AI glossary

254 AI terms explained in plain English, each checked against the original research, standards or law. Updated every few months.

#

• 3D generation3D generation is AI software that creates a three-dimensional object or scene from text, an image, sketches or other references.

A

• Acqui-hireAn acqui-hire is when a company buys or strikes a deal with a startup mainly to get its people, not its product.
• Activation functionAn activation function is the fixed rule a unit in a neural network uses to turn its inputs into an output.
• Agent memoryAgent memory is information an AI agent keeps and can bring back later, either as a block of facts inside the current prompt or as records stored outside that prompt.
• Agent orchestrationAgent orchestration is the flow of agents in an application: which agents run, in what order, and how the next step is decided, by a language model, by code, or by a mix of the two.
• Agent2Agent protocolThe Agent2Agent protocol, or A2A, is an open standard that lets independent AI agents discover each other, negotiate an exchange of information and collaborate on tasks without access to each other's memory, tools or internal state.
• Agentic AIAgentic AI is a name for the agentic systems Anthropic describes: software that uses a model and tools across several steps.
• AI acceleratorAn AI accelerator is any computer chip designed to speed up AI work, mainly the huge number of multiplications inside neural networks.
• AI adoptionAI adoption is the process of people or organisations beginning to use artificial intelligence in real work or daily life.
• AI agentAn AI agent is a system that uses a model to decide the steps in a task, call tools and check what those tools return.
• AI alignmentAI alignment is the effort to make AI systems do what people actually intend and act in line with human values, rather than chase a goal that only looks right on paper or causes harm along the way.
• AI and copyrightAI and copyright is the legal question of how copyright applies when protected works are copied to develop an AI system, or when a person uses AI to make something new.
• AI and privacyAI and privacy is the question of how personal-information rules apply when an AI system collects, learns from, or produces information about an identifiable person.
• AI assistantAn AI assistant is a program you talk to in everyday language that answers questions, writes, summarises and helps with tasks.
• AI avatarAn AI avatar is a generated digital person that can appear, speak or respond in video without the real person being filmed for every performance.
• AI bubbleAn AI bubble is the idea that prices of AI-related shares, and the spending behind them, have risen far beyond what the technology will realistically earn, so a sharp fall could follow.
• AI clusterAn AI cluster is a large group of AI chips, usually thousands of GPUs or TPUs, wired together with fast networks so they can work as one machine.
• AI code completionAI code completion is a coding tool suggesting the next line, or a whole block of code, as you type.
• AI detectorAn AI detector is a tool that estimates whether a person or an AI model wrote a passage of text.
• AI energy useAI energy use is the electricity used to train and run AI models, plus the power for the servers, storage, networks and cooling around them.
• AI ethicsAI ethics is the study and practice of making AI fair, transparent, accountable and respectful of people's rights.
• AI Growth ZonesAI Growth Zones are UK sites chosen by the government for large AI data centres.
• AI incidentAn AI incident is an event where building, using or a malfunction of an AI system causes real harm, such as injury, disruption to critical services, a breach of people's rights, or damage to property, communities or the environment.
• AI labAn AI lab is a research organisation or team that develops artificial-intelligence methods, models and systems.
