AI Glossary

The effort to ensure that AI systems do not create unjustified differences between groups of people; it encompasses various, sometimes mutually incompatible, definitions of fairness.

The requirement that it be intelligible how, and on what basis, an algorithm reaches its decisions, particularly where those decisions affect people's rights.

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 2025 settlement of a class action brought by authors against Anthropic, considered the largest known copyright settlement in the United States to date. It concerned claims related to the downloading of pirated copies of works used for training.

The provision defining prohibited AI practices, such as subliminal manipulation, disproportionate social scoring, certain forms of biometric categorization and the untargeted scraping of facial images. Applicable from 2 February 2025.

The provision imposing transparency obligations, such as informing people that they are interacting with AI and labeling artificially generated or manipulated content. Applicable from August 2026.

The provision setting out the obligations of GPAI model providers regarding transparency and copyright compliance, which the code of practice serves to implement.

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.

Regulation (EU) 2024/1689, the world's first comprehensive legal framework for AI, based on a risk-tiered approach.

We break down the risk categories, developer obligations and penalties in our article AI Compliance – Do you develop or use artificial intelligence?

A comprehensive set of policies, processes, and roles through which an organization plans, operates, monitors, and improves its use of AI; analogous to management systems under standards for quality or information security.

A machine-based system designed to process inputs with varying levels of autonomy and generate outputs such as predictions, recommendations, decisions or content that can influence physical or virtual environments.

A person established in the EU whom a provider from a third country appoints to fulfill the obligations under the AI Act and who serves as a contact point for supervisory authorities.

A lawsuit brought by an association of authors; one of the leading cases representing writers' interests against the use of their works to train language models.

The tendency of people to overvalue and uncritically accept the outputs of an automated system, which undermines the effectiveness of human oversight.

AI systems that drive vehicles or assist with driving, regulated both by the AI Act and by product safety and road traffic legislation.

A new prohibition targeting AI systems that create or alter intimate imagery without the consent of the depicted person (so-called nudifiers), included among prohibited practices.

A new prohibition being inserted into Article 5 of the AI Act which, as of 2 December 2026, prohibits placing on the market systems designed to generate child sexual abuse material.

A measure by a supervisory authority prohibiting the further use of an AI system that violates legal requirements.

A dispute brought by authors (A. Bartz, Ch. Graeber, K. W. Johnson) against Anthropic. In June 2025, Judge William Alsup held that training a model on books is itself exceedingly transformative and constitutes fair use, but that creating and retaining a permanent library of pirated works does not.

The gradual change in the behavior of a deployed AI system during use, where it is difficult to distinguish expected adaptation from a substantial modification that would trigger a new conformity assessment.

Distortion of evaluation results caused by test tasks inadvertently ending up in the training data, so that the model achieves artificially better scores.

A systematic deviation in the data or model that may lead to unfair or discriminatory outcomes.

Assigning people to groups on the basis of biometric data; some forms (e.g. inferring sensitive characteristics) are prohibited or strictly limited by the AI Act.

The difficulty that the internal decision-making processes of complex models are very hard for humans to explain, which complicates both oversight and the proving of liability.

A party's obligation to prove the facts it asserts; in the case of harm caused by AI, its allocation is crucial, because the internal workings of the system are often difficult to prove.

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.