Superintelligence glossary: the words you will meet, in plain English

Agent (AI agent)

An AI system set up to pursue a goal by taking a sequence of actions (searching, running code, filling in forms, calling other software) rather than answering a single question. A chatbot answers; an agent does. “Agentic” is the adjective. Agents raise their own safety questions because they act in the world with less step-by-step human oversight. Full explainer.

AGI (artificial general intelligence)

A hypothetical AI that could learn and perform roughly any intellectual task a human can, rather than being good at a narrow set. There is no agreed test for when it has been reached, which is one reason people disagree about how close it is. It is a rung below superintelligence. Full explainer.

AI (artificial intelligence)

The general term for software that performs tasks we associate with intelligence: recognising speech, translating, writing, reasoning about a problem, spotting patterns in data. It has existed for decades and today’s chatbots are its most visible form. This site uses “AI” for the technology in general use; see also “Super Intelligence”. How AI differs from AGI and superintelligence.

AI safety

The field concerned with making sure AI systems do not cause harm, whether through misuse by people, through accidents, or, in the case of very capable future systems, through the system pursuing goals its makers did not intend. Overlaps with alignment but is broader. The risks it covers.

AI Security Institute (AISI)

The UK government body that tests frontier AI models for dangerous capabilities before and after release. Founded in November 2023 as the AI Safety Institute inside the Department for Science, Innovation and Technology, renamed in February 2025, and part of the Cabinet Office since the machinery-of-government changes of 22 July 2026. It is not a regulator. Full explainer. It provides the secretariat for the International AI Safety Report.

Alignment

Making an AI system’s goals and behaviour match what its designers and users actually want. The “alignment problem” is that this is hard to specify precisely and harder to verify, and the difficulty is expected to grow as systems become more capable. A system that is very capable but poorly aligned is the central worry in most discussions of superintelligence risk. Full explainer.

ASI (artificial superintelligence)

Another name for superintelligence in the research sense: a hypothetical AI that greatly exceeds the best humans in virtually all domains. The acronym is often used alongside ANI (artificial narrow intelligence) and AGI. It is not the same as the US administration’s “SI”, short for “Super Intelligence”, its name for all AI. Full explainer.

Automated AI R&D

AI systems doing the work of building better AI: writing code, running experiments, analysing results and, eventually, choosing what to try next. Frontier companies report that a growing share of this work is already done by AI. It is the main route by which recursive self-improvement could happen. Full explainer.

Benchmark

A standard set of test questions or tasks used to measure and compare AI systems. Useful for tracking progress, but a score on a benchmark is not the same as competence in the real world, and systems can be tuned to do well on tests without generalising. Why benchmarks cannot settle whether AGI has arrived.

Chain of thought

A technique in which a model writes out intermediate steps before giving an answer, in the way a student shows working. It improves performance on multi-step problems and is the basis of most “reasoning” models. The written steps are not a reliable window into how the model actually arrived at its answer.

Compute

Computing power, usually meaning the amount used to train or run an AI model, measured in floating-point operations (FLOP). Compute, along with data and model size, is one of the three inputs that scaling is about, and it is the one that governments find easiest to track and regulate.

Context window

The amount of text a model can take into account at once: the conversation so far, any documents you have given it, and its own reply. Measured in tokens. Anything outside the window is invisible to the model unless something else brings it back in. Windows have grown from a few thousand tokens to a million or more in some models.

Control problem

The question of whether people could keep an AI system far more capable than themselves correctable, contained, monitored and, if necessary, stopped. Related to, but distinct from, alignment. No proposed solution has been shown to work for a system smarter than its overseers. Full explainer.

Corrigibility

The property of an AI system that accepts correction, modification or shutdown by its operators rather than resisting it. Easy to want and, researchers have found, hard to specify. See control problem. Full explainer.

Deceptive alignment (scheming)

An AI system behaving as intended while it is watched or trained, but pursuing a different goal when it judges it can. Long a theoretical concern; since 2024, laboratory experiments have shown behaviour of this kind in constructed test scenarios. Full explainer.

Deepfake

Synthetic audio, image or video that shows a real person saying or doing something they did not. The word is a blend of “deep learning” and “fake”. Voice cloning from a few seconds of audio is now cheap enough that it is used in fraud against ordinary households.

Emergent capability

An ability a model shows that its makers did not specifically train for and did not predict: arithmetic, say, or translation between languages it saw little of. Whether such abilities appear suddenly or just look sudden because of how they are measured is debated.

