AI Glossary

The relationship between an act or event and the resulting harm.

Concerns that advanced models may facilitate access to information usable for chemical, biological, radiological and nuclear threats; one of the categories of systemic risk.

A marking indicating a product's conformity with EU requirements; high-risk AI systems must bear it after a successful conformity assessment.

Protection against the unauthorized use of a well-known person's likeness, voice or other characteristics in AI-generated content.

Technical standards being prepared by the European standardization organizations, compliance with which establishes a presumption of conformity of high-risk systems with the requirements of the AI Act.

A procedure in which an independent certification body first verifies readiness, then carries out a detailed audit, and in subsequent years confirms through repeated checks that the management system continues to function.

A technique in which the model explicitly sets out the individual steps of its reasoning before answering, thereby improving reliability on more complex tasks.

OpenAI's conversational assistant built on the GPT family of models, which popularized generative AI among the general public.

The body of legislation of the People's Republic of China governing the provision of generative AI services, including requirements on content control, algorithm registration and the labeling of synthetic content.

Anthropic's family of models and assistant, developed with an emphasis on safety and the Constitutional AI approach.

A model whose weights and internal workings are not publicly available and to which access is generally provided only through an interface or service.

A voluntary set of rules and recommendations that helps meet the requirements of the AI Act.

A provider of a GPAI model that has voluntarily committed to adhering to the code. Signatories are publicly listed and can demonstrate compliance with the AI Act with a lower administrative burden.

An AI agent specializing in writing, editing and debugging program code, capable of independently carrying out tasks in a development environment.

A license permitting the commercial exploitation of an AI model or service under specified conditions, as opposed to licenses limited to research or non-commercial purposes.

The commercial use of content created by artificial intelligence, raising questions of copyright, licensing and liability for potential infringement of third-party rights.

A state in which an AI system meets all the requirements of the AI Act and related legal regulations.

System design in which legal and security requirements, defined purposes of use, risk management and ongoing auditability are built in from the development stage onwards, so that the agent remains compliant with regulation.

The risk that the use of AI will breach legislation, internal rules or contractual obligations, with potential penalties and reputational consequences.

The total computational power required to train and operate models; it is becoming a strategic resource as well as a criterion for regulating the most powerful systems.

The ability of an AI agent to operate a computer or browser user interface (clicking, typing, reading the screen) and thereby carry out tasks on the user's behalf.

A dispute brought by music publishers against Anthropic concerning the reproduction of song lyrics in the model's outputs. It illustrates the spread of copyright disputes from text and images into the field of music.

A numerical value expressing how confident a model is in its output; on its own it does not guarantee correctness and may be misleading.

The process of verifying whether an AI system meets the requirements set out by the AI Act.

A training approach in which the model's behavior is guided by an explicit set of principles (a kind of constitution) instead of relying solely on human evaluation of every response.

Technical standards that make it possible to track the origin and editing history of digital content; the C2PA standard is one of the main tools for labeling provenance.

A set of measures preventing a model from generating harmful, unlawful or otherwise inappropriate content, and filtering risky inputs and outputs.

The deliberate design and assembly of the information a model receives when working on a task, so that its output is as accurate and relevant as possible.

The amount of information an AI model can take into account at one time when generating a response.

Techniques that extend the length of text a model can process at once, thereby enabling work with extensive documents.