AI Glossary

A set of voluntary principles of conduct for organizations developing advanced AI systems, agreed within the cooperation of the G7 countries.

Google DeepMind's family of multimodal models, integrated into Google services and developer tools.

A model capable of performing a wide range of tasks without being limited to a specific field or use, and of being integrated into various AI systems; referred to in the AI Act as a GPAI model.

We break down the obligations of GPAI model developers and the systemic-risk regime in our article AI Compliance – Do you develop or use artificial intelligence?

An AI system based on a general-purpose model that can be used for a wide range of applications and sectors without being limited to a single specific purpose.

An architecture with two competing networks (a generator and a discriminator), one of which creates content and the other evaluates it; historically used, among other things, for deepfakes.

Artificial intelligence that creates new content, such as text, images, audio or video, based on learned patterns.

A lawsuit filed in 2023 in which Getty Images accuses Stability AI of the unauthorized use of more than 12 million photographs, their captions, and metadata to build the Stable Diffusion and DreamStudio image generators.

A tool built into development environments that suggests and completes program code based on context; it has been the subject of copyright disputes over its training code.

A document adopted at the United Nations in 2024 that sets out shared goals and principles for digital cooperation among states, including a commitment to govern emerging technologies and AI for the benefit of humanity.

The goal of having individual national and international AI frameworks complement and recognize one another, thereby reducing the burden on organizations operating across multiple jurisdictions.

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.