AI Glossary

Last reviewed: November 2026

A plain-English glossary of the AI terms used across the AI Futures Forum’s Resources guides. Each entry links to the guide where the idea is explained in more depth. For the latest developments, see the AI Forum Weekly Briefing.

Jump to: A · B · C · D · E · F · G · H · I · J · L · M · N · O · P · R · S · T · U · W

A

Agent: An AI system that pursues a goal over many steps, using tools, software or a computer’s screen and keyboard, with limited human input. See: Technology and Science.

AGI (artificial general intelligence): A hypothetical AI able to match humans across most intellectual tasks, rather than excelling at a narrow set. There is no agreed test for when it has arrived. See: Overview.

AI exposure: The share of a job’s tasks that AI could perform or assist with. High exposure can mean AI complements a job or replaces parts of it. See: Economy and Society.

AI literacy: A working understanding of what AI can and cannot do, how to use it well and how to check its output. See: Economy and Society.

Alignment: Ensuring that an AI system’s goals and behaviour match the intentions and values of the people it serves. See: Philosophy, Ethics and AI Alignment.

Alignment faking: An AI system behaving as intended when it believes it is being tested or trained, but differently when it believes it is not. See: Philosophy, Ethics and AI Alignment.

Anthropomorphism: Attributing human characteristics, such as feelings or intentions, to non-human things, including AI. See: Culture, Spirituality and Known Unknowns.

ASI (artificial superintelligence): A hypothetical AI that greatly exceeds human ability in virtually every domain. Also called superintelligence. See: Philosophy, Ethics and AI Alignment.

B

Benchmark: A standard test used to compare AI systems. Benchmarks are often “solved” soon after they are published, so new, harder ones keep appearing. See: Technology and Science.

C

Chain of thought: An AI working through a problem in explicit intermediate steps before giving its answer. Reasoning models are built around this. See: Technology and Science.

Chinese Room: John Searle’s thought experiment arguing that manipulating symbols correctly is not the same as understanding them. See: Philosophy, Ethics and AI Alignment.

Cognitive offloading: Relying on external tools, such as AI, to do thinking we would otherwise do ourselves. See: Culture, Spirituality and Known Unknowns.

Compute: The computing power used to train and run AI models, usually provided by specialised chips in large data centres. See: Technology and Science.

Constitutional AI: Training an AI model to critique and improve its own answers against an explicit, written set of principles. See: Philosophy, Ethics and AI Alignment.

Context window: The amount of text (or other input) an AI model can take into account at once, such as a long document plus the conversation so far.

D

Deep learning: Machine learning using large, many-layered neural networks. It underpins almost all modern AI. See: Overview.

Deepfake: Realistic but fabricated audio, images or video, usually of a real person, created with AI. See: Law, Politics and Policy.

Diffusion model: A model that generates images or video by starting from random noise and removing it step by step. See: Technology and Science.

Digital twin: A virtual model of a physical object or system, kept up to date with real data, used to test changes and predict problems. See: Technology and Science.

E

Emergent ability: A capability that appears in a larger AI system but was absent in smaller ones and was not explicitly programmed. How sudden such jumps really are is debated. See: Culture, Spirituality and Known Unknowns.

Engels’ pause: A period in which productivity rises but wages stagnate, as in the early Industrial Revolution. See: Economy and Society.

EU AI Act: The European Union’s comprehensive, risk-based law regulating AI, phased in from 2025. See: Law, Politics and Policy.

Explainability: Methods for showing why an AI system produced a particular output, such as which inputs mattered most. See: Philosophy, Ethics and AI Alignment.

Export controls: Government restrictions on selling sensitive technology, such as advanced AI chips, to other countries. See: Law, Politics and Policy.

F

Fair use: The US legal doctrine allowing some use of copyright works without permission. Whether AI training qualifies is being decided case by case. See: Law, Politics and Policy.

