France, Germany and Italy have put forward a joint proposal for how the European Union should regulate artificial intelligence. Their approach would ask AI providers to follow a common code of conduct, while focusing regulation on what AI systems are used for rather than applying broad rules to foundation models themselves.
The proposal comes as the European Commission, the European Parliament and the EU Council negotiate the bloc’s position on AI regulation. It offers a way to set expectations for providers while avoiding an initial sanctions regime.
A shared code providers can sign up to
The three governments support voluntary rules for both large and small AI providers in the EU. The rules would become binding for companies that sign up to them, but there would be no initial sanctions for violations.
This arrangement combines a common set of expectations with a choice for providers about whether to join. It would apply across company sizes, rather than making the code relevant only to the largest AI developers.
The governments have raised concerns about a narrower approach. If a code were binding only on major providers, they argue, customers might trust smaller European providers less and those companies could lose business. Extending the voluntary rules to large and small providers is intended to avoid that distinction.
Regulate uses of AI, not foundation models
France, Germany and Italy argue that horizontal rules for foundation models would conflict with the AI Act’s technology-neutral, risk-based approach. Their alternative is to regulate applications of AI, with attention to the risks tied to particular uses.
Germany’s Digital Minister Volker Wissing explained the position to Reuters: “We need to regulate the applications and not the technology if we want to play in the top AI league worldwide.” The proposal therefore draws a line between rules for specific applications and broad requirements imposed on the underlying technology.
The European Parliament’s AI Act, unveiled in June, aims to manage risks from AI applications, prevent discriminatory effects and promote innovation. The three governments’ argument is that regulation should follow those application-level risks rather than classify foundation models under a single horizontal set of rules.
Model cards would describe capabilities and limits
Although the proposal resists broad binding rules for foundation models, it does call for information about them. Developers would define model cards that describe a model’s capabilities and limitations.
An AI governance body would develop guidelines for those cards and review their use. In practice, this would create a way to document what a model can and cannot do, while also giving a governance body a role in setting expectations for that documentation.
The model cards are part of the proposal’s emphasis on oversight and transparency. They would not replace rules for AI applications; instead, they would provide information about the foundation models on which applications may rely.
Monitoring could come later
The proposal does not rule out stronger oversight in the future. Reuters reports that a European authority could be established to monitor compliance with the standards if violations of the code of conduct are found after a certain period.
The non-paper seen by Euractiv says that “any suspected violation in the interest of transparency should be made public by the authority.” That would make public disclosure part of the proposed oversight, though the authority is described as a possible future step rather than an immediate enforcement body.
For now, the plan pairs sign-up rules without initial sanctions with documentation for foundation models and the prospect of later monitoring. Its details will be considered as the European Commission, the European Parliament and the EU Council continue negotiating the EU’s approach to AI regulation.