responsible AI Designing, deploying, and running AI with attention to fairness, transparency, safety, privacy, and accountability throughout its life, not just at launch.

EU AI Act The European Union's law regulating AI by risk level, from banned practices through high-risk systems with strict obligations to transparency duties for systems such as chatbots. It applies to organizations serving people in the EU, wherever they are based. Parts already apply (prohibitions and the AI literacy duty since February 2025), and a 2026 political agreement on the Digital Omnibus was reported to push most standalone high-risk requirements to December 2027 and August 2028. See The Global AI Regulatory Landscape.

Key takeaways​

  • The core vocabulary (token, context window, hallucination, effort) explains most of what you see when you use an assistant.
  • The agent vocabulary (tool use, MCP, skills, harness, agent loop, eval) describes how assistants now act, not just answer.
  • Business and risk terms (AI literacy, shadow AI, governance, blast radius) are what leaders need to buy and roll out AI safely.
  • Terms keep evolving; when a vendor uses a new word, ask what it lets the system read, do, and remember.

Next steps​

SponsoredYour product hereReach people actively learning AI — flat monthly rates, no ad networks