AI Glossary

The degree to which a model's outputs correspond to reality; in generative AI it is crucial because of the risk of convincing but incorrect statements.

An informal label for the state of U.S. case law on fair use in AI training, reflecting the inconsistency of decisions to date depending on how the data was obtained and the nature of the system.

A training method in which data remains on users' devices and only model updates are shared, thereby reducing the transfer of personal data.

A situation in which AI outputs influence the future data on which the system learns, thereby further reinforcing the original bias.

A model created by fine-tuning a base model on a narrower set of data for a specific task or domain, thereby achieving better results in that area.

A process in which an already trained model is further trained on purposefully selected data so that it better handles a specific task or behaves in a desired way.

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 general-purpose AI model trained on large-scale data that can then be used or adapted for many different purposes.

A comparison of the concepts and requirements of different frameworks (AI Act, NIST, ISO, OECD) so that it is clear where they overlap and where they differ, and so that compliance can be addressed together.

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.

Voluntary commitments by leading AI developers regarding safety testing and the disclosure of risk thresholds, adopted at international AI safety summits.

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.

The ability of a model to invoke a specified function or tool and pass it structured parameters, thereby connecting the language model with external systems.

Rights protected by European Union law, such as the right to privacy, the protection of personal data, non-discrimination, or freedom of expression.

An analysis of an AI system's impact on the fundamental rights of individuals prior to its deployment.