AI Glossary

Google's research and development division behind the Gemini family of multimodal models and video-generation models; a major player among providers of GPAI models.

The overarching component of the NIST framework that establishes an organizational culture of risk management, along with the roles, responsibilities, and policies within which the other activities are situated.

A comprehensive framework of rules, roles, and processes that helps an organization systematically manage the development, deployment, and oversight of AI in compliance with the law and internal policies.

A voluntary instrument published by the EU AI Office on 10 July 2025 that helps providers of GPAI models demonstrate compliance with the AI Act. It was prepared by independent experts in a multi-stakeholder process and is divided into three chapters.

A general-purpose model whose capabilities or impact reach a level at which they may pose a risk to society as a whole. It is subject to the obligations under Article 55 of the AI Act.

A series of large language models from OpenAI that power ChatGPT, among other things; newer generations add multimodal capabilities and more advanced reasoning.

A conversational model from xAI integrated into the X social network platform.

Anchoring a model's responses in verified sources or data, which reduces the risk of hallucinations and increases the factual accuracy of its outputs.

Protective mechanisms that define what a model may and may not do, preventing harmful, inappropriate, or dangerous outputs and behavior.

An interpretive document issued by the Commission in July 2025 that helps providers understand when a model is considered a general-purpose AI model and what obligations apply to it.

An indicator of how often a model produces false or fabricated information; used to compare the reliability of models and to manage risks.

Specialized hardware (e.g., graphics or tensor processors) designed to accelerate the computations typical of the training and operation of models.

A technical standard adopted by the European standardization organizations, compliance with which establishes a presumption of conformity of a high-risk AI system with the requirements of the AI Act.

A common framework shared by management system standards (quality, information security, AI), which facilitates their simultaneous implementation within a single organization.

Editorial annotations and summaries of points of law attached to court decisions in the Westlaw database. In the Ross Intelligence case, the court recognized their originality and copyright protection.

A group of class action lawsuits by book authors against Apple, which place it among the technology companies facing copyright disputes over AI training.

An AI system falling into the high-risk category under the AI Act, subject to strict requirements on risk management, data quality, documentation, transparency and human oversight.

A platform and community repository of models and datasets, key to the sharing and deployment of open models.

An arrangement in which the AI proposes a course of action or output, but its execution must be approved by a human before it is carried out; a form of human oversight.

A requirement that fundamental employment decisions not be left solely to AI.

A value whose protection the AI Act emphasizes, particularly in prohibiting manipulative practices and disproportionate social scoring.

Human evaluation of a model's outputs, used to fine-tune its behavior toward usefulness and safety.

A set of measures enabling a human to monitor the operation of an AI system and to intervene where necessary.

An approach to the design and use of AI that places the human being, their rights, interests and dignity at the center.

An arrangement in which a human remains part of the AI decision-making process and can intervene in it, review it or override the system's output.

A search combining classic keyword-based search with vector-based search by meaning, so that the results are more precise and relevant.

The sequence of hardware, software, and service suppliers needed to operate AI; its vulnerabilities can be a source of security and legal risks.

An entity that places on the EU market an AI system originating from a provider established outside the Union; the AI Act imposes verification and documentation obligations on it.

The ability of a model to use information contained in the current context without additional training.

Public or sector-specific collections of recorded failures and harms caused by AI systems, serving for learning and prevention.