An 80 percent success rate on short text is what OpenAI is calling compliance. That number matters, because watermarking is about to become mandatory for every major AI company operating in Europe.
OpenAI has announced plans to add invisible watermarks to text and code generated within the European Union, using a proprietary system it calls textGrain. The system works by embedding a statistical signal into the model’s word choices during generation. A separate detector can then read those signals and confirm whether a watermark is present. The move follows a similar rollout by Anthropic earlier this year, and both companies are responding to the same deadline: Article 50 of the EU’s AI Act, which requires providers of generative AI to make text outputs machine-identifiable. Pre-existing companies have until December 2 to comply.
OpenAI claims textGrain “matched or exceeded” the performance of comparable approaches, including Google DeepMind’s SynthID for text. But the company also acknowledges that strong benchmark performance doesn’t mean the system works reliably in real-world conditions. Short text is harder to watermark accurately, hitting around 80 percent detection. Mathematical content also underperforms. And editing watermarked text can degrade the signal further. So the system is functional, but not airtight.
The watermark will be on by default only for eligible ChatGPT and Codex outputs in the EU. Outside that region, customers can opt in for select models, though OpenAI hasn’t specified which models will be available globally. Access to the detection tool is being kept tight for now, limited to approved researchers and expert organizations. Importantly, the detector doesn’t expose user identity, prompts, or conversation history.
This matters beyond just OpenAI. The December 2 deadline applies across the board, meaning Microsoft, Google, Meta, and every other company with a generative AI product in the EU market is on the clock. Some, like Google with SynthID, already have watermarking infrastructure in place. Others are still figuring it out. OpenAI’s plan to open-source textGrain is notable here. It could become a shared standard rather than just a compliance checkbox, though that depends on how widely it gets adopted and whether the accuracy issues get resolved.
Watermarking text is genuinely hard. Unlike images, where tools like C2PA and SynthID have made real progress, language is fluid. People paraphrase, edit, and reformat AI output constantly. Any system that degrades under light editing has a real adoption problem in practice. OpenAI knows this, which is probably why it’s being cautious about who gets access to the detector. Still, having something is better than nothing, and the regulatory clock isn’t stopping for anyone.