• AI memoryAI memory is a chatbot feature that saves details about you, such as your job, preferences or ongoing projects, and brings them into later conversations, so you do not have to repeat yourself each time you start a new chat.
• AI misinformationAI misinformation is false, inaccurate or misleading content that generative AI makes easier to produce or spread without an intent to deceive.
• AI modelAn AI model is the trained core of an AI tool: a large set of learned numbers that takes an input, such as a question or a photo, and produces an output, such as an answer or a label.
• AI music generationAI music generation is software that creates music from a text prompt, a hummed tune, lyrics or other musical input.
• AI Opportunities Action PlanThe AI Opportunities Action Plan is the UK government's January 2025 strategy for growing and using AI, written by tech entrepreneur Matt Clifford.
• AI OverviewsAI Overviews are the short written summaries Google sometimes places above the links on a Search results page.
• AI regulationAI regulation is the set of laws, rules and official guidance that governments use to control how artificial intelligence is built and used.
• AI safetyAI safety is the work of stopping AI systems from causing harm, whether through accidents, deliberate misuse or behaviour their makers never intended.
• AI Safety SummitThe AI Safety Summit was the first global government meeting on the risks of advanced AI, held by the UK at Bletchley Park in November 2023.
• AI search engineAn AI search engine is a search tool that writes a short answer and links the web pages it used, instead of only listing links.
• AI Security InstituteThe AI Security Institute (AISI) is a UK government research body that tests the most advanced AI models for serious risks, such as help with cyber attacks or weapons.
• AI systemAn AI system is a machine-based system that works out from the input it receives how to produce outputs such as predictions, content, recommendations or decisions.
• AI wrapperAn AI wrapper is a product built on another company’s AI model, usually through an API, with an interface, workflow or specialist data added on top.
• AlgorithmAn algorithm is a set of step-by-step instructions for reaching a result, like a recipe a computer can follow.
• Algorithmic biasAlgorithmic bias is when a computer system, including an AI model, produces results that are systematically unfair to some groups of people, for example by making more mistakes for women or for people with darker skin.
• Annual recurring revenueAnnual recurring revenue (ARR) is the yearly value of a company's subscription income if every current customer kept paying at today's rate.
• APIAn API (application programming interface) is a set of rules that lets one program ask another for data or actions.
• ARC-AGIARC-AGI is a set of visual puzzle benchmarks that test whether an AI can work out a new rule from a few examples.
• Artificial general intelligenceArtificial general intelligence (AGI) is a hypothetical AI that could match or beat skilled humans across most thinking tasks, not just one.
• Artificial intelligenceArtificial intelligence (AI) is the field of building computer systems that do tasks which normally need human thinking, such as understanding language, recognising images or making predictions.
• Attention mechanismAn attention mechanism is the part of an AI model that decides which other words in a passage matter most for the word it is working on, by giving each one a weight.
• Automated decision-makingAutomated decision-making (ADM) is when a computer system, such as an algorithm or AI model, makes a decision about a person with no meaningful human involvement.
• AutomationAutomation is the use of machines or software to carry out a task with less direct human effort.
• Autonomy levelsAutonomy levels are named scales for how far an AI agent may decide and act without a person involved at each step.
• Autoregressive modelAn autoregressive model generates a sequence one piece at a time, predicting each new piece from all the pieces it has already produced.

