A view from Brussels: White smoke on AI Omnibus, but are lessons really learned?

White smoke emerged from the European institutions 7 May as negotiators reached a political agreement on the AI Omnibus set to amend the EU Artificial Intelligence Act. This happened a week after a 28 April first failed attempt at reaching an informal deadline. The ink is barely fresh on this agreement — or on the AI Act for that matter — so the wisest thing to do right now is to take a few steps back. 

By all accounts, the original AI Act negotiations were rushed. European elections were looming and the European Commission self-inflicted a hard deadline to ensure its flagship AI Act proposal would be agreed before the 2024 elections. It was understandable politically: some people needed the win; the momentum might have suffered from the lame duck season; and AI uptake was real, so was the need for guardrails.

With the AI Omnibus, the pressing factor is also in part self-inflicted. From a calendar perspective, it is based on retro-planning against the AI Act high-risk provisions kicking in 2 Aug. 2026. From a political perspective, the pressure came in part from the explosive combination of competitiveness ambitions and AI race geopolitics, mounting pressure on the Commission to deliver on some of the recommendations outlined in the Draghi report. 

The original AI Act course of negotiations was also impacted by knee-jerk reactions and event-driven positioning. The most impactful example happened a year into the negotiations. The wider public became aware of this little thing called "general-purpose AI" with the launch of ChatGPT in November 2022. This prompted for including general-purpose AI and foundational models in the AI Act, leading to a notable expansion of its material scope. 

With the AI Omnibus, this happened to some extent. The agreement expands the material scope of the AI Act by introducing a ban on "nudifier" apps and child sexual abuse material, a strong ask that came from Parliament in the wake of several scandalous headlines during the negotiations.

Legislating is hard. It is unforgiving. I am not opining on the substance of the AI Omnibus agreement and whether outcomes are good or bad. But I am questioning what this exercise tells us about legislative work given how much it means for governance professionals right now. 

Is it possible to simplify legislation or is it too complex to make things simpler? Does this agreement guarantee that the AI Act regime is now set, irrespective of upcoming guidelines, and provides certainty and clarity to operators? 

I can't help wondering what the AI Act would have looked like if back two years, legislators afforded themselves a bit more time to come to a finished product, perhaps allowing some of the shortcomings addressed not to be there in the first place. With ifs and buts…