---
title: "General-purpose AI (GPAI): definition and example – Carl-Uwe Würfel"
description: "GPAI explained: what the EU AI Act classifies as a general-purpose AI model, which obligations and deadlines apply, and what it means for companies using them."
url: https://carl-wuerfel.de/en/glossary/gpai
author: Carl-Uwe Würfel
---

1. [Home](https://carl-wuerfel.de/en)
  2. [Knowledge](https://carl-wuerfel.de/en/knowledge)
  3. [Glossary](https://carl-wuerfel.de/en/glossary)
  4. General-purpose AI (GPAI)

# General-purpose AI (GPAI)

GPAI (general-purpose AI) is the EU AI Act’s legal term for general-purpose AI models: broadly usable models whose providers carry their own set of obligations.

## On this page

  1. What counts as a GPAI model
  2. Two tiers: with and without systemic risk
  3. What providers have to do
  4. Deadlines and the state of play in November 2026
  5. What it means for companies that only use GPAI models
  6. Where the term stops working
  7. Frequently asked questions
  8. Sources

Clarifying your own role is the first step.

_Figure: A decision path using the example of a logistics company. It builds an assistant that answers delivery inquiries on top of a large vendor 's language model and uses the model through the vendor's API. During procurement it requests the information for downstream providers and records which model runs for which task. The deciding question is its own role: a company that substantially modifies a model, or puts a system built on one to a high-risk use, can become a provider itself. Otherwise the Article 53 obligations stay with the vendor, as they generally do when a model is used through an API or a subscription. The company also takes care of AI literacy for its staff._

**Example**

A logistics company builds an assistant that answers delivery inquiries on top of a large vendor’s language model. During procurement it requests the information for downstream providers and records which model runs for which task. The Article 53 obligations stay with the vendor; the company takes care of AI literacy for its staff and checks whether its use falls under the high-risk or transparency rules.

The EU [AI Act](https://carl-wuerfel.de/en/ai-act) regulates AI at two levels: AI systems, meaning applications with a specific purpose, and the models many of those applications are built on. For the models, it created a category of its own, the general-purpose AI model, or GPAI. Article 3(63) defines it as a model that displays significant generality, can competently perform a wide range of distinct tasks and can be integrated into a variety of downstream systems. Models used only for research, development or prototyping before being placed on the market are excluded.

In technical terms, this mostly covers large language, image and video models, which researchers call foundation models; the [foundation model](https://carl-wuerfel.de/en/glossary/foundation-model) entry covers how they work. GPAI, by contrast, is a legal classification. It determines who carries which obligations: providers of these models must document them, respect copyright and disclose what they were trained on. Particularly capable models with systemic risk come with additional testing and reporting duties. Oversight lies with the European Commission and its AI Office.

The term matters to executives because almost every AI tool a company buys runs on a GPAI model. The model obligations generally fall on the vendor, not on the company that uses a model through an API or a subscription. But a company that substantially modifies a model, or puts a system built on one to a high-risk use, can become a provider itself. Clarifying your own role is the first step.

## What counts as a GPAI model.

The definition in Article 3(63) has three elements: significant generality, the ability to competently perform a wide range of distinct tasks, and suitability for integration into many downstream systems. How the model reaches the market, whether through an API, as an open-weights download or inside an application, expressly does not matter.

Because these elements are fuzzy, the European Commission published guidelines in July 2025. As an indicative criterion, a model counts as a GPAI model if its training required more than 10²³ floating-point operations (FLOP) and it can generate language as text or audio, images from text or video from text. The criterion is not a hard line: a model above it may exceptionally be too narrow, and one below it may still be general enough.

The model has to be kept apart from the general-purpose AI system (Article 3(66)). It is based on a GPAI model and can serve many purposes, used directly or built into other systems, such as a chat assistant with a user interface. The model is subject to Chapter V of the Act; the system falls under the rules for AI systems, up to the high-risk obligations depending on its purpose.

