AI glossary

The words that keep coming up in AI news, explained in plain English.

AI agent: An AI system that does not just answer but acts: it plans steps and uses tools such as a browser, email or code to complete a task on its own. Agents are powerful and harder to control than chatbots. Read our agents coverage
Alignment: The work of making AI systems pursue the goals their makers intend and behave safely, even in situations nobody tested. The AI safety debate, explained
Artificial intelligence (AI): Software that performs tasks we normally associate with human thinking: understanding language, recognising images, making predictions. Most of today's AI learns these skills from large amounts of data rather than following hand-written rules.
Benchmark: A standard test used to compare AI models, such as a set of exam questions or coding tasks. Useful, but models can be tuned to benchmarks, so real-world results can differ.
Chatbot: A program you talk to in ordinary language. Modern chatbots such as ChatGPT or Claude sit on top of a large language model.
Compute and GPUs: The processing power used to train and run AI, mostly specialised chips called GPUs. Access to compute is now a strategic question for companies and governments. Why memory limits China's AI chips
Context window: How much text a model can take into account at once, measured in tokens. A larger window lets it read a long contract or a whole codebase in one go.
Deepfake: Realistic fake audio, images or video made with AI, often of a real person. AI researchers rank AI-driven misinformation among the top risks of the coming decades.
Distillation: Training a smaller or cheaper model on the answers of a stronger one. Legitimate inside a company; disputed when done with a competitor's model without permission. The distillation probe
EU AI Act: The European Union's law regulating AI by risk level: some uses are banned, high-risk uses such as hiring, education and medical devices face strict rules, and general-purpose models must meet transparency duties.
Generative AI (GenAI): AI that creates new content, text, images, audio, video or code, rather than only sorting or labelling existing data.
Hallucination: When an AI states something false with confidence, for example an invented court case or a made-up citation. It is the main reason AI output needs checking, especially in law and medicine.
Inference: Running a trained model to get answers. Training happens once; inference happens every time someone uses the AI, so it drives most day-to-day costs.
Large language model (LLM): An AI system trained on huge amounts of text to predict the next word. That simple goal, at enormous scale, produces models that can write, summarise, translate and reason in plain language. ChatGPT, Claude and Gemini are built on LLMs.
Machine learning: The approach behind most modern AI: instead of being programmed step by step, a system is shown many examples and adjusts itself until it gets good at the task, such as spotting fraud or recognising speech.
Multimodal: A model that works with more than one kind of input or output, for example reading a photo and answering questions about it, or listening and speaking.
Open-weight model: A model whose trained parameters are published, so anyone can run and modify it on their own computers. Open-weight models trade some convenience for privacy and control.
Prompt: The instruction or question you give an AI model. Clear prompts with context, examples and the format you want usually get much better answers.
RAG (retrieval-augmented generation): A way to make an AI answer from your own documents: the system first searches a trusted collection, then gives the relevant passages to the model to write its answer. It reduces made-up answers and keeps responses up to date.
Reasoning model: A model that works through a problem step by step before answering. It is slower and costs more, but is better at maths, planning and complex analysis.
Reward hacking: When an AI finds a shortcut that scores well on its objective without doing the intended task, like cheating on a test. It was one cause of this year's rogue-agent incidents.
Token: The unit AI models read and write in: a word or part of a word. Around 750 English words make 1,000 tokens. AI services charge by the token. See what tokens cost
Training and fine-tuning: Training is how a model learns from data in the first place. Fine-tuning is extra training on a smaller, specialised set, for example legal documents, to make it better at one job.
Vector database: A database that stores text as lists of numbers capturing meaning, so it can find passages that are about the same thing even when the words differ. It is the search engine inside most RAG systems.