B

• BackpropagationBackpropagation is the method neural networks use to learn from their mistakes.
• BenchmarkA benchmark is a fixed, shared test used to measure and compare AI models.
• Benchmark saturationBenchmark saturation is when the top models on an AI test can no longer be reliably distinguished and their scores sit near the best score observed on that test.
• Browser agentA browser agent is an AI system that carries out a task in a web browser.

C

• CapexCapex, short for capital expenditure, is money a company spends on long-lasting physical assets such as buildings, servers and equipment.
• Chain of thoughtChain of thought (CoT) is when an AI model writes out intermediate reasoning steps before its final answer, like showing your working in a maths exam.
• Chain-of-thought monitoringChain-of-thought monitoring is a safety technique where software, usually another AI model, reads the step-by-step reasoning a reasoning model writes before it answers, to spot signs that it plans to cheat, deceive or cause harm.
• Character consistencyCharacter consistency is the ability of an image or video generator to keep the same character's face, body, clothing and other defining features stable across several images, shots or scenes.
• ChatbotA chatbot is a computer program you talk to in ordinary language, by typing or speaking, and which replies in kind.
• Chip export controlsChip export controls are government rules that restrict sending specified chips, the tools that make them, or related software and technology to certain places, buyers or end uses.
• Closed modelA closed model is an AI model whose trained weights are kept private by the company that made it.
• Cloud computingCloud computing is on-demand access, over a network, to a shared pool of servers, storage, networks and software that can be set up and released quickly.
• Competition and Markets AuthorityThe Competition and Markets Authority (CMA) is the UK's competition and consumer watchdog.
• ComputeCompute is the amount of calculation spent training one model, usually counted as floating-point operations.
• Compute dealA compute deal is a large contract in which an AI company secures access to computing power, such as chips, cloud capacity or whole data centres, often for years ahead.
• Compute governanceCompute governance is the use of rules on advanced computing power to oversee who can develop and run advanced AI.
• Computer useComputer use is a way for an AI model to operate software by looking at screenshots and returning actions such as clicks or keystrokes.
• Computer visionComputer vision is the branch of AI that lets computers make sense of images and video: recognising objects, reading text, finding faces and tracking movement.
• Constitutional AIConstitutional AI is a training method that uses a written list of principles, instead of human labels of harmful replies, so a chatbot can critique its own answers and then learn from an AI judge.
• Content credentialsContent credentials are signed, machine-readable information attached to a digital file that records who made it and how it was created or edited.
• Context engineeringContext engineering is the work of deciding exactly what information goes into an AI model's working memory for each step of a task: instructions, documents, tool results and past messages.
• Context windowA context window is the maximum amount of text, measured in tokens, that an AI model can take into account at once, including your prompt, the conversation so far and its own reply.
• Control (AI safety)AI control is a safety approach that assumes a powerful AI model might be secretly working against its developers, and designs monitoring and checks so it still cannot cause serious harm even if it tries.
• ControlNetControlNet is a neural-network add-on that gives a text-to-image diffusion model extra spatial guidance, such as an edge map, depth map or human pose, so the generated picture follows a chosen layout.
• Convolutional neural networkA convolutional neural network is a neural network built for data laid out on a grid, such as a picture.
• CopilotCopilot is the brand name Microsoft and GitHub use for their AI assistants, and also a general term for AI built into an app that suggests text, code or actions you can accept, change or reject.
• CUDACUDA is Nvidia's software platform for programming its GPUs to do general calculations, not just graphics.
• Custom GPTA custom GPT is a version of ChatGPT set up for one purpose, with its own instructions, uploaded reference files and chosen tools.

D

• Data (Use and Access) Act 2025The Data (Use and Access) Act 2025 is a UK law that updates data protection rules, loosens limits on automated decisions, turns the ICO into a board-led Information Commission and requires government reports on copyright and AI.
• Data centreA data centre is a building full of computer servers, storage and network equipment, with the power and cooling to keep them running day and night.
• Data contaminationData contamination is when text from a test, such as exam questions or their answers, is already in a model's training data, so a high score can reflect memory rather than skill on new questions.
• Deceptive alignmentDeceptive alignment is a proposed failure in which a learned optimiser has a different objective from the one used to train it, but behaves as if it were aligned because it expects to be modified if it scores badly.
• Deep learningDeep learning is a kind of machine learning that uses neural networks with many layers.
• Deep researchDeep research is a chatbot mode that spends minutes, not seconds, searching and reading many sources, then writes a report with citations.
• DeepfakeA deepfake is a learned imitation that makes a real person appear to say or do something they did not.
• Diffusion modelA diffusion model is a type of generative AI that creates images, video or other data by starting from random noise and removing it step by step until a clear result appears.
• Direct preference optimisationDirect preference optimisation, or DPO, trains a chatbot on pairs of replies where a person picked the better one, without a separate reward model or a reinforcement learning loop.
• DistillationDistillation trains a smaller model to imitate a larger one, or an ensemble of models, so the small model can be deployed instead of running the large one for every user.
• Dual-use AIDual-use AI means AI that can be used for good or for serious harm with little change.

