ChatGPT in the EU is about to carry an invisible watermark. On October 5, OpenAI announced textGrain, a system that embeds a hidden statistical signature into ChatGPT and Codex text for users in the European Union — rolling out over the coming weeks. The reason is regulation: the EU AI Act's transparency rules (Article 50(2), in force since August 2) require AI-generated content to be identifiable in a machine-readable way. OpenAI is the latest to comply — Anthropic shipped its own text watermarking in August, and Google DeepMind has had one for years.
How it works
textGrain subtly nudges the model's word choices against a secret key, weaving a statistical pattern into the text itself. There's nothing visible — no hidden characters, no odd spacing, no metadata trickery. A detector can then test a passage for that pattern. Because the signal lives in the words, it survives copy-paste, and OpenAI says it doesn't change answer quality or readability in any meaningful way. It also doesn't identify who generated the text — it's a content signature, not a user fingerprint.
The honest limits
OpenAI's own tests show the watermark bends fast under editing: swapping just 10% of words for synonyms dropped detection from about 92% to 66%, and replacing a quarter of the words collapsed it to 17%. Short texts, math answers, and translated text are harder to watermark in the first place. That's why the detector isn't public — OpenAI is giving access only to approved researchers and expert organizations for now, and it cautions that a missing watermark "does not prove human authorship." In other words: this is a compliance tool, not a cheat-proof plagiarism detector.
EU-only for now
The default stops at the EU border — OpenAI is explicitly not making this global at launch, and the US has no equivalent federal rule. API developers worldwide can opt in from October 5, but it's off by default there.
AiPost's take
The interesting part isn't the technology — statistical watermarking is old research. It's that regulation is now what ships features. The EU AI Act is forcing every major lab to build provenance into their products, and the watermark race between OpenAI, Anthropic, and DeepMind is a direct product of that law. But a watermark that a paraphrase pass defeats is a speed bump, not a wall. Treat it as what OpenAI says it is: a transparency label, not proof.
Comments
Post a Comment