Plain English
Plain-English answers to AI jargon
New AI words seem to arrive every week, and not knowing them can make the whole topic feel out of reach. This glossary explains 80 AI terms in plain English, one short definition at a time, so you can look up just the word you need and get back to work.
80 AI terms
15 explained in depth
A
- Agentic workflow
- Also called: agentic AI, AI workflow
- A multi-step process in which AI models and tools work through a task together, with some steps decided by the AI itself.
- AI acceptable use policy
- Also called: AI policy, AI usage policy
- A short written guide telling staff which AI tools they may use, for what tasks, and what information must never be entered.
- AI agent
- Also called: agent, autonomous agent
- An AI system that works toward a goal by planning steps and using tools, such as search or email, with limited human input.
- Read more: AI agent
- AI governance
- Also called: AI risk management, responsible AI
- The rules, roles and checks an organization uses to decide which AI to use, how, and who is responsible when things go wrong.
- AI literacy (EU AI Act Article 4)
- Also called: AI skills, AI training
- Having the skills, knowledge and understanding to use AI systems sensibly and to recognize their opportunities, risks and possible harms.A plain-English summary, not legal advice.
- AI model
- Also called: model
- The trained core of an AI system: the part that takes an input and produces an answer, prediction or piece of content.
- Anonymization
- Also called: anonymisation, anonymized data
- Changing data so that the people in it can no longer be identified by anyone, in a way that cannot be reversed.
- API (application programming interface)
- Also called: API, API key
- A set of rules that lets one piece of software request data or actions from another, without a person clicking through screens.
- Artificial intelligence (AI)
- Also called: AI, AI system
- Computer systems that produce predictions, content, recommendations or decisions from the information they are given, tasks that used to need human judgment.
- Read more: Artificial intelligence (AI)
- Automated decision-making (GDPR Article 22)
- Also called: ADM, solely automated decisions, profiling
- Decisions about people made entirely by technology, with no meaningful human involvement.A plain-English summary, not legal advice.
- Automation bias
- Also called: over-reliance on AI
- The tendency to trust an automated system's answer too much, even when it is wrong or other evidence disagrees.
B
- Benchmark
- Also called: AI benchmark, eval
- A standard test used to compare how well different AI models perform on the same set of tasks.
- Bias
- Also called: AI bias, algorithmic bias, harmful bias
- When an AI system's results are unfairly skewed for or against certain people or groups, often because of patterns in its data.
- Biometric data
- Also called: biometrics
- Personal data about a person's body or behavior, such as their face, fingerprints or voice, used to recognize or identify them.
C
- Chatbot
- Also called: AI assistant, conversational AI, virtual assistant
- A program you talk to in a chat window or by voice, which answers questions or completes simple tasks in conversation.
- Read more: Chatbot
- Computer vision
- Also called: image recognition
- The branch of AI that lets computers find useful information in photos, videos and other images.
- Content provenance and watermarking
- Also called: watermarking, content credentials, provenance
- Ways of recording or marking where a piece of digital content came from and whether AI created or changed it.
- Context window
- Also called: context length, context limit
- The maximum amount of text, measured in tokens, that a model can consider at once, including your files, the conversation and its reply.
- Read more: Context window
- Copilot
- Also called: AI copilot, Microsoft Copilot
- An AI assistant built into software you already use, which suggests text, answers or actions while you stay in control.
- Custom assistant (custom GPT)
- Also called: custom GPT, custom AI assistant, custom agent
- A version of a general AI assistant that you set up with your own instructions, files and tools for a specific job.
D
- Data controller
- Also called: controller
- The person or organization that decides why and how personal data is used, and carries main responsibility for it under data protection law.
- Data leakage
- Also called: data leak, data exposure
- When private or confidential information ends up somewhere it should not, for example in an AI tool or in the answers it gives.
- Data minimization
- Also called: data minimisation
- The principle of collecting and sharing only the personal data you actually need for a specific purpose.
- Data processing agreement (DPA)
- Also called: DPA, data processing addendum, Article 28 contract
- A written contract between a business and a provider that handles personal data for it, setting out how that data may be used and protected.
- Data processor
- Also called: processor, sub-processor
- An outside organization that handles personal data only on behalf of, and following the instructions of, the controller.
- Deepfake
- Also called: deep fake, synthetic media
- AI-made or AI-altered image, audio or video that falsely makes a real person or event look or sound genuine.
- Read more: Deepfake
- Deployer (EU AI Act)
- Also called: AI deployer, AI user
- Under the EU AI Act, a person or organization that uses an AI system under its own authority, other than for purely personal activity.A plain-English summary, not legal advice.
