Skip to content
BENIBIX Learn

General-Purpose AI Model with Systemic Risk

Also known as GPAI Model with Systemic Risk, GPAI with Systemic Risk, GPAISR, Systemic Risk (EU AI Act)

The EU AI Act's top tier for the most powerful AI models. They carry extra duties.

Think of it like

It is like a very big bridge. Every bridge gets checked. The biggest get more checks, since so many cross.

Example

A lab trains a new chat model with more compute than the line allows. Within two weeks it tells the Commission. Before launch, it tests whether the model could help with cyber attacks.

Why it matters

A flaw in one huge model can spread to every app built on it. These duties aim to catch it at the source.

Definition

A general-purpose AI model with systemic risk is a model so strong that the EU AI Act gives it extra rules. A model counts as one if its training used more than 10^25 computing steps. Its maker must then tell the Commission within two weeks. It must test the model for harm, even by attacking it on purpose. It must also cut the risks it finds and report serious incidents. And it must guard the model from hackers.
Every model maker has the base duties. Past the compute line, the extra duties start.

Common mistake

Thinking the compute line is the only test. The Commission can also name a smaller model if its reach or skills are high enough.

Key takeaways

  1. 1
    A compute line. Over 10^25 steps means it is presumed.
  2. 2
    Extra duties. Test, cut risks, report and secure.
  3. 3
    Two weeks to tell. The Commission must hear within two weeks.