AI Glossary

A means by which one can seek review or amendment of a decision, including decisions of supervisory authorities in the field of AI.

One of the legal bases for processing personal data, frequently relied upon for data processing in the development and operation of AI, provided it overrides the interests and rights of the data subjects.

A gradation of an agent's degree of autonomy according to the role a human holds toward it, from a mere observer through an approver and consultant to a collaborator and a fully autonomous executor of tasks.

The question of who bears responsibility for content created by AI and for its potential legal consequences, for example the infringement of third-party rights or the dissemination of false information.

The obligation to compensate for harm caused to another; with AI, it is crucial to determine which entity in the value chain is liable for the damage.

An authorization to use a copyrighted work under specified conditions.

A contract by which a rights holder grants authorization to use a work.

A trend where rights holders (e.g. publishers, music and image agencies) reach an agreement with AI providers on a paid license to use content for training or operating models, instead of going to court.

Rights protecting a person's likeness and identity against unauthorized commercial use.

Protection against the unauthorized use of a person's appearance or identity, including in the context of digital replicas and AI-generated content.

A family of Meta models released with open weights, widely used for self-hosted deployment and fine-tuning; it featured in the Kadrey v. Meta case.

The distinction between explaining one specific output (local) and explaining a model's overall behavior across many cases (global).

Automatically created records of an AI system's activity that enable retrospective checking, auditing and incident handling; they are mandatory for high-risk systems.

Techniques aimed at removing the influence of specific data from an already trained model without having to retrain it from scratch.

A technical means of expressing a reservation of rights (e.g., in metadata or a file on a website) that enables automated systems to recognize and respect the reservation.

A principle under which a human must have real, not merely formal, control over a high-risk system, including the ability to intervene in and override the system.

A component of the NIST framework in which an organization, based on identified and measured risks, selects and implements measures to address them and continuously reassesses them.

The responsibility of an organization's management for governance, oversight and ensuring that the use of AI complies with rules and strategy.

A regular evaluation of the functioning of the AI management system by top management, which is part of management system standards and a basis for improvement.

Practices in which AI uses subliminal or deceptive techniques to influence people's behavior in a way that may harm them; the AI Act prohibits some of them.

A component of the NIST framework focused on mapping the context: what the system is used for, how it works, whom it affects, and where it may fail.

An argument according to which generative AI harms the market not through direct copying, but through its ability to flood the market with an endless amount of competing outputs that displace the original works. In the Kadrey case, Judge Chhabria described it as potentially decisive in future disputes.

An authority that checks whether products and AI systems placed on the market meet legal requirements and is empowered to impose corrective measures.

The exercise of public oversight over whether AI systems placed on the market and put into service meet the requirements of legal regulations, including the possibility of imposing measures and sanctions.

An open standard enabling the secure connection of language models to external data sources, tools, and enterprise systems; it unifies the way assistants call functions and obtain context from applications.

A component of the NIST framework devoted to measuring and evaluating identified risks using both quantitative and qualitative methods.

A type of attack that determines whether a specific record was part of a model's training data, which may endanger the privacy of the individuals contained in the data.

A system layer that stores and provides access to the memory of an AI agent or application so that they can build on previous interactions and information.

Microsoft's AI assistant integrated into office and developer tools that helps with content creation, data analysis, and other tasks.

A text-to-image generator that has become the subject of copyright disputes over the reproduction of protected characters and visual works in its outputs.