AI Glossary

AI used in agent form inside real workflows: systems that take actions across software, not just generate text for a person to copy.

How widely people or organisations actually use AI, usually measured by surveys asking whether they used an AI tool in a given period.

An AI system that carries out a multi-step task on its own: it plans, uses tools such as a browser or code, checks results and keeps going until the goal is met.

A processor built for the maths of AI, mainly multiplying large grids of numbers. GPUs are the most common type; others include TPUs and custom chips from the large cloud companies.

A chatbot designed for ongoing personal conversation and emotional connection rather than for tasks or information.

A real-world event in which an AI system caused or nearly caused harm, such as a wrongful arrest, a dangerous instruction or a deepfake scam, as recorded by trackers like the AI Incident Database.

A patent whose claims cover an AI method or application, identified by patent classification codes and keywords.

Laws and binding rules that govern how AI is built and used, such as transparency duties, bans on certain uses, safety testing or rules for AI in hiring and elections.

Abilities such as machine learning, natural language processing or using generative AI tools, tracked through job postings and online profiles.

The share of a country's workers, on a platform such as LinkedIn, who have AI engineering skills or AI job titles.

A medical device that uses AI, for example to read scans or flag heart rhythms, and has been authorised by a regulator such as the US FDA.

A company's current monthly revenue multiplied by 12. AI companies often report it because their revenue is growing too fast for annual figures to keep up.

A benchmark of visual puzzles that people find easy but AI finds hard, designed to test whether a system can learn a new rule from a few examples.

Computer systems that perform tasks that normally need human intelligence, such as understanding language, recognising images or making decisions. Most AI today is machine learning: systems that learn patterns from data rather than following hand-written rules.

A standard test used to compare AI systems, such as a set of exam questions, coding problems or tasks, scored the same way for every model.

A program you talk to in ordinary language. Modern chatbots such as ChatGPT, Claude and Gemini run on large language models.

A public leaderboard where people compare two anonymous chatbot answers and vote for the better one; the votes are turned into Elo-style ratings.

A building full of servers, storage and networking. AI data centres pack in thousands of AI chips and need large amounts of electricity and cooling.

Rules that require certain data to be stored or processed inside a country, or limit sending it abroad.

Realistic fake audio, images or video of a real person, made with AI.

A computer model of a real thing, such as a patient's heart or a factory, kept up to date with real data so that changes can be tested on the model first.

A score that rises when you beat stronger opponents and falls when you lose to weaker ones, first used in chess. AI leaderboards use it to rank models from head-to-head votes.

The European Union's law on artificial intelligence, which sets rules by risk level: some uses are banned, high-risk systems face strict duties, and general-purpose models must meet transparency requirements. It entered into force in August 2024 and applies in stages.

A single arithmetic operation on decimal numbers. Training compute is counted in total FLOP; chip speed in FLOP per second (FLOP/s).

A large model trained on broad data that can be adapted to many tasks, such as a large language model that powers both a chatbot and a coding assistant.

One of the most capable AI models available at a given time, usually from a handful of leading labs.

A model's ability to call outside tools, such as a search engine, calculator or company database, by producing a structured request that software then runs.

AI that creates new content, such as text, images, audio, video or code, in response to a prompt.

A benchmark of hard science questions written by PhD experts; the 'Diamond' set contains the questions experts answer correctly but skilled non-experts usually miss, even with web search.

A chip originally designed for computer graphics that can do many calculations at once, which makes it well suited to training and running AI models. Nvidia makes most of those used for AI.

When an AI model states something false or unsupported as if it were true, such as a made-up citation or a wrong figure.

The score people achieve on a benchmark, used as a reference point for AI. Depending on the test it may be average people or domain experts.

A programmable machine with movable arms used in manufacturing for tasks such as welding, assembly and handling, as counted by the International Federation of Robotics.

Running a trained model to produce an answer. Training happens once; inference happens every time someone uses the model.

A model trained on very large amounts of text to predict the next word, which lets it write, summarise, translate, answer questions and write code. ChatGPT, Claude and Gemini are built on LLMs.

The main approach to AI today: instead of following hand-written rules, a system learns patterns from examples and uses them to make predictions on new data.

Massive Multitask Language Understanding: a benchmark of multiple-choice questions across dozens of school and professional subjects. MMLU-Pro is a harder version with more answer options.

A model that works with more than one kind of input or output, such as text, images, audio and video, in the same system.

A model made of layers of simple connected units whose connection strengths (parameters) are adjusted during training. Deep learning means neural networks with many layers.

A model that Epoch AI includes in its database because it was state of the art, highly cited, historically important or widely used.

A model whose trained parameters are published so anyone can download and run it. Open weights are not the same as open source: training data and code may still be closed.

AI released with its code, and ideally its weights and training data details, under a licence that lets others use, study and change it.

The numbers inside a neural network that are adjusted during training; together they store what the model has learned. Model size is usually given as a parameter count.

Money that venture capital and other private investors put into AI companies, as tracked from funding rounds.

A model trained on protein sequences, in the way language models learn from text, used to predict protein structure and function or design new proteins.

A model trained to work through a problem step by step before answering, spending more computation at answer time on hard questions.

Practices for building and using AI safely and fairly: testing for bias and errors, protecting privacy, being transparent and having human oversight.

A self-driving car that carries paying passengers with no safety driver at the wheel.

A benchmark that tests whether an AI system can fix real bugs from open-source projects on GitHub; 'Verified' is a subset checked by human engineers.

A list of numbers that represents the meaning of a piece of text, so that similar texts get similar numbers. Embeddings power search and retrieval systems.

The total amount of calculation used to train a model, counted in FLOP.