AI Glossary

An EU regulation governing the conduct of large digital platforms designated as gatekeepers, with the aim of ensuring fair competition; it also touches on the integration of AI services with platforms.

We explain who the DMA applies to, what rights it gives businesses and how it differs from the DSA in our article The DMA Regulation in a Nutshell.

An artificially created digital copy of a specific person's appearance, voice or expression, the use of which without consent may interfere with personality rights.

An artificially created depiction of a specific person (image, voice) whose creation and use without consent runs up against personality rights and other protections.

An EU regulation governing the responsibility and obligations of providers of online services and platforms, including rules on content moderation and the transparency of recommender systems, which intersect with the use of AI.

You will find practical guidance in our article How to bring your terms and conditions into line with the DSA, and a comparison with the DMA in The DMA Regulation in a Nutshell.

The monitoring and evaluation of employees using AI, which touches on the protection of privacy, personal data and dignity and is particularly sensitive from the perspective of labor law.

A virtual model of a real object or process that is continuously updated with real-world data and used for simulation and optimization.

The phase of a dispute in which the parties exchange evidence. In AI disputes it has become apparent that operational records and product outputs can be requested by a court (e.g. an order to hand over millions of ChatGPT conversations).

Unjustified disadvantaging of individuals or groups as a result of AI decision-making, for example due to biased data or inappropriately set criteria.

A lawsuit by major film studios against the image generator Midjourney concerning the reproduction of protected characters and visual works in its outputs. It is among the leading disputes focused not on the inputs but on the outputs of generative AI.

A procedure in which a smaller model learns to imitate the behavior of a larger model in order to achieve similar results at a lower cost.

An entity in the supply chain that makes an AI system available on the market without being its provider or importer; the AI Act imposes certain obligations on it.

An entity that integrates a third party's GPAI model into its own system; it relies on the documentation and information provided by the original model provider.

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.

The risk that a model's capabilities usable for legitimate purposes may be misused for harmful activities, for example in the area of cyberattacks or dangerous substances.

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 author's rights to use a work and to decide on its commercial exploitation, which can be licensed or transferred.

Running models directly on end devices (phone, sensor) without the need to send data to the cloud, which can improve both privacy and responsiveness.

Interpretive guidance from the European Data Protection Board on when a model can be considered anonymous and how to assess legitimate interest during training.

The European framework for electronic identification and trust services, relevant for identity verification, electronic signatures, and trustworthy proof of origin, including in AI environments.

A platform for voice synthesis and voice cloning, associated with questions of voice rights protection and fraud using imitated voices.

A numerical representation of the meaning of a word, text or image in a multidimensional space that makes it possible to compare the similarity of content; the basis of vector search.

Capabilities or behaviors that appear unexpectedly in larger models and were not directly programmed or intended during training.

An AI system designed to infer a person's emotions or psychological states from their expressions.

A particularly sensitive use of emotion-inference systems, which the AI Act prohibits in the workplace and education settings, subject to exceptions.

The part of a model that converts input (e.g. text) into an internal representation suitable for further processing, such as classification or search.

The process of converting information into a form readable only by authorized persons.

The energy and water consumption associated with training and operating large models and running data centers; a growing topic in regulation and corporate responsibility.

A predefined procedure for how and to whom to report identified AI risks or incidents so that they reach the appropriate level of decision-making.

An office established within the European Commission that oversees in particular general-purpose AI models, coordinates the application of the AI Act and prepares guidance on it.

A coordination body composed of representatives of the Member States that supports the consistent application of the AI Act and cooperation in its implementation.