## Two tiers: with and without systemic risk.

The Act distinguishes GPAI models with and without systemic risk. Under Article 3(65), a systemic risk stems from high-impact capabilities and has a significant effect on the EU internal market because of the model’s reach or possible consequences for health, safety, fundamental rights or society.

Under Article 51, a model is presumed to have such capabilities if its training required more than 10²⁵ floating-point operations. The provider must notify the Commission within two weeks under Article 52. It can argue that its model, despite crossing the threshold, exceptionally poses no systemic risk; the Commission decides. Conversely, the Commission can designate a model below the threshold, using criteria from Annex XIII such as the number of parameters, the size of the training data or the model’s reach. High impact on the market is presumed once a model is available to at least 10,000 registered business users established in the EU.

The Commission publishes a list of models with systemic risk and can adjust the thresholds by delegated act to keep pace with the technology. Systemic risk is not a ban. These models may be offered; their providers carry additional obligations.

## What providers have to do.

Under Article 53, every GPAI model provider must keep technical documentation and hand it to the AI Office on request. It must give companies that build the model into their own systems information on its capabilities and limitations. It needs a policy for complying with EU copyright law, in particular for honoring rights holders’ opt-outs from text and data mining under Article 4(3) of Directive (EU) 2019/790. And it publishes a summary of the training content using a template the Commission released in July 2025. Providers outside the EU appoint an authorized representative in the Union (Article 54).

For models released under a free and open-source license with public weights, the first two obligations fall away; the copyright policy and the summary remain. The exemption does not apply to models with systemic risk, and Article 55 adds model evaluations including adversarial testing, assessment and mitigation of systemic risks, reporting of serious incidents to the AI Office, and adequate cybersecurity.

The voluntary General-Purpose AI Code of Practice of July 10, 2025 describes how to meet these obligations, with chapters on transparency, copyright, and safety and security. Signatories include Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI (as of November 2026). For violations, the Commission can impose fines under Article 101 of up to 3 percent of worldwide annual turnover or 15 million euros, whichever is higher.

## Deadlines and the state of play in November 2026.

The obligations for GPAI model providers have applied since August 2, 2025. Models placed on the market before that date must comply by August 2, 2027 under Article 111(3). The Commission’s enforcement powers, including fines under Article 101, have applied since August 2, 2026.

Regulation (EU) 2026/1744, the Digital Omnibus on AI of July 8, 2026, entered into force on July 27, 2026. It did not move the GPAI deadlines. It did reorganize supervision: the AI Office is now solely responsible for AI systems based on a GPAI model from the same provider or the same corporate group, with exceptions such as certain high-risk areas. Providers of systems, including general-purpose AI systems, that generate synthetic content and were placed on the market before August 2, 2026 must implement the marking obligation in Article 50(2) by December 2, 2026. The amendment pushes the high-risk rules back to December 2, 2027 for Annex III and August 2, 2028 for Annex I.

The rules keep moving through delegated acts and harmonized standards, so deadlines should be checked against the current consolidated text.

## What it means for companies that only use GPAI models.

Most companies do not provide a GPAI model; they use one, through a vendor’s API, a cloud platform or as an open model on their own hardware. The obligations in Articles 53 and 55 do not apply to them. Under the Commission’s guidelines, a company that fine-tunes or modifies a model only becomes a model provider itself if the modification uses more than a third of the original training compute.

Other obligations remain. A company that builds a model into its own AI system is a downstream provider (Article 3(68)) and can request the Article 53 information from the model provider. A company that puts a system, including a general-purpose AI system, to a high-risk use it was not intended for becomes the provider of a high-risk system under Article 25. On top of that come AI literacy for staff under Article 4 and the transparency duties in Article 50, for example for deepfakes.

In practice, that means keeping an inventory of which model runs in which application for what purpose; requesting the downstream documentation during procurement; agreeing in the contract who notifies whom when the model changes; and checking every new [use case](https://carl-wuerfel.de/en/glossary/use-case) against the high-risk list.

