Anthropic just did something no other major AI lab has done: it watermarked every Claude model released on or after August 2, 2026 — worldwide, not just in Europe. The move came three days after the EU AI Act's Article 50(2) transparency rules kicked in, and Anthropic framed it as compliance. But here's the twist: the watermark is invisible, the detector is not public, and nobody outside the company can verify it works. That's like a restaurant posting a health grade it printed itself — and refusing to let the inspector inside.
The technical lineage is deliciously ironic. Anthropic reportedly deployed SynthID-Text, the watermarking scheme built by Google DeepMind. Google spent years publishing watermark research that many assumed would stay academic. Now a direct competitor has baked it into production compliance. Google wins the standards war without shipping the product. Meanwhile, OpenAI — the market leader — hasn't watermarked ChatGPT at all. So the industry's most safety-forward lab is the only one wearing the invisible scarlet letter.
But the real problem is bigger than brand rivalry. A new arXiv paper, 'Watermarks Without Verification: AI Text Watermarking After the EU AI Act,' argues that neither the user complaints nor the company promises can be checked, because the actual systems are black boxes. The paper empirically tests SynthID-Text on prose quality and code generation — exactly where users say things get worse. If the mark measurably degrades code, that's not a compliance win. That's a commercial wound.
And the mark itself is fragile. Copy text into a different editor, reformat it, translate it, paraphrase it — the signal degrades or vanishes. Cybernews notes watermarks fail after edits or platform changes. So the platform specifically worried about AI text laundered through humans is exactly where the mark stops working. One widely-shared blog post claims watermarking 'fails 98% of the time in 2026.' Treat that number with suspicion — no methodology is visible — but the direction matches the arXiv findings.
The sharpest contrast comes from California. The state's AI Transparency Act, taking effect in 2027, will require large platforms to provide a free tool to detect watermarks. The EU mandates the mark and says nothing about the reader. California is demanding verification; Brussels settled for theater. That's the difference between a lock and a lock you can actually open.
So who loses? Paying Claude users who see quality drops and can't opt out. Researchers who can't test the claims. And the EU's credibility, if this becomes the template for future AI rules. Who wins? Google, whose watermark is now the de facto standard. And possibly OpenAI, which gets to watch Anthropic absorb the backlash.
Trends suggest this could become the defining fight of AI transparency: not whether we mark synthetic text, but whether anyone can prove it. If California's approach spreads, companies might have to open their detectors. If it doesn't, we'll have a generation of invisible marks that mean nothing. Your move, Brussels.