AI Glossary

Verification of an AI system outside the laboratory environment, but under the supervision of the competent authority and under specified conditions, often within a regulatory sandbox.

A use of an AI system that, although not intended, can be reasonably expected.

The ability of a model to solve tasks through multi-step inference; so-called reasoning models devote part of their computation to thinking before generating an answer.

A model optimized for more complex logical and analytical tasks.

Systematic testing of a model using targeted attacks and provocations in order to uncover weaknesses, risks and vulnerabilities.

Targeted testing of a model by a team that attempts to elicit harmful or dangerous behavior in order to identify and eliminate weaknesses before deployment.

The obligation to enter a high-risk AI system into a public EU database before it is placed on the market or put into service, serving transparency and oversight.

The official designation of the Artificial Intelligence Act, the world's first comprehensive legal framework for AI, based on a risk-tiered approach. It entered into force in 2024 and applies gradually.

A risk arising from a change in legislation or its interpretation.

A controlled environment that allows the testing of innovative AI solutions under the supervision of a regulator.

A controlled environment created by a supervisory authority that allows the testing of innovative AI systems before they are placed on the market.

A method of learning in which a system improves based on rewards and penalties for its actions; in language models it is used especially to fine-tune behavior.

The concurrent application of both regulations, where the AI Act complements rather than replaces the rules on personal data protection; in the case of biometrics and profiling the regimes overlap.

The recognition of persons at a distance on the basis of biometric data; its use in publicly accessible spaces is fundamentally restricted by the AI Act.

An agent focused on searching for, gathering and evaluating information from various sources.

An approach to the development and use of AI based on the principles of accountability, transparency and fairness.

A comprehensive set of principles for developing and deploying AI in a way that is ethical, safe, transparent and fair.

A sequence of steps through which a system retrieves and prepares relevant information before generating an answer – from the query, through retrieval and ranking, to passing the most suitable materials to the model; the foundation of RAG-type systems.

Deployment of RAG in enterprise applications, where the model retrieves relevant information from company documents before answering, thereby reducing hallucinations.

Updated EU legislation that explicitly includes software and AI systems among products and makes it easier for injured parties to prove a defect and causation.

The contested question of how to fulfill a data subject's right to erasure when their data is already baked into a trained model from which it cannot be easily removed.

The right of a person affected by an automated decision to obtain an understandable explanation of its underlying logic and consequences.

An expression of will by a rights holder reserving that their work may not be used for text and data mining; it must be expressed in a machine-readable manner.

A set of processes through which a provider continuously identifies, assesses, monitors and mitigates the risks associated with an AI system.

A record of identified risks, including their assessment, owner and the measures adopted; a tool linking AI governance with enterprise-wide risk management.

The level of risk that an organization is willing to knowingly accept; it should be set out explicitly and reflected in decisions about deploying AI.

The core regulatory logic of the AI Act, which classifies systems according to their level of risk into categories ranging from unacceptable risk (prohibition) through high risk (strict obligations) and limited risk (transparency obligations) to minimal risk.

A method of improving models using feedback from people who rate the quality of the outputs.

A file on a website expressing rules for automated bots; in the AI context it has become one of the tools by which websites limit the collection of content for training.

A platform using generative AI for the creation and editing of video, animation, and multimedia content.