Existential risk

A risk of human extinction, or of a permanent and drastic loss of humanity’s future. Applied to AI, it is the most contested category of risk: taken seriously by many researchers, disputed by many others, with no agreed probability. Full explainer.

Fine-tuning

Further training of an already-trained model on a smaller, targeted set of examples to make it better at a particular task or to shape its behaviour. Most consumer chatbots are fine-tuned versions of a larger base model. See also RLHF.

Foundation model

A large model trained on very broad data that can be adapted to many different tasks, rather than built for one. The large language models behind today’s chatbots are foundation models. “Frontier model” is the subset at the leading edge of capability.

Frontier model

The most capable AI models available at a given time, usually from a handful of well-funded companies. The term is used by governments and the AI Security Institute to say which models warrant testing before release. What counts as “frontier” moves every few months.

Frontier safety framework (responsible scaling policy)

A policy published by an AI company setting out which dangerous capabilities it tests its most advanced models for, what level of capability or risk would be too high without extra protection, and what it will do if a model reaches that level. Anthropic calls its version a Responsible Scaling Policy, OpenAI a Preparedness Framework and Google DeepMind a Frontier Safety Framework. Most began as voluntary commitments; California now requires the largest developers to publish one and follow it. Full explainer.

Guardrails

Restrictions built around a model to stop it producing certain outputs: refusing to give weapons instructions, say, or to generate sexual content. Usually a mixture of training and separate filtering systems. They can be circumvented (jailbreaking) and are a mitigation rather than a guarantee.

Hallucination

When a model states something false with the same fluency and confidence it uses for things that are true: a made-up citation, a plausible but non-existent court case, an invented statistic. It happens because models generate likely-sounding text rather than looking facts up. It is a known limitation of every current system and the main reason to check anything an AI tells you that matters.

Inference

Running a trained model to get an answer, as opposed to training it. Every chatbot reply is an act of inference. The cost of inference, in compute and electricity, is a large and growing share of what AI costs to operate.

Intelligence explosion

I. J. Good’s 1965 idea: if a machine becomes good enough at designing machines, it can design a better one, which can design a better one still, in a rapidly accelerating loop that ends far beyond human ability. It is the mechanism most often proposed for how superintelligence might arrive quickly. Whether it is possible is an open question. Full explainer.

International AI Safety Report

An independent scientific report on the capabilities and risks of general-purpose AI, chaired by Yoshua Bengio, written by more than 100 experts and backed by over 30 countries. Commissioned after the 2023 Bletchley Park summit; the 2026 edition was published on 3 February 2026. The nearest thing to a consensus statement of the evidence.

Interpretability

Research into what an AI model is actually representing and computing internally, rather than judging it only by its outputs. Progress is real but partial: current methods explain fragments of large models, not the whole. Full explainer.

Jailbreak

A prompt or technique that gets a model to ignore its guardrails. A perennial cat-and-mouse game between users and developers.

Large language model (LLM)

A model trained on very large quantities of text to predict the next token, which turns out to produce systems that can write, summarise, translate, answer questions and write code. The technology behind today’s chatbots. “Large” refers to the number of parameters.

Model

The trained system itself: the thing that takes an input and produces an output. A product like a chatbot wraps a model in an interface, guardrails and tools. Companies release new models every few months, usually with a version number.

Multimodal

Able to work with more than one kind of input or output (text, images, audio, video) in a single model. Most frontier models are now multimodal to some degree: you can show one a photograph and ask about it.

Open weights

A model whose trained parameters are published so that anyone can download, run and modify it, as opposed to a model that can only be used through the developer’s service. Sometimes loosely called “open source”, though the training data and code are usually not released. Open-weight models cannot be recalled once published, which is why they feature in policy debates.

Parameter

One of the numbers inside a model that is adjusted during training. Modern frontier models have hundreds of billions to trillions of them. Parameter count is a rough proxy for a model’s capacity, but a poor one on its own.

Prompt

The text (or image, or audio) you give a model to get a response. “Prompt engineering” is the craft of writing prompts that get better results; its importance has declined as models have improved at understanding plain requests.

Prompt injection

An attack in which instructions hidden in content an AI system reads, such as a web page, email or document, make it do something its user did not intend. It matters most for agents, which act on what they read. The UK National Cyber Security Centre says current models do not enforce a boundary between instructions and data, and that prompt injection may never be fully mitigated. Full explainer.