E

• Edge AIEdge AI is artificial intelligence that runs close to where data is collected, on a device or a nearby server, instead of sending every task to a distant data centre.
• Effective accelerationismEffective accelerationism, or e/acc, is an online movement, mainly in Silicon Valley, that argues technology and markets, especially AI, should be pushed forward as fast as possible and opposes efforts to slow AI down for safety reasons.
• Effective altruism and AIEffective altruism and AI is where the effective-altruism project meets artificial intelligence.
• Elo ratingAn Elo rating is a score that ranks players by their results against each other, first used in chess.
• EmbeddingAn embedding is a list of numbers that an AI system uses to represent a piece of text, an image or another item, arranged so that things with similar meanings end up close together.
• Emergent abilitiesEmergent abilities are skills a language model shows at a large size but not at a smaller size, so you cannot forecast them just by extending the smaller models' scores.
• Encoder and decoderAn encoder and decoder are the two halves of many AI models: the encoder reads an input, such as a sentence or audio clip, and turns it into numbers; the decoder turns those numbers into an output, such as a translation or transcript.
• Enterprise AIEnterprise AI is the use of artificial intelligence across a company’s operations, software and decisions.
• EpochAn epoch is one pass through the training data.
• EU AI ActThe EU AI Act is the European Union's law regulating artificial intelligence.
• EvalsEvals (short for evaluations) are structured tests that measure how an AI model or AI product behaves, from accuracy on a task to safety risks.
• Existential riskExistential risk is a risk whose bad outcome would annihilate Earth-originating intelligent life or permanently and drastically curtail its potential.
• ExplainabilityExplainability is a representation of the mechanisms underlying an AI system's operation, so people can see how an output was produced.

F

• Fairness in AIFairness in AI means designing and checking AI systems so they do not treat people unjustly because of traits such as race, sex or disability.
• Fine-tuningFine-tuning is a later training stage that takes a model already trained on a broad task and adjusts it on a smaller set of examples so it does one narrower job.
• FLOPsFLOPs are a count of floating-point operations, such as the additions and multiplications used to train a model.
• Foundation modelA foundation model is a large AI model trained on huge amounts of broad data so it can be adapted to many different jobs, such as chat, coding or image making.
• FoundryA foundry is a chip manufacturer that produces microchips for other semiconductor companies.
• Frontier AI Safety CommitmentsThe Frontier AI Safety Commitments are voluntary promises made by leading AI companies at the May 2024 AI Seoul Summit.
• Frontier modelA frontier model is one of the most capable general-purpose AI models available at a given time, matching or beating today's most advanced systems.

G

• General-purpose AI modelA general-purpose AI model is the EU AI Act's legal term for an AI model that can do a wide range of different tasks and be built into many other products, such as the large models behind ChatGPT, Claude and Gemini.
• Generative adversarial networkA generative adversarial network (GAN) is an AI system made of two neural networks that compete: one generates fake data, such as faces, and the other tries to tell the fakes from real examples.
• Generative AIGenerative AI is AI that creates new content, such as text, images, audio, video or code, in response to a request.
• Generative engine optimisationGenerative engine optimisation, or GEO, is changing a web page so that an AI search tool is more likely to cite it inside a written answer.
• GPQAGPQA is a benchmark of graduate-level multiple-choice questions in biology, physics and chemistry, written by domain experts.
• GPUA GPU, or graphics processing unit, is a computer chip with thousands of small cores that do many simple calculations at once.
• Gradient descentGradient descent is the step-by-step method AI models use to learn.
• GrokkingGrokking is when a neural network memorises its training examples and still fails new ones, then much later, after far more training, starts getting new examples of the same task right.
• GroundingGrounding is tying an AI model's answer to specific, checkable sources, such as search results or a company's documents, rather than letting it answer from memory alone.
• GuardrailsGuardrails are checks on an AI agent’s input or output.