E
- Embedding
- Also called: vector embedding, vector
- A list of numbers that represents the meaning of a piece of text or an image, so a computer can find similar items.
- EU AI Act
- Also called: AI Act, Regulation (EU) 2024/1689
- The European Union's law on artificial intelligence, which sets rules according to how much risk an AI use poses to people.A plain-English summary, not legal advice.
- Read more: EU AI Act
- Explainability
- Also called: explainable AI, XAI, interpretability
- How well people can understand why an AI system produced a particular result.
F
- Fine-tuning
- Also called: model customization
- Extra training that adjusts an existing model with your own examples so it gets better at a specific task or style.
- Read more: Fine-tuning
- Foundation model
- Also called: base model
- A large model trained on broad data that can be adapted for many different tasks instead of being built for just one.
G
- GDPR (General Data Protection Regulation)
- Also called: GDPR, UK GDPR, EU data protection law
- The European Union's main data protection law, which sets rules for how organizations collect, use and protect personal data.A plain-English summary, not legal advice.
- General-purpose AI model (GPAI)
- Also called: GPAI, GPAI model
- An EU AI Act term for an AI model trained on broad data that can handle many different tasks and be built into many products.A plain-English summary, not legal advice.
- Generative AI
- Also called: GenAI, GAI
- AI that creates new content, such as text, images, audio or video, based on patterns learned from large amounts of existing material.
- Read more: Generative AI
- Guardrails
- Also called: safety filters, safeguards, content filters
- Rules and checks placed around an AI tool to keep its outputs safe, accurate and within the limits you set.
H
- Hallucination
- Also called: confabulation, AI making things up
- When an AI tool states something false or made up as if it were true.
- Read more: Hallucination
- High-risk AI system
- Also called: high-risk AI
- An EU AI Act category for AI uses that could seriously affect people's health, safety or rights, which must meet strict requirements.A plain-English summary, not legal advice.
- Human in the loop
- Also called: HITL, human review
- A setup where a person reviews, approves or can correct an AI system's output before it has a real effect.
- Human oversight (EU AI Act Article 14)
- Also called: human oversight
- An EU AI Act requirement that high-risk AI systems be designed so people can monitor them, understand their limits and step in or stop them.A plain-English summary, not legal advice.
I
- Inference
- Also called: model inference, running a model
- The moment a trained model is actually used to produce an answer, as opposed to the earlier stage when it was trained.
J
- Jailbreak
- Also called: jailbreaking
- A deliberate attempt to get an AI tool to ignore its safety rules or instructions by phrasing requests in a tricky way.
K
- Knowledge cutoff
- Also called: training cutoff, training data cutoff
- The date after which a model has little or no information, because its training data stops around then.
L
- Large language model (LLM)
- Also called: LLM, language model
- An AI model trained on huge amounts of text so it can read, write, summarize and answer questions in everyday language.
- Read more: Large language model (LLM)
- Latency
- Also called: response time
- The waiting time between sending a request to an AI tool and getting its response.
M
- Machine learning
- Also called: ML
- A way of building AI where a computer finds patterns in example data instead of following rules a person wrote by hand.
- Read more: Machine learning
- Model Context Protocol (MCP)
- Also called: MCP, connectors
- An open standard that lets AI assistants connect to outside tools and data, such as files, calendars or databases, in a consistent way.
- Multimodal AI
- Also called: multimodal model
- AI that can work with more than one type of input or output, such as text, images, audio or video, in the same task.
N
- Natural language processing (NLP)
- Also called: NLP
- The branch of AI that helps computers read, understand and produce human language, written or spoken.
- No-code and low-code
- Also called: no-code, low-code
- Tools that let people build apps, automations or AI assistants with visual editors and plain language instead of writing programming code.
O
- Open-weight model
- Also called: open model, open-source model
- An AI model whose trained internal settings, called weights, are published so anyone can download and run it themselves.
- Optical character recognition (OCR)
- Also called: OCR, text recognition
- Technology that reads printed or handwritten text in images or scans and turns it into text a computer can edit and search.
P
- Personal data
- Also called: personal information, personally identifiable information, PII
- Any information that relates to an identified or identifiable living person, such as a name, email address, photo or customer number.
- Pilot project
- Also called: proof of concept, POC, trial run
- A small, time-limited test of an AI tool on one real task, to see whether it works before a wider rollout.