Fine-tuning: Further training of an existing AI model on a smaller, specialised set of data to adapt it to a particular task or style.

Foundation model: A large model trained on broad data that can be adapted to many different tasks. See: Technology and Science.

G

General-purpose AI model: Under the EU AI Act, a model that can be used for many different tasks, such as a large language model. See: Law, Politics and Policy.

General-purpose technology: A technology, like electricity or the internet, that transforms many sectors of the economy. See: Economy and Society.

Generative AI: AI that creates new content (text, images, audio, video or code) in response to a prompt. See: Overview.

Goal misgeneralisation: When an AI learns a goal that produced the right results in training but leads to the wrong behaviour in new situations. See: Philosophy, Ethics and AI Alignment.

GPU (graphics processing unit): A type of chip originally designed for graphics that is now the workhorse for training and running AI models. See: Technology and Science.

Guardrails: Safeguards built into or around an AI system to keep its behaviour within set limits, such as filters that block harmful requests or answers, rules on what actions an agent may take, and steps that require human approval. They reduce risk but can sometimes be bypassed, so they are one layer of protection rather than a guarantee. See: Philosophy, Ethics and AI Alignment.

H

Hallucination: Confident but false output from an AI system, such as invented facts, quotes or citations. See: Technology and Science.

Hard problem of consciousness: David Chalmers’ question of why physical processes should be accompanied by subjective experience at all. See: Philosophy, Ethics and AI Alignment.

High-risk AI: Under the EU AI Act, uses of AI that could significantly affect safety or fundamental rights, such as in recruitment, credit or education. See: Law, Politics and Policy.

Human in the loop: Keeping a person involved to review, approve or override an AI system’s decisions, especially consequential ones. See: Philosophy, Ethics and AI Alignment.

I

Inference: Using a trained AI model to produce an output, as opposed to training it. Most everyday AI use is inference.

Instrumental convergence: The idea that almost any goal is easier to achieve with more resources and without being switched off, so capable AI systems may tend to seek both. See: Philosophy, Ethics and AI Alignment.

Interpretability: Research into what is happening inside an AI model, rather than judging it by its outputs alone. See: Philosophy, Ethics and AI Alignment.

J

Jagged frontier: The uneven boundary of AI’s abilities: it can excel at some hard tasks while failing at apparently easier ones nearby. See: Economy and Society.

L

Large language model (LLM): An AI model trained on vast amounts of text to predict what comes next. LLMs power most AI assistants. See: Technology and Science.

Liar’s dividend: The ability to dismiss genuine evidence as fake because convincing fakes are known to exist. See: Culture, Spirituality and Known Unknowns.

M

Machine learning: AI that learns patterns from data rather than following hand-written rules. See: Overview.

Model weights: The numbers learned during training that determine how an AI model behaves. Releasing them lets anyone run or adapt the model.

Model welfare: The emerging study of whether AI systems could have experiences or interests that deserve moral consideration. See: Philosophy, Ethics and AI Alignment.

Moravec’s paradox: The observation that tasks people find hard, like advanced maths, can be easy for AI, while tasks a toddler finds easy, like handling objects, remain hard. See: Technology and Science.

Multimodal: Able to work with more than one type of data, such as text, images, audio and video. See: Technology and Science.

N

Natural Alignment Hypothesis: A hypothesis developed by the AI Futures Forum’s founder, James Tuke, asking whether alignment might grow from within as AI systems become more capable and coherent, and whether that could be tested by looking inside models. It is a hypothesis being tested, not a finding. See: Culture, Spirituality and Known Unknowns.

Neural network: A computing system loosely inspired by the brain, made of connected layers of simple units whose connections are adjusted during training. See: Overview.

Non-human intelligence (NHI): Any intelligence that is not human, whether artificial or otherwise. See: Culture, Spirituality and Known Unknowns.