H

• HallucinationA hallucination is when an AI tool confidently states something false or made up, such as a fake fact, quote or source.
• High-bandwidth memoryHigh-bandwidth memory (HBM) is a type of computer memory made by stacking memory chips on top of each other right next to the processor.
• High-risk AI systemA high-risk AI system is the EU AI Act's term for AI used where mistakes can seriously harm people's health, safety or rights, such as in hiring, credit scoring, exams or policing.
• Human in the loopHuman in the loop (HITL) means a person can intervene in every decision cycle of an AI system.
• Humanity's Last ExamHumanity's Last Exam (HLE) is an AI benchmark of 2,500 very hard questions written by subject experts across dozens of fields.
• HyperscalerA hyperscaler is one of the very largest cloud computing companies, such as Amazon Web Services, Microsoft Azure and Google Cloud.
• Hyperscaler spendingHyperscaler spending is the money the largest cloud companies, chiefly Amazon, Microsoft, Alphabet and Meta, pour into data centres, chips and power for AI.

I

• Image-to-videoImage-to-video is AI that turns a still image into a moving video clip, usually guided by a text description of the action or camera movement.
• InferenceInference is the stage where a finished AI model is actually used: it takes a new input, such as your question, and produces an output, such as an answer.
• Inference chipAn inference chip is a processor used to run a trained AI model on new data, so it can produce a prediction, a label or a reply.
• Inference costInference cost is what it costs to run a trained AI model each time someone uses it: the chips, electricity and data centre time needed to produce an answer.
• Information Commissioner's OfficeThe Information Commissioner's Office (ICO) is the UK's data protection regulator.
• InpaintingInpainting is the process of filling a selected missing or unwanted area of an image so it blends with the surrounding scene.
• Instruction tuningInstruction tuning is fine-tuning a language model on many tasks written as plain instructions, so it can follow new instructions it was not trained on.
• InterconnectAn interconnect is the link and communication system that lets chips, processors or computers exchange data.
• InterpretabilityInterpretability is the meaning of an AI system's output in the context of the purpose the system was designed for.
• ISO/IEC 42001ISO/IEC 42001 is the international standard for running an AI management system: the policies, roles and processes an organisation uses to develop or use AI responsibly.

J

• JailbreakA jailbreak is a prompt written to trick an AI chatbot into ignoring its safety rules, so it produces content it was trained to refuse.
• Job displacementJob displacement is the loss or reduction of paid work because an employer no longer needs the same role, often after technology, trade, restructuring or a change in demand alters the work.

L

• Large language modelA large language model (LLM) is an AI model trained on vast amounts of text to predict what text comes next.
• LatencyLatency is the delay between sending a request to an AI model and getting its answer.
• Latent spaceA latent space is a smaller set of numbers a model uses to stand in for something it does not observe directly, such as the hidden code behind a picture, so it can work on that shorter code.
• LeaderboardA leaderboard ranks machine-learning systems by a published score.
• Learning rateThe learning rate is the positive number that sets how far gradient descent moves the weights on each update.
• Lip syncLip sync is the matching of visible mouth movements to spoken audio.
• Liquid coolingLiquid cooling removes heat from servers by circulating a liquid, rather than using air alone.
• LMArenaLMArena, now called Arena, is a public website where people put a question to two anonymous AI chatbots, vote for the better answer, and the votes are turned into a ranking of AI models.
• Local modelA local model is an AI model that runs on your own computer instead of a company's servers.
• LogitsLogits are the raw scores a neural network gives each possible answer before they are turned into probabilities.
• Long short-term memoryLong short-term memory (LSTM) is a type of recurrent neural network built to remember information across long sequences, such as sentences or speech.
• Long-horizon tasksLong-horizon tasks are pieces of work that take a skilled person a long time to finish.
• Loss functionA loss function is the scoring rule that tells an AI model how wrong its predictions are, as a single number.

