AI Glossary

An entity that intermediates the sharing of data between those who provide it and those who use it, under the conditions laid down by EU law.

A requirement to process only the data necessary for a given purpose, which conflicts with the tendency of models to benefit from as large a volume of data as possible.

A targeted attack in which malicious or misleading samples are deliberately inserted into training data in order to disrupt the behavior of the resulting model.

The ability to move data between different service providers.

An agreement under data protection rules governing the processing of data by an AI service provider on behalf of the controller.

A requirement that data used for the development and operation of AI be relevant, representative, sufficiently accurate and, as far as possible, free of systematic bias.

The provision of data to another entity under conditions laid down by contract or by law.

A natural person whose personal data is processed.

A freely given, specific, informed, and unambiguous expression of will regarding the processing of personal data.

An entity that, by law or contract, gains access to data and uses it under conditions laid down by the Data Act.

The sui generis right of a database maker to protect content in which they have made a substantial investment; relevant when large datasets are used for training.

A structured collection of data used for training; the conditions of its use are set by its license, the breach of which may give rise to liability.

The question of whether, and to what extent, a dataset used for training is protected by copyright or by the sui generis database right, and on what conditions it may be used.

License terms determining how a dataset may be used, shared and further distributed; breaching them may give rise to liability.

The dataset equivalent of a model card; it documents the origin, composition, collection method and limitations of the data so that its suitability can be assessed.

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.

The part of a model that generates output, such as text, from an internal representation; typical of models geared towards content creation.

A subfield of machine learning that uses neural networks with many layers, capable of recognizing complex patterns in large datasets.

Targeted tailoring of content to an individual based on extensive data about their behavior, which can increase the effectiveness of manipulative practices.

Realistic-looking but artificially created or altered image, audio or video content depicting a person saying or doing something that never happened.

Advertising using a deepfake likeness or voice of a real person, often without their consent, raising questions of personality rights, misleading advertising and liability.

A Chinese model provider that attracted attention for training powerful models efficiently and sparked debate about the costs and geopolitics of AI.

A product that does not provide the safety that one is entitled to expect from it; for AI incorporated into products it is relevant to liability for damage.

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 work created by adapting or altering an existing work, the creation of which generally requires the consent of the author of the original work.

A mathematical privacy framework that, by adding controlled noise, limits the ability to infer information about an individual from outputs.

A type of generative model that creates content by progressively removing noise from a random input; it underpins many image and video generators.

A term for an AI agent deployed within an organization to perform tasks independently in a manner similar to a human employee, raising questions of liability and oversight.

An artificially created rendering of a specific person's appearance or voice; creating and using it without consent conflicts with personality rights and other rights.