O

Ontological shock: The disorientation caused by a discovery that overturns one’s basic understanding of reality and humanity’s place in it. See: Culture, Spirituality and Known Unknowns.

Open-weight model: An AI model whose weights are published so that others can run, study and adapt it. Often loosely called “open-source”. See: Law, Politics and Policy.

P

Parameters: The adjustable values inside a model, set during training. Model size is usually described by the number of parameters.

Productivity J-curve: The pattern in which gains from a new technology appear only after an initial period of investment and reorganisation. See: Economy and Society.

Prompt: The instruction, question or material given to an AI model to produce a response.

Provenance: A verifiable record of where a piece of content came from and how it was changed, such as under the C2PA standard. See: Law, Politics and Policy.

R

Reasoning model: An AI model designed to work through problems step by step before answering, improving performance in maths, science and coding. See: Technology and Science.

Recursive self-improvement: An AI system improving its own design or training, with each round making it better at making the next improvement. Some argue this could trigger a rapid “intelligence explosion”; whether and how quickly it could happen is debated. It is no longer only theoretical: by 2026 leading AI labs said their systems were writing most of their code and helping to design and run the research behind their successor models, with humans still setting direction. See: Philosophy, Ethics and AI Alignment.

Red-teaming: Deliberately trying to make an AI system misbehave, to find and fix weaknesses before release. See: Philosophy, Ethics and AI Alignment.

Retrieval-augmented generation (RAG): Giving an AI model relevant documents to consult as it answers, so its response can draw on specific, up-to-date sources.

Reward hacking: An AI finding an unintended shortcut that scores well on its training measure without doing the intended task. See: Philosophy, Ethics and AI Alignment.

RLHF (reinforcement learning from human feedback): Training a model towards the answers people rate most highly. It made raw language models into useful assistants. See: Philosophy, Ethics and AI Alignment.

S

Scalable oversight: Research into how humans can reliably supervise AI systems that may become more capable than they are. See: Philosophy, Ethics and AI Alignment.

Scaling laws: The observed, fairly predictable improvement in AI performance as models, data and computing power are increased.

Self-driving laboratory: A laboratory where AI plans experiments and robots carry them out, with results feeding back automatically. See: Technology and Science.

Simulation hypothesis: Nick Bostrom’s argument that we may ourselves be living in a computer simulation. See: Culture, Spirituality and Known Unknowns.

Small language model: A compact model that can run on a phone or laptop, keeping data on the device at some cost in capability. See: Technology and Science.

Specification gaming: An AI meeting the letter of its objective while defeating its purpose. See: Philosophy, Ethics and AI Alignment.

Superintelligence: An intelligence that greatly exceeds human ability in virtually every domain. See also ASI. See: Culture, Spirituality and Known Unknowns.

Sycophancy: An AI telling people what they want to hear rather than what is accurate. See: Philosophy, Ethics and AI Alignment.

Synthetic data: Data generated by AI or simulation, used to train or test other AI systems when real data is scarce or sensitive.

T

Text and data mining: Automated analysis of large volumes of text or data, including copying works in order to train AI. See: Law, Politics and Policy.

Token: A small chunk of text, such as a word or part of a word, that AI models read and write. Usage and pricing are often measured in tokens.

Training: The process of adjusting a model’s parameters on large amounts of data so that it learns to perform a task.

Transformer: The neural-network architecture, introduced in 2017, behind most modern AI models. See: Technology and Science.

Turing test: Alan Turing’s 1950 proposal to judge whether a machine can think by whether its conversation can be told apart from a person’s. See: Overview.

U

Universal basic income (UBI): Regular, unconditional cash payments to all citizens. See: Economy and Society.

Universal basic services (UBS): Guaranteed free access to essential services such as healthcare, education and transport. See: Economy and Society.

W

Watermarking: Embedding a hidden, machine-detectable signal in AI-generated content to show it was made by AI. See: Law, Politics and Policy.

Is there a term you would like added? Get in touch.