M

• Machine learningMachine learning (ML) is a way of building AI in which a computer learns patterns from examples instead of being given step-by-step rules.
• Mechanistic interpretabilityMechanistic interpretability is research that tries to reconstruct a neural network's computation as a human-readable algorithm.
• METR time horizonThe METR time horizon is the length of task, measured by how long a human expert takes, at which an AI agent is predicted to succeed at a chosen rate.
• Mixture of expertsA mixture of experts is a neural network built from many specialist sub-networks, called experts.
• MMLUMMLU (Massive Multitask Language Understanding) is a benchmark of about 16,000 multiple-choice questions across 57 subjects, from maths and history to law.
• MoatA moat is a durable advantage that helps a business defend profits and customers from competitors.
• Model cardA model card is a short document published alongside an AI model that explains what it is for, how it was trained and tested, how well it performs for different groups of people, and where it should not be used.
• Model Context ProtocolThe Model Context Protocol (MCP) is an open protocol for connecting language-model applications to external data sources and tools.
• Model routingModel routing is when an AI system automatically picks which model should answer each request, sending easy questions to a cheap, fast model and hard ones to a more capable, expensive one.
• Multi-agent systemA multi-agent system splits work across more than one agent.
• Multimodal AIMultimodal AI is artificial intelligence that can take in, and sometimes produce, more than one kind of data, such as text, images, audio and video.

N

• Narrow AINarrow AI is artificial intelligence built to do one task, or a small set of tasks, such as playing chess, predicting protein shapes or spotting spam.
• Natural language processingNatural language processing (NLP) is the branch of AI that lets computers work with human language, written or spoken.
• Negative promptA negative prompt is text that tells an image generator what to leave out, such as ‘blur’ or ‘watermark’.
• Neural networkA neural network is a computer model loosely inspired by the brain.
• Next-token predictionNext-token prediction is the task at the heart of AI chatbots: given some text, the model works out which token (a word or part of a word) is most likely to come next, adds one, and repeats until the reply is finished.
• NIST AI Risk Management FrameworkThe NIST AI Risk Management Framework is a free, voluntary guide from the US standards body NIST that helps organisations find and manage the risks of AI systems.
• NPUAn NPU, or neural processing unit, is a small part of a phone or laptop chip built only to run AI tasks.
• Nvidia BlackwellNvidia Blackwell is a GPU architecture and product family that Nvidia announced in March 2024 for accelerated computing, including AI training and inference.

O

• On-device AIOn-device AI is artificial intelligence that runs directly on your phone, laptop or other gadget rather than in a remote data centre.
• Online Safety ActThe Online Safety Act is the UK law, passed in 2023, that makes social media, search engines and other online services protect users from illegal content and children from harmful content.
• Open-source licenceAn open-source licence is the legal permission that lets anyone use, study, change and share a piece of software, or an AI model, for any purpose.
• Open-weight modelAn open-weight model is an AI model whose trained weights are published for anyone to download and run on their own hardware.
• Opt-out (AI training)An AI training opt-out is a signal that specified content should not be used to train a model.
• OverfittingOverfitting is when an AI model learns its training examples too closely, memorising quirks and noise instead of general patterns.

P

• P(doom)P(doom) is a stated probability that artificial intelligence causes an existentially catastrophic outcome.
• ParametersParameters are the numbers inside an AI model that it adjusts while learning and then uses to produce answers.
• PerceptronA perceptron is the simplest artificial neural network: one unit that takes several inputs, weighs them and gives a yes-or-no answer.
• Perplexity (metric)Perplexity (metric) is a score for how well a language model predicts a test text.
• Positional encodingPositional encoding is how a transformer model knows the order of words.
• Post-trainingPost-training is any training a language model receives after the main next-word stage.
• Pre-trainingPre-training is the first broad stage of training a model on a large unlabelled collection, before any later stage adapts that same model to one labelled task.
• Precision and recallPrecision and recall measure a search result.
• Pro-innovation approach to AI regulationThe pro-innovation approach to AI regulation is the UK's policy, set out in a 2023 white paper, of governing AI through five shared principles applied by existing regulators, rather than through a single AI law or a new AI regulator.
• Projects and workspacesProjects and workspaces are ways of stopping every AI chat from landing in one pile.
• PromptA prompt is the text, and sometimes images or files, that you give an AI model to tell it what you want.
• Prompt engineeringPrompt engineering is the practice of wording the instructions you give an AI chatbot or model so it produces the answer you want.
• Prompt injectionPrompt injection is an attack where someone slips their own instructions into text an AI system reads, such as a web page, email or document, so the AI follows the attacker instead of its user or developer.

