AI glossary
47 terms explained in plain language. Tap any of them for examples and related terms.
- AI agent
An AI system that doesn’t just answer but takes steps on its own to complete a task: searching the web, filling in a form, using other apps. It’s worth supervising, because it can make mistakes when it acts.
- AI literacy
Basic knowledge for using AI with judgement: what it can do, its limits and its risks. The EU AI Act requires it of staff at organisations deploying AI systems (Art. 4).
- AI search engine
A tool that answers your question with written text and links to the pages each fact comes from, instead of a plain list of results.
- AI watermark
A visible or invisible signal embedded in AI-generated content so it can be identified later.
- API
A gateway that lets one program use another’s services. AI companies sell access to their models through APIs so other apps can build them in.
- API key
A secret code that identifies your account with an AI service so you can use it from other programs. Whoever has it can spend in your name: do not share it.
Example: The key you paste into an agent to connect it to a model.
- ATS (applicant tracking system)
Software some companies use to receive, sort and filter CVs. A clear CV using the job ad’s vocabulary is easier to process.
- Bias
A model’s tendency to give unbalanced or unfair answers, often inherited from the data it was trained on.
- Chatbot
A program you talk to in writing or by voice. Today’s chatbots, such as ChatGPT or Gemini, run on language models.
- Content credentials (C2PA)
An open standard that adds verifiable information about an image’s or video’s origin, for example whether it was generated or edited with AI.
Example: Images created with Copilot include these credentials, according to Microsoft.
- Context window
How much text (measured in tokens) a model can take into account at once: your conversation, the documents you attach and its reply.
- Deep research
A feature in some assistants that searches and reads many sources over several minutes and delivers a report with citations. It saves time, but you need to check the sources.
Example: Asking for a report on public grants for home renovation.
- Deepfake
An image, audio clip or video generated or altered with AI to look real, for example putting words in someone’s mouth. The EU AI Act requires them to be labelled as artificial content.
- Embedding (semantic vector)
A way of representing text as a list of numbers so that texts with similar meanings end up close together. It’s used to search by meaning, not just by words.
- EU AI Act
Regulation (EU) 2024/1689, the first general law on artificial intelligence. It classifies systems by risk level and sets transparency duties, such as labelling deepfakes.
- Fine-tuning
Extra training of an existing model with specific examples to specialise it in a task or style.
- Gateway
An intermediary program that connects your messaging apps or other channels to an AI agent and controls who can talk to it.
Example: The OpenClaw Gateway receives your Telegram messages and passes them to the agent.
- GDPR
The EU General Data Protection Regulation (2016/679). It governs how personal data is handled and gives you rights such as access and erasure.
- Generative AI
A type of artificial intelligence that creates new content (text, images, audio, video or code) from a request.
- GPT
Short for “Generative Pre-trained Transformer”, OpenAI’s family of models used in ChatGPT. As of October 2026 the latest version in ChatGPT is GPT-6.
- Hallucination
When an AI presents made-up or wrong information as fact, often with complete confidence. That’s why you need to check facts, figures and quotes.
- High-risk AI system
An EU AI Act category for uses that can affect people’s rights, such as assessing students or selecting staff. They carry stricter obligations.
Example: A system that automatically scores job applications.
- Inference
The moment a trained model is used to answer. Every reply a chatbot gives you is an inference and has a computing cost.
- LLM (large language model)
An AI model trained on huge amounts of text to predict and generate language. It’s the “engine” behind chatbots such as ChatGPT, Gemini or Claude.
- Local model
An AI model that runs on your own computer or phone, without sending your data over the internet. It’s usually less powerful than cloud models.
- MCP (Model Context Protocol)
Open standard for connecting AI applications to external tools and data such as files, calendars or databases. Its website compares it to a USB-C port for AI.
- Memory
A feature of some assistants that keeps information about you between conversations to personalise replies. You can review it and turn it off in settings.
- Multimodal
A model’s ability to understand or generate several kinds of content: text, images, audio or video.
- Node
In programs such as ComfyUI, each box on the canvas that performs a single task: loading a model, writing the description, generating or saving the image.
Example: A “Save image” node at the end of the workflow.
- Open source
Software whose source code is public and can be read, modified and shared under its licence. In AI people also talk about “open weights”: models you can download and run.
- Open weights
An AI model whose trained parameters are published so anyone can download and run it. It doesn’t always include the training data or code.
- Parameters
The internal values a model learns during training. They’re counted in billions; more parameters doesn’t always mean better answers.
- Prompt
The instruction or request you write to an AI. A good prompt includes context, task, format and tone.
- Prompt injection
An attack in which a text the AI reads (a web page, an email, a document) hides instructions to make it do something the user did not ask for. It is a particular risk for agents.
- Quantisation
A technique that reduces the precision of a model’s numbers so it takes up less memory and runs on more modest computers, at the cost of some quality.
- RAG (retrieval-augmented generation)
A technique in which the AI first looks up information in specific documents and then answers based on them. It reduces, but doesn’t eliminate, made-up answers.
- Reasoning model
A model that spends more time “thinking” before answering, breaking the problem into steps. It tends to do better on complex tasks, at the cost of being a little slower.
- Sandbox
A space separated from the rest of the system where a program or agent can run actions without touching your files or your computer.
Example: An agent that runs commands inside a Docker container.
- Skill
A package of instructions and resources that teaches an agent to do a specific task in a repeatable way.
Example: A skill for summarising meeting minutes in a fixed format.
- System instructions
Fixed instructions given to an assistant before the conversation to set its role, tone or limits. Many apps let you add custom instructions.
Example: “Always answer in British English and with short sentences.”
- Temporary chat
A conversation mode that isn’t saved to your history or used to train models. In ChatGPT, OpenAI may keep a copy for up to 30 days for safety.
- Text-to-image
Models that create an image from a written description. The main assistants include this feature, with limits on their free plans.
Example: “Generate a watercolour of a lighthouse at sunset.”
- Token
A small chunk of text (a word or part of one) that models work with. AI limits and prices are measured in tokens.
- Training
The process by which a model learns from large amounts of data. Many services let you stop your conversations from being used for it.
- Verification
Checking what an AI tells you against a reliable source (an official website, a public body, the original document) before you use it.
- VRAM (video memory)
The graphics card’s memory. When running AI models locally, the more of the model fits in VRAM, the faster it responds.
Example: A model that recommends 8 GB of VRAM.
- Workflow
A set of connected nodes that defines, step by step, how a result is generated. It can be saved as a file and reused.
Example: A text-to-image workflow saved as JSON.