The European Union just turned every AI text generator into a digital forensics lab. Starting August 2, 2026, AI-generated or modified text that touches public-interest topics must be machine-detectably watermarked. And if you think you can dodge this by being based outside Europe, think again: the rules apply to any provider whose system is accessible to EU citizens. That includes open-source projects like Ollama and LM-Studio, and every chatbot that can be reached from a Paris café.
Google is already playing by the new rules. Its SynthID system, which now covers text, returns one of three labels: watermarked, not watermarked, or uncertain. That third state is a problem: short AI snippets often fall into "uncertain" territory, which could mean the regulation's promise of clear identifiability has a gap you could drive a llama through. Short tweets? Quick replies? Grey zone.
Then there's OpenAI. The company has an internal text watermarking tool that it claims is 99.9% effective. It works by biasing token selection during ChatGPT generation, creating a statistical fingerprint. But OpenAI hasn't deployed it. Internal documents reveal a careful weighing of accuracy against user friction and fear of circumvention. With the deadline looming, the delay looks less like prudence and more like a high-stakes game of chicken with regulators.
And even if every provider complies, the arms race is already on. Commercial tools exist that can strip SynthID watermarks from images. Text watermark removers are almost certainly being built right now. The EU's mandate may create a cat-and-mouse dynamic where enforcement becomes a reactive game.
All of this unfolds against a backdrop of regulatory uncertainty. The Digital Omnibus amendments, which could clarify technical requirements, were still being adopted as of June 2026. Providers face a compliance cliff without a clear ladder.
So who wins? Google, for now, looks like the compliant darling. OpenAI could become a test case for enforcement. Small providers and open-source communities might struggle to implement cryptographic metadata and statistical watermarking without turning their models into identity cards. The real winners might be the watermark removal services — they'll have a thriving business if the regulation creates more detection than prevention.
Here's my prediction: expect a last-minute scramble from companies still unprepared, the first enforcement actions by late 2026, and a new sub-industry of watermark auditing. The EU's ambition to make AI text traceable is admirable, but the devil is in the tokens. And right now, the tokens are winning.