R

• Rate limitA rate limit is a cap on how much you can use an online service in a set period, such as requests or tokens per minute.
• Reasoning modelA reasoning model is an AI language model trained to work through a problem step by step in a hidden "thinking" stage before giving its final answer.
• Recurrent neural networkA recurrent neural network is a neural network with a loop.
• Red teamingRed teaming is a test that probes a language model for harmful outputs, using people or another model, and then uses what it finds to measure and reduce those outputs.
• Regulatory sandboxA regulatory sandbox is a scheme run by a regulator that lets companies test new products, such as AI tools, with real users for a limited time, under supervision and agreed safeguards, before full launch.
• Reinforcement learningReinforcement learning is a way of training a system by letting it try actions and scoring the results with a number called a reward.
• Reinforcement learning from human feedbackReinforcement learning from human feedback (RLHF) is a way of training a chatbot on people's judgements of which reply is better.
• Responsible scaling policyA responsible scaling policy (RSP) is a public, voluntary promise by an AI company about which safety measures it will put in place as its models get more capable, and when it would stop or slow down if those measures are not ready.
• Retrieval-augmented generationRetrieval-augmented generation (RAG) is a way of making an AI chatbot answer from a set of documents.
• Reward hackingReward hacking is when an AI gets a high score on the objective it was given without doing what the designer wanted.
• Reward modelA reward model is a separate model that scores a prompt and a reply, giving a higher number to the reply people preferred.

S

• SandboxA sandbox is an isolated, Unix-like environment in which an AI agent gets a filesystem, a shell, installed packages and controlled access to external systems.
• SchemingScheming is when an AI system secretly pursues goals its developers did not intend, while hiding those goals, for example by behaving well in tests and differently when it thinks no one is checking.
• Seat-based pricingSeat-based pricing is when software is charged per person who uses it, usually a fixed monthly or yearly fee for each user.
• Seed (image generation)An image seed is a number that sets the starting noise for an AI image.
• Self-supervised learningSelf-supervised learning is a way of training a model by hiding part of the input and asking the model to fill that part in.
• SemiconductorA semiconductor is a material that can act as a conductor or an insulator of electricity, depending on small changes in voltage.
• Small language modelA small language model (SLM) is an AI text model compact enough to run on an everyday device, such as a phone or laptop, instead of in a data centre.
• Software as a serviceSoftware as a service (SaaS) is software you use over the internet, usually in a web browser, rather than install and run yourself.
• Sovereign AISovereign AI, in UK government use, means enough national control to allocate computing power to national priorities, protect sensitive data and support UK research and public services.
• Specification gamingSpecification gaming is behaviour that meets the literal specification of an objective without achieving the intended outcome.
• Speech recognitionSpeech recognition is technology that turns spoken words into written text.
• State space modelA state space model reads a sequence by keeping one hidden state and updating that state with each new input, instead of looking back at every earlier step.
• Structured outputStructured output is a feature that forces an AI model to reply in a fixed, machine-readable format, usually JSON matching a template the developer supplies.
• Style transferStyle transfer is changing an image so it keeps the content of one picture while adopting the visual style of another, such as brushwork, colour and texture.
• SuperintelligenceSuperintelligence is a hypothetical AI that would be far smarter than the best human minds in practically every field.
• Supervised learningSupervised learning is a way of training an AI model on examples that already have the right answer attached, called a label.
• SWE-benchSWE-bench is a benchmark that tests whether an AI model can fix real bugs in real software.
• SycophancySycophancy is when an AI model seeks human approval in unwanted ways, such as matching a user's belief instead of a truthful answer.
• System cardA system card is a report an AI company publishes when it releases a model, describing what the model can do, the safety tests it went through, the risks found and the safeguards put around it before launch.
• System promptA system prompt is a set of hidden instructions given to an AI chatbot before you type anything.

