general-purpose AI model posing systemic risk (GPAISR)

in English: general-purpose AI model with systemic risk

professionalspreading

A high-performance AI model capable of handling many tasks, which under EU rules is considered so impactful that its errors could affect society as a whole.

In plain terms

The EU's Artificial Intelligence Act (AI Act) places the most powerful AI models in a separate category. These are the foundation models on which chatbots, coding agents, and thousands of other services are built. If such a model fails or is misused, the problem can show up in every product built on it.

It is like the separate handling of systemically important institutions at banks. Rules apply to small banks too, but they are stricter for the largest ones, because their collapse would shake the entire economy. Similarly, every general-purpose AI model has basic obligations, while the largest ones have additional tasks.

Classification does not mean the model is dangerous or banned. It means the developer must assess and manage large-scale risks in advance, such as assistance for cyberattacks, advice on producing dangerous materials, or loss of control.

How it works

The regulation first checks whether the model is general-purpose, meaning it can be used for various tasks and embedded in many products.
If the total computational volume used for training exceeds 10^25 floating-point operations, the regulation presumes the model is high-impact and therefore poses a systemic risk.
The developer (in the regulation's terminology: provider) must report this to the European Commission within two weeks at the latest. The Commission can also classify a model by its own decision, for example based on the number of users or its capabilities, and the developer can argue that its model does not pose such a risk after all.
The developer of a designated model must evaluate it and test it with adversarial tests, assess and mitigate systemic risks, report serious incidents, and ensure adequate cybersecurity.
Compliance with the rules is monitored by the AI Office operating within the Commission. For non-compliance, fines can reach up to 3 percent of global annual turnover or 15 million euros; the higher of the two amounts is the upper limit.

An everyday example

A domestic company is building a customer service agent on a large American developer's language model. The developer of the model must assess and manage its systemic risks, not the company. The company sees the model's capabilities and limitations from the developer's documentation. However, the company remains responsible for its own product: for example, it must inform customers that they are speaking with an AI.

Who is affected

developerIf you build on such a model, the provider's documentation and the model's limits determine what you can responsibly offer in your own product. → Code

entrepreneur, executiveWhen choosing a supplier, it matters whether the model developer meets EU obligations, as this affects the reliability of your service.

lawyer, citizenThis category identifies which AI developers the EU expects to provide enhanced transparency and risk management. → Law

everyoneThe chatbots you use daily typically run on such models, so the rules indirectly protect your safety too.

A common misconception

MythA systemic-risk classification means the model is dangerous and the EU wants to ban it.

In realityThe classification is not a ban but an additional obligation: the model remains usable, but the developer must test, document, and protect it more strictly. The AI Act bans certain usage practices, not models.

Limits and risks

The compute threshold is a rough metric: a model with lower compute requirements can still have a major impact, and the boundary may become obsolete over time due to more efficient training methods.
The obligations apply to the model developer, so they do not guarantee that a specific application or agent built on the model is also safe.
Compliance largely relies on developers' own assessments and reports, which are difficult to fully verify from the outside.
The rule applies only to models available on the EU market, and its practical enforcement is still evolving.

Similar, but not the same

High-risk AI system – There, the use case is decisive (for example, recruitment or credit assessment); here, it is the base model's power and prevalence.
general-purpose AI model – This broader category covers all general-purpose models; the systemic-risk group is only its highest-impact slice, with additional obligations.
foundation model – Base model is a technical term, while model of systemic risk is a legal classification carrying specific obligations.

Related concepts

Broader

general-purpose AI model, EU AI Act, AI regulation, Risk-based approach

Narrower

compute threshold, red teaming, reporting of a serious incident

On the edge

High-risk AI system, foundation model, AI Hivatal, code of practice, AI safety

A brief history

The category was introduced by the EU Artificial Intelligence Act, Regulation (EU) 2024/1689, which entered into force on 1 August 2024. The obligations for general-purpose models apply from 2 August 2025, and the voluntary code of practice to help with compliance was published in July 2025.

Where things stand

The obligations have been in force since August 2025, and the Commission has been able to impose fines since August 2, 2026. Models placed on the market before August 2, 2025, must comply by August 2, 2027. Practice is still taking shape: professional debate continues regarding classification, thresholds, and the method of verification.

Status: 2026 November

Also known as

general-purpose AI model posing systemic risk · systemic-risk general-purpose AI model · GPAI model with systemic risk · GPAI with systemic risk · GPAISR · general-purpose AI model with systemic risk · systemic-risk GPAI model · systemic-risk foundation model

Sources

Artificial Intelligence Act (Wikipedia)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.