AI Glossary

A dataset set aside to verify a model's performance and reliability, which is not used during its training so that the evaluation is objective.

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

The ability to understand and clearly explain how an AI system arrived at a result or decision.

An approach that makes it possible to explain the functioning and results of an AI system so that a human can understand them.

A feature of a legal framework whereby it applies also to entities outside a given territory if their systems affect persons or the market in that jurisdiction; typical of the European AI Act.

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.

A set of voluntary principles of conduct for organizations developing advanced AI systems, agreed within the cooperation of the G7 countries.

Google DeepMind's family of multimodal models, integrated into Google services and developer tools.

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?

An AI system based on a general-purpose model that can be used for a wide range of applications and sectors without being limited to a single specific purpose.

An architecture with two competing networks (a generator and a discriminator), one of which creates content and the other evaluates it; historically used, among other things, for deepfakes.

Artificial intelligence that creates new content, such as text, images, audio or video, based on learned patterns.

A lawsuit filed in 2023 in which Getty Images accuses Stability AI of the unauthorized use of more than 12 million photographs, their captions, and metadata to build the Stable Diffusion and DreamStudio image generators.

A tool built into development environments that suggests and completes program code based on context; it has been the subject of copyright disputes over its training code.

A document adopted at the United Nations in 2024 that sets out shared goals and principles for digital cooperation among states, including a commitment to govern emerging technologies and AI for the benefit of humanity.

The goal of having individual national and international AI frameworks complement and recognize one another, thereby reducing the burden on organizations operating across multiple jurisdictions.