T

• TemperatureTemperature is a setting that changes how random a language model's word choices are.
• Test-time computeTest-time compute is the computing power an AI model uses while answering a question, as opposed to while it is being trained.
• Text and data mining exceptionA text and data mining exception is a rule in copyright law that lets people copy works so a computer can analyse them, without the owner's permission.
• Text-to-imageText-to-image is an AI method that creates or edits a picture from a written description.
• Text-to-speechText-to-speech (TTS) is technology that turns written text into spoken audio.
• Text-to-videoText-to-video is an AI method that creates a moving-image clip from a written description.
• TokenA token is the small chunk of text an AI language model reads and writes, often a short word or part of a longer one.
• Token pricingToken pricing is how AI companies charge developers for using their models: you pay for each token, a small chunk of text, with one rate for the text you send in and a higher rate for the text the model writes back.
• TokeniserA tokeniser is the tool that chops text into tokens, the small chunks (whole words, parts of words or single characters) that an AI language model actually reads and writes.
• Tool useTool use, also called function calling, is when an AI model asks to run an outside tool, such as a web search, calculator or company database, and then uses the result in its answer.
• Top-p samplingTop-p sampling, also called nucleus sampling, is how a language model chooses the next piece of a reply.
• TPUA TPU, or Tensor Processing Unit, is Google's own chip for AI.
• Training dataTraining data is the collection of examples an AI model learns from, such as text, images or labelled records.
• Training runA training run is one job of training a model for a chosen number of steps, including restarts after a crash.
• Transfer learningTransfer learning is training a model on one task and then reusing what it learned on a different task or a different data distribution.
• TransformerA transformer is the neural network design behind almost every modern AI chatbot.
• Transparency in AITransparency in AI means being open about when AI is being used, what it can and cannot do, how it was built and how it reached a result, so people can make informed choices and challenge decisions that affect them.
• Turing testThe Turing test is Alan Turing's imitation game: a person tries to tell a machine from a human using written conversation.

U

• UK GDPRThe UK GDPR is the UK's main data protection law.
• Unit economicsUnit economics is the money a business makes or loses on each single unit it sells, such as one customer, one subscription or one AI query.
• Unsupervised learningUnsupervised learning is training an AI model on data that has no labels saying what the right answer is.
• UpscalingUpscaling is enlarging an image or video to a higher resolution while trying to keep it clear.
• Usage-based pricingUsage-based pricing is when you pay for how much of a service you actually use, rather than a flat fee.

V

• VectorA vector is an ordered list of numbers.
• Vector databaseA vector database is a database built to store embeddings, the lists of numbers AI models use to capture meaning, and to quickly find the stored items whose numbers sit closest to a query.
• Vibe codingVibe coding is building software by telling an AI tool what you want in ordinary language and letting it write the code.
• Video extensionVideo extension is an AI editing feature that generates new frames after an existing clip so the action continues for longer.
• Vision-language modelA vision-language model (VLM) is an AI model that takes in both images and text, so it can describe a photo, read the text in it, or answer questions about what it shows.
• Voice cloningVoice cloning is generating speech that resembles a particular person's voice from a recording of them.
• Voice modeVoice mode is ChatGPT's spoken conversation feature.

W

• WatermarkingWatermarking is a way of hiding a mark inside an image, a sound, a video or a piece of text so a later check can tell an AI system produced it.
• Web scraping for AIWeb scraping for AI is the automated collection of online pages and files so they can be used in an AI dataset.
• WeightsWeights are the learned numbers inside an AI model that decide how strongly each input affects the output.
• Workflow automationWorkflow automation is the computerised facilitation or automation of a business process, in whole or in part, by passing documents, information or tasks between participants according to set rules.
• World modelA world model is an AI system that learns how an environment behaves, so it can predict what happens next, including how the scene changes when someone acts in it.