AI Glossary

The controlled shutdown of an AI system that is causing disproportionate harm or behaving undesirably, including the ability to override or correct it.

A measure imposed by a supervisory authority to eliminate the risk or non-compliance of an AI system, for example withdrawal from the market or restriction of operation.

The most extensive part of the code, applying only to providers of GPAI models with systemic risk, which describes the process of identifying, assessing, and mitigating systemic risks before major deployment decisions.

A document that a signatory must produce before placing a model with systemic risk on the market; it contains evaluation triggers, risk categories, mitigation strategies, forecasting methods and the allocation of responsibilities.

A component of a product whose failure may endanger health or safety. Following revisions to the Act, systems serving only for optimization, automation or convenience are not regarded as safety components, provided their failure does not endanger health.

A protective mechanism that filters or modifies a model's inputs and outputs in order to prevent harmful or inappropriate content.

A fine or other measure imposed for a breach of AI-related obligations.

Fines for violations of the AI Act, graduated according to severity; the highest apply to violations of the ban on unacceptable practices and can reach a significant share of worldwide turnover.

An entity that participates in the testing of its AI system in a regulatory sandbox under the supervision of the competent authority.

A risk-identification method based on thinking through possible scenarios of misuse or failure of a model, supplemented by a risk inventory and consultations with internal and external experts.

An advisory body established at the EU level, composed of independent experts, that supports the AI Office and supervisory authorities in particular in the assessment of models with systemic risk.

The automated retrieval of data from web pages or other sources.

The question of whether, and under what conditions, data publicly available on the internet may be used for training without violating privacy protection rules.

The manipulation of the sources a system draws on during retrieval, so as to smuggle false or harmful information into its responses.

A form of attention mechanism in which each element of the input sequence compares its relevance to all other elements of the same sequence.

Search based on the meaning of the query rather than merely on keyword matching.

An event that leads or potentially leads to significant harm to health, safety, or fundamental rights.

The obligation of signatories to document and report serious incidents to the AI Office and the relevant national authorities, with graduated deadlines depending on severity.

The use of AI tools by employees without the knowledge or approval of the organization.

An unauthorized repository of digital copies of works; the use of data from such sources for training was at the core of both the Bartz and Kadrey lawsuits.

A contractual agreement on the quality, availability, and support of the provided service.

A smaller and less demanding language model that can be run at lower cost, on end devices or for narrowly defined tasks.

The assessment of persons based on their behavior or personal characteristics, which may lead to disproportionate disadvantage.

A procedure envisaged by the Czech draft, whereby for less serious violations the authority may first issue a warning and set a reasonable period for remedy (at least 15 days), and only then proceed to a sanction.

A model developed by OpenAI focused on generating videos from text instructions.

An effort by a state or region to secure its own capacities for the development and operation of AI, including computing infrastructure, data, and models, with the aim of reducing dependence on foreign providers.

Sensitive data, such as data on health, biometric data, political opinions, or religious beliefs, which enjoy enhanced protection.

An open-source model designed for creating and editing digital images using artificial intelligence.

A person or entity that is affected by the development or deployment of AI or who can influence it; their involvement is a common requirement of international frameworks.

Model provisions prepared primarily for public sector purchasers that facilitate the procurement of AI systems in compliance with the AI Act.