The EU spent three years building the world's most comprehensive AI law. This week, it spent one night tearing it apart.

EU legislators struck a provisional deal early Thursday morning that postpones high-risk AI restrictions by more than a year and largely exempts industrial AI from the law's scope entirely. The negotiations lasted until 4:30 AM.

What Actually Changed

  • High-risk AI rules delayed from August 2026 to December 2027 — a 16-month pushback
  • Industrial AI exempted from the AI Act. Companies using AI in manufacturing now only face machinery regulations
  • Watermarking grace period cut to three months instead of six
  • Sexual deepfake ban added after global outrage over Grok's misuse

Germany got exactly what it wanted. Chancellor Friedrich Merz pushed hard to exempt industrial AI so Siemens and Bosch wouldn't face what officials called a "double regulatory burden."

Why This Happened

The official story: the EU was moving too fast while the rest of the world watched. Ursula von der Leyen called it "a simple, innovation-friendly environment."

The real story: pressure from three directions.

Industry: German manufacturers, French AI labs, and European tech companies spent months lobbying for lighter rules. The message was consistent — strict regulation puts Europe at a competitive disadvantage.

The US: The Trump administration has been hammering EU tech regulation as anti-American. The AI Act was a target. Rolling it back is diplomatic currency.

Competition: China isn't regulating AI like this. The US isn't either. Europe was alone, and von der Leyen admitted it: "only a couple of countries around the world followed the EU's lead."

What This Means

The EU's AI Act was supposed to be the global standard. The Brussels Effect — where EU regulation becomes de facto world law — has worked for data privacy and product safety.

Not this time. The EU passed the AI Act, the world shrugged, and now the EU is walking it back.

That precedent is dangerous. If the EU can't make its own AI law stick, what regulatory model works? The US voluntary approach? China's state-controlled model? Something else entirely?

The Deepfake Exception

The one new restriction — banning AI-generated sexual deepfakes — was added after Elon Musk's Grok was used to create abusive content at scale. Even in a rollback, some harms are too visible to ignore.

But the ban is narrow. It covers "identifiable" people and child sexual abuse material. It doesn't address the broader deepfake ecosystem. That's not a comprehensive policy — it's a reaction to a specific scandal.

The Verdict

The EU's AI Act was always a bet that Europe could regulate its way to AI leadership. This week's deal is an admission that the bet didn't pay off.

The question now isn't whether AI needs regulation. It's who writes the rules, and whether any single jurisdiction can enforce them alone. The EU just proved it can't.

Published May 8, 2026. See something off? Drop us a note.