AI Glossary

A threshold of computing power used for training beyond which a GPAI model is presumed to pose systemic risk and becomes subject to the strictest obligations.

An independent review of an AI system in terms of regulatory compliance, data quality, risks and impacts on the persons concerned.

The use of AI for diagnostics, decision support or the processing of health data; often falling under both the medical devices regime and that for high-risk systems.

A continuously maintained overview of all AI systems used in an organization, enabling oversight to be managed according to the risk level of individual systems.

A continuously maintained inventory of all AI systems used within an organization, enabling oversight to be managed according to the level of risk.

The ability to continuously monitor the behavior of a deployed system (inputs, outputs, errors, costs) for the purposes of debugging, auditing, and risk management.

The contractual and statutory definition of which entity bears which obligations and risks; particularly crucial for high-risk systems and the integration of third-party models.

A reference to EU harmonization legislation; AI systems embedded in regulated products (e.g., medical devices, machinery) are classified as high-risk under it, with a longer transition period.

A list of areas in which standalone AI systems are considered high-risk, such as recruitment, credit scoring, education, law enforcement, migration and border management, and critical infrastructure.

The company developing the Claude family of models, with an emphasis on safety and the Constitutional AI approach; a party to the Bartz case and its settlement, and a participant in discussions about GPAI models with systemic risk.

The provision laying down the specific obligations of providers of GPAI models with systemic risk, in particular risk assessment and mitigation, incident reporting and ensuring security.

The provision setting out the rules under which an AI system is considered high-risk, including exemptions for systems that do not pose a significant risk of harm.

The distinction between evaluating what a model can do (its capabilities) and evaluating whether its use is safe; both form part of managing systemic risks.

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.

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.

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.

A structured set of controls from which an organization selects those that match the risks of a particular AI system and the requirements of several frameworks at once.

A step taken to eliminate an identified non-compliance or risk associated with an AI system.

Fixed deadlines that the Czech draft sets for the processes of authorizing the operation and testing of high-risk AI systems in real-world conditions.

A change from the Digital Omnibus, under which the obligations for standalone high-risk systems under Annex III are deferred to 2 December 2027, and for systems embedded in regulated products to 2 August 2028.

A natural or legal person that uses an AI system in the course of its activities without being its provider; the AI Act imposes specific obligations on it, particularly for high-risk systems.

A data protection impact assessment that a controller must carry out for processing that poses a high risk to the rights and freedoms of individuals, often relevant to AI systems.

A document by which the provider declares that a high-risk AI system meets the requirements of the AI Act and other relevant regulations.

The hypothetical risk that very advanced AI systems could cause catastrophic impacts on humanity; a subject of both expert and political debate.

A unit measuring the amount of computation. The total number of FLOP used during training serves as one of the criteria for classifying models with high impact or systemic risk.

A rule under which certain obligations of the code do not apply to free and open models, unless they are classified as models with systemic risk.

The most advanced and most powerful current models, which push the boundaries of AI capabilities and are subject to special attention with regard to safety and systemic risks.

A model capable of performing a wide range of tasks without being limited to a specific field or use, and of being integrated into various AI systems; referred to in the AI Act as a GPAI model.

We break down the obligations of GPAI model developers and the systemic-risk regime in our article AI Compliance – Do you develop or use artificial intelligence?

A general-purpose model whose capabilities or impact reach a level at which they may pose a risk to society as a whole. It is subject to the obligations under Article 55 of the AI Act.