AI Glossary

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.

A series of lawsuits related to the collection of content (e.g., videos) for training models, illustrating the expansion of disputes to hardware and platform providers as well.

An approach of the Czech legislator whereby the adaptation act regulates only the necessary elements and leaves detailed obligations directly to the text of the AI Act, so as not to increase the administrative burden on companies.

A European (French) provider of language models, often highlighted in connection with European competitiveness in the field of AI.

A model architecture in which only a part of the specialized subnetworks (experts) is activated for each input, which enables larger models at lower computational cost per response.

An effort to ensure that a model's goals and behavior correspond to human values and intentions, thereby preventing undesirable or harmful conduct.

A structured document describing the purpose, capabilities, limitations, and both recommended and inappropriate uses of a model, serving transparency toward users.

The gradual degradation of the quality of models trained predominantly on content created by another AI, where diversity and fidelity to the original data are lost.

A standardized model documentation form that a GPAI provider must prepare before placing a model on the EU market and which summarizes the information required by Annexes XI and XII of the AI Act.

The gradual deterioration in the accuracy of a deployed model caused by real-world data gradually diverging from the data on which the model was trained.

Systematic testing of a model's capabilities, safety and reliability using task suites; the basis for deployment decisions and for demonstrating compliance.

A type of attack in which an attacker mimics the behavior of a target model through repeated queries and creates a functional copy of it.

The ability to understand the internal logic of how an AI model works.

A type of attack that attempts to reconstruct the sensitive input data used during training from a model's outputs.