- Privacy by design
- Also called: data protection by design, privacy by default
- Building privacy protections into a product or process from the start, rather than adding them after problems appear.
- Prohibited AI practices
- Also called: banned AI practices, unacceptable risk
- AI uses that the EU AI Act bans outright because they are considered an unacceptable risk to people's safety or rights.A plain-English summary, not legal advice.
- Prompt
- Also called: instruction, query
- The instruction, question or material you give an AI tool to get a response.
- Read more: Prompt
- Prompt engineering
- Also called: prompt design, prompting
- The practice of writing and testing prompts so an AI tool produces the results you need more reliably.
- Read more: Prompt engineering
- Prompt injection
- Also called: indirect prompt injection
- An attack where hidden instructions in text, a web page or a file trick an AI tool into doing something you did not intend.
- Provider (EU AI Act)
- Also called: AI provider
- Under the EU AI Act, whoever develops an AI system or model, or has one developed, and offers it under their own name.A plain-English summary, not legal advice.
- Pseudonymization
- Also called: pseudonymisation, de-identification
- Replacing names and other direct identifiers with codes, while the key that links them back to real people is kept separately.
R
- Rate limit
- Also called: usage limit, quota, spend limit
- A cap on how many requests, or how much text, you can send to an AI service within a set period of time.
- Reasoning model
- Also called: thinking model
- A type of AI model that works through a problem step by step internally before giving its final answer.
- Red teaming
- Also called: red team, adversarial testing
- Testing an AI system by deliberately trying to make it fail, misbehave or leak information, so problems are found before real users find them.
- Retrieval-augmented generation (RAG)
- Also called: RAG, grounding
- A method where an AI tool first looks up relevant information from a trusted source, such as your documents, and then uses it to answer.
- Read more: Retrieval-augmented generation (RAG)
- Return on investment (ROI)
- Also called: ROI
- A measure of whether something pays off: what you gain from it compared with what it costs.
S
- Seat license
- Also called: per-user license, per-seat pricing, subscription seat
- Paying a fixed fee for each named person who is allowed to use a software tool, usually monthly or yearly.
- Sentiment analysis
- Also called: opinion mining
- Using AI to judge whether a piece of text sounds positive, negative or neutral.
- Shadow AI
- Also called: unapproved AI use, bring your own AI
- Staff using AI tools for work without the knowledge or approval of the business, often through personal accounts.
- Synthetic content labeling (EU AI Act Article 50)
- Also called: AI labeling, AI disclosure, transparency obligations
- Rules that make AI visible: people must be told when they deal with an AI system, a deep fake or certain AI-written text on public matters, and AI-generated content must be marked in a machine-readable way.A plain-English summary, not legal advice.
- System prompt
- Also called: system message, custom instructions
- Background instructions, set before a conversation starts, that shape how an AI assistant behaves, such as its role, tone and rules.
T
- Temperature
- Also called: randomness setting
- A setting that controls how predictable or varied an AI model's answers are: low is more consistent, high is more creative.
- Token
- Also called: tokens
- A small chunk of text, often part of a word, that language models read and write, and that providers use to measure usage and price.
- Read more: Token
- Total cost of ownership (TCO)
- Also called: TCO
- The full cost of a tool over time, including setup, training, maintenance and staff time, not just the subscription price.
- Training data
- Also called: training set, dataset
- The examples, such as text, images or records, that an AI model learns from before it is used.
- Training opt-out
- Also called: opt out of model training, data usage controls
- A setting or contract term that stops an AI provider from using your prompts and files to train or improve its models.
- Transcription
- Also called: speech-to-text, automatic speech recognition
- Turning spoken audio, such as a meeting, call or voice note, into written text using AI.
U
- Usage-based pricing
- Also called: pay-as-you-go, consumption-based pricing, per-token pricing
- Paying for AI according to how much you use it, such as per token or per request, rather than a flat fee.
V
- Voice cloning
- Also called: synthetic voice, AI voice
- Using AI to copy a real person's voice so it can say new things they never said.
W
- Workflow automation
- Also called: business process automation, automation
- Using software to carry out a repeatable series of business steps automatically, with or without AI.
No terms match that search, so try a shorter word or clear the filter.
Last checked How we check this
How we check this
Each definition is written in plain English from named sources, such as the EU AI Act, NIST and OECD publications, and the providers' own documentation. Those sources are listed on each term page, with the date the term was last checked. Definitions are kept short on purpose, so the sources are the place for the full detail. Meanings shift as tools change, so a definition that looks out of date is worth flagging through the Tell us about a change link on any term page.