EU countries strike deal on landmark AI rulebook

BRUSSELS — European Union member countries on Friday unanimously reached a deal on the bloc’s Artificial Intelligence Act, overcoming last-minute fears that the rulebook would stifle European innovation. 

EU deputy ambassadors green-lighted the final compromise text, hashed out following lengthy negotiations between representatives of the Council, members of the European Parliament and European Commission officials. 

The law would ban some applications of AI technology, impose strict limits on use cases considered high-risk, and hem in the most advanced software models with obligations of transparency and stress-testing. 

The EU is the first out of the blocks in laying down binding rules for fast-moving AI  technology. Even if many countries and international clubs — from the OECD to the G7 — have spent the past few years pondering how to regulate AI, most have stuck to voluntary guidelines or codes of practice. 

When EU policymakers announced they had found a final compromise on the AI Act’s content in December, the breakthrough was hailed as a pioneering step Europe should celebrate amid the rise of ubiquitous AI tools such as OpenAI’s ChatGPT and Google’s Bard. 

But the achievement rubbed some EU countries the wrong way. Over the past few weeks, the bloc’s top economies Germany and France, alongside Austria, hinted that they might oppose the text in Friday’s vote. 

While Vienna’s beef was with data protection provisions, Paris and Berlin warned that rules for advanced AI models would hamstring Europe’s budding AI champions, such as France’s Mistral and Germany’s Aleph Alpha. With Italy — sometimes an AI Act critic — keeping mum on its intentions, the AI Act’s fate was suddenly in question, as four opposing countries would be enough to derail the law for good.

The cabinet of French Economy Minister Bruno Le Maire called for a further round of negotiations with the European Parliament to address his concerns. The prospect horrified the Belgian Council presidency, due to the lack of time for further negotiations. Making things worse, the Parliament itself was dealing with a simmering row over the AI Act’s facial-recognition rules, triggered by privacy hawk Svenja Hahn, an MEP.

Eventually, the matter was resolved through the EU’s familiar blend of PR offensive and diplomatic maneuvering. The Commission ramped up the pressure by announcing a splashy package of pro-innovation measures targeting the AI sector, and in one fell swoop created the EU’s Artificial Intelligence Office — a body tasked with enforcing the AI Act. 

Austria, France and Germany were lured back into the fold with promises that the Commission would make formal declarations reassuring them on their pet issues, two people involved in the negotiations told POLITICO. These declarations are not legal acts, but as the Commission will oversee the AI Act's implementation, they would provide a guarantee of sorts.

A spokesperson for German Digital Minister Volker Wissing, the foremost AI Act skeptic within Germany’s coalition government, told POLITICO: "We asked the EU Commission to clarify that the AI Act does not apply to the use of AI in medical devices."

The AI Act still needs the formal approval of the European Parliament. The text is slated to get rubber-stamped at the committee level in two weeks, with a plenary vote expected in April.

Disgruntled pro-privacy lawmakers might still try to hamper the law’s progression by proposing amendments — which, if adopted, will need additional negotiations with the Council.  But most people who worked on the AI Act within Parliament are confident that the law will pass with no change.