AIGP key terms

A field of computing concerned with systems that carry out tasks associated with human intelligence, such as recognising patterns, reasoning, translating languages, and generating text or images.

A branch of AI in which a system learns patterns from examples rather than following only hand-written rules, so its performance on a task improves with exposure to data.

The examples used to teach a model statistical patterns. Its quality, representativeness, licensing and provenance strongly shape how the resulting system behaves and which risks it carries.

A large model trained on broad, general data that can be adapted to many downstream tasks through fine-tuning or prompting instead of being built for a single purpose.

A foundation model trained on very large bodies of text to predict and produce language, enabling tasks such as summarising, translating, answering questions and writing code.

AI that creates new content, such as text, images, audio, video or code, by reproducing patterns learned from training data rather than only labelling or scoring existing items.

An output in which a generative model states something false or unsupported as if it were true, because it predicts plausible-sounding content instead of checking facts.

Systematic and unfair differences in how an AI system treats people or groups, often caused by unrepresentative data, poor labels or design choices, which can entrench discrimination.

The extent to which people can understand why a system produced a given output, including which inputs mattered and how confident the result is.

The laws, policies, roles, processes and controls used by an organisation or government to steer the development and use of AI toward lawful, ethical and accountable outcomes.

The principle that identified people and organisations must answer for the AI systems they build or operate, supported by clear responsibilities, records and consequences.

The principle that providers and users of AI should disclose enough about a system's purpose, data, capabilities and limits for people to understand and challenge how it is used.

A European Union law that sorts AI systems into risk tiers and attaches duties accordingly, from outright bans on prohibited practices to detailed requirements for high-risk and general-purpose AI.

A way of regulating or governing AI in which obligations grow with the chance and seriousness of harm, so riskier uses face stricter requirements and closer oversight.

An AI use that meets danger criteria set in law, such as safety components, biometric identification, critical infrastructure, hiring, education or access to essential services, and so attracts extra obligations.

A model trained on broad data that can perform many different tasks and be built into numerous downstream products; EU rules require documentation and, for the most capable, extra testing.

A voluntary United States framework organising AI risk work into govern, map, measure and manage functions, accompanied by profiles including one for generative AI.

An international standard setting requirements for an AI management system, giving organisations a certifiable structure of policy, roles, risk treatment and continual improvement for responsible AI.

Intergovernmental recommendations for trustworthy AI that call for inclusive growth, human rights protection, transparency, robustness, accountability and sustained investment in research.

Decisions about people made by a system with little or no human involvement; under the GDPR such decisions with significant effects carry special safeguards and a right to contest.

The full sequence a system passes through - problem definition, data work, building, testing, deployment, monitoring and retirement - each stage needing its own governance and controls.

The continuing process of finding, assessing, prioritising, treating and watching the risks AI poses to people, the organisation and society, so decisions rest on evidence.

A structured look at what could go wrong with a system, how likely and serious the harm would be, who is affected and which controls would reduce it.

The amount and kind of risk an organisation is willing to carry in pursuit of its goals, which decides whether a proposed AI use is approved, modified or refused.

The risk left once planned controls are in place; leaders must decide whether to accept it or ask for further mitigation before the system goes live.

A design where a person reviews, approves or can override a system's output before it takes effect, giving meaningful control where errors could seriously harm people.

Adversarial testing in which specialists deliberately try to make a system fail, behave unsafely or be misused, so weaknesses are found before harm occurs.

The gradual decline in a model's accuracy as live conditions and data move away from what it learned during training; catching it requires ongoing measurement and retraining.

Ongoing observation of a live AI system's performance, outputs and effects so that errors, drift, misuse or emerging harms are detected and fixed quickly.

A short published document describing a model's intended uses, training data, evaluation results, limits and known risks, helping downstream users judge whether it fits their purpose.

The study of what is right and fair in building and using AI, covering values such as fairness, autonomy, non-harm, dignity, privacy and sustainability.

The problem of making an AI system's behaviour match human intent and values, especially as models grow more capable and their outputs become harder to predict.

The contested possibility that highly capable AI could cause harm severe enough to threaten civilisation or human survival, a claim that shapes precautionary policy proposals.

Images, audio or video produced or altered by AI to show people saying or doing things they never did, creating risks of fraud, defamation and election manipulation.

Methods such as watermarking and signed metadata that record how a piece of media was created, helping people and platforms tell AI-generated content from authentic material.

The unresolved question of whether copying protected works to train models is lawful, and how licensing, attribution and the status of generated outputs should work.

A model whose trained parameters are released for others to download, inspect or adapt, which spreads benefits and scrutiny but removes the original developer's control over misuse.

AI that can plan and carry out sequences of actions toward a goal with limited human involvement, raising questions about control, liability and unintended side effects.

The energy, water and hardware that training and running AI consume, together with the resulting emissions and land use, now a factor in sustainability decisions.

Changes in the number, mix and quality of jobs as AI automates or augments tasks, shifting which skills are valued and who bears the transition costs.