## Where the term stops working

GPAI is a legal classification of the model and says nothing about the quality or danger of an application. Whether a use is high-risk depends on the purpose of the AI system under Article 6, not on the model underneath; the two can coincide. The underlying technology is described by the term foundation model, which overlaps largely, but not completely, with GPAI. This entry is not legal advice for individual cases.

## Frequently asked questions

What does GPAI mean?

GPAI stands for general-purpose AI; in the EU AI Act it refers to general-purpose AI models. These are broadly usable models that can perform many different tasks and be built into a wide range of applications. Their providers have specific obligations under Articles 53 and 55.

What is the difference between GPAI and a foundation model?

Foundation model is a technical term from research; GPAI is a legal classification under the EU AI Act. Both largely describe the same models. The legal definition, however, focuses on generality and placing on the market, and attaches obligations to it.

When do the GPAI obligations apply?

Since August 2, 2025 for new models, and from August 2, 2027 for models that were already on the market. The Commission has been able to impose fines since August 2, 2026. The 2026 Digital Omnibus on AI did not move these deadlines (as of November 2026).

Which GPAI models pose a systemic risk?

Systemic risk is presumed for models whose training required more than 10²⁵ floating-point operations. The Commission can also designate models based on other criteria such as reach or capabilities, and it publishes a list. Systemic risk means additional obligations, not a ban.

Does a company that uses ChatGPT or another GPAI model have to meet the Article 53 obligations?

Usually not, because Article 53 is aimed at model providers. Under the Commission’s guidelines, a company only becomes a provider if it modifies a model using more than a third of the original training compute. Its own use is still covered by the rules for AI systems, such as AI literacy and transparency.

What is the GPAI Code of Practice?

The General-Purpose AI Code of Practice is a voluntary rulebook from July 2025 that providers can use to demonstrate compliance. It has chapters on transparency, copyright, and safety and security. Providers who do not sign it must show compliance by other means.

Is GPAI related to the Global Partnership on AI?

No. The Global Partnership on Artificial Intelligence is an initiative of governments on AI policy, founded in 2020 and working jointly with the OECD since 2024. It shares the acronym but creates no obligations for companies.

## Further reading

  * [**AI glossary**](https://carl-wuerfel.de/en/glossary)
  * [**EU AI Act for mid-sized companies**](https://carl-wuerfel.de/en/ai-act)
  * [**Foundation model**](https://carl-wuerfel.de/en/glossary/foundation-model)
  * [**AI labeling and disclosure**](https://carl-wuerfel.de/en/glossary/ki-kennzeichnung)
  * [**LLM: Large Language Model**](https://carl-wuerfel.de/en/glossary/llm)
  * [**Fine-Tuning** Term](https://carl-wuerfel.de/en/glossary/fine-tuning)
  * [**Generative AI** Term](https://carl-wuerfel.de/en/glossary/generative-ki)

## Sources

  1. [EU: Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 3(63) to 3(68), 25, 51 to 55, 101, 111 and 113, 2024](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)
  2. [EU: Regulation (EU) 2026/1744 amending the AI Act (Digital Omnibus on AI), 2026](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng)
  3. [European Commission: Guidelines for providers of general-purpose AI models, 2025](https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers)
  4. [European Commission: Q&A on the guidelines on obligations for general-purpose AI providers, 2025](https://digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-gpai-providers)
  5. [European Commission: The General-Purpose AI Code of Practice, with signatories, 2025](https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai)
  6. [European Commission: Explanatory Notice and Template for the Public Summary of Training Content, 2025](https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models)
  7. [EU: Directive (EU) 2019/790 on copyright in the Digital Single Market, Article 4, 2019](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32019L0790)
  8. [OECD.AI: About the Global Partnership on Artificial Intelligence (GPAI)](https://oecd.ai/en/about/about-gpai)