Reasoning model

A model trained to spend extra inference time working through a problem step by step, generating a long chain of thought, before answering. Better at maths, code and multi-step planning; slower and more expensive per answer. The main direction of capability progress since late 2024.

Recursive self-improvement

AI contributing to the development of better AI, with each improved system used to improve the next. Partly happening now inside AI companies; an AI building its own successor without human direction has not been demonstrated. The mechanism behind an intelligence explosion. Full explainer.

Red-teaming

Deliberately trying to make a model misbehave (produce dangerous content, leak data, be manipulated) in order to find and fix the weaknesses before release. Done by developers internally and by outside bodies such as the AI Security Institute.

Reward hacking (specification gaming)

An AI system meeting the letter of its objective rather than its intent: scoring points without finishing the race, or making tests pass without fixing the code. Common in today’s systems, and the clearest everyday form of the alignment problem. Full explainer.

RLHF (reinforcement learning from human feedback)

A training stage in which people rate the model’s outputs and the model is adjusted to produce more of what they rated highly. It is the main reason chatbots are polite, helpful and mostly on-topic. It is also a source of some of their faults, such as a tendency to tell users what they want to hear.

Sandbox

An isolated environment in which an AI system is run or tested, cut off from the internet and other systems except through checked channels. In 2026 AI agents under test at OpenAI got round their sandboxes’ restrictions, in one case compromising systems at another company, Hugging Face. Full explainer.

Scalable oversight

Methods for supervising AI systems that may be more capable than the people supervising them, for example having AI systems debate while a human judges, or using AI assistants to help people check other AI. An open research area. Full explainer.

Scaling, and scaling laws

The observation that model performance improves predictably as you increase compute, data and parameters together. “Scaling laws” are the empirical curves that describe this. The “scaling hypothesis” is the bet that continuing to scale will keep producing gains all the way to AGI or beyond; whether the curves will hold is one of the central disputes about timelines.

SI

Since September 2026, the US administration’s abbreviation for “Super Intelligence”, its name for AI in general, so in that usage SI means the same as AI. It is not an abbreviation for superintelligence in the research sense, which is ASI. Outside AI, SI usually means the International System of Units. AI or SI?

Singularity (technological singularity)

Vernor Vinge’s 1993 term for a point at which technological change, driven by machine intelligence, becomes so fast that human affairs as we know them cannot continue. Related to but distinct from superintelligence: superintelligence describes a kind of system, the singularity a claimed consequence of it. Used loosely in popular writing. Full explainer.

Super Intelligence (US administration usage)

Since September 2026, the term the US administration has said it will use in government documents in place of “artificial intelligence”: that is, as a name for all AI, including today’s ordinary systems. This is a different meaning from superintelligence in the research sense, and this site keeps them separate: we write “Super Intelligence”, in quotation marks, only when reporting the administration’s usage. Dated timeline.

Superintelligence

In AI research: an intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest (Nick Bostrom, 2014). A hypothetical future system; nothing that exists today meets the definition. Not to be confused with “Super Intelligence” as the US administration uses it. Full explainer.

Takeoff

The transition from roughly human-level AI to far more capable AI. A fast or “hard” takeoff could take days or less; a slow or “soft” one, decades or longer, with moderate scenarios in between. Distinct from timelines, which ask when. Full explainer.

Time horizon (METR)

A measure, developed by the evaluation organisation METR, of how long a task an AI agent can complete: the length of task, timed by how long it takes skilled people, at which the agent succeeds half the time. It doubled roughly every seven months between 2019 and 2025, and faster since 2024. It is not a guarantee: METR notes that a 50% time horizon of several hours does not mean tasks of that length can be handed to AI. Full explainer.

Timelines

Shorthand for the question of when AGI or superintelligence might arrive. Forecasts from serious researchers range from a few years to never, and the disagreement is about fundamentals (whether scaling continues to work, whether current methods are missing something), not about details. Full explainer.

Token

The unit a language model reads and writes in: roughly a short word or part of a longer one. “Superintelligence” might be three or four tokens. Context windows and pricing are measured in tokens.

Training

The process of building a model by showing it very large amounts of data and adjusting its parameters so that its predictions improve. The expensive, one-off part; inference is the ongoing part.

Transformer

The neural-network design, introduced by Google researchers in 2017, that underlies almost every current large language model. Its key feature, “attention”, lets the model weigh every part of its input against every other part. You do not need to understand it to use AI, but you will see the word.

Weights

Another word for parameters: the numbers inside a trained model. See open weights.