OpenAI Starts Watermarking ChatGPT Text in the EU Under AI Act Rules
OpenAI announced textGrain, an invisible statistical watermark added to eligible ChatGPT and Codex text output in the EU to comply with AI Act provenance rules. API customers worldwide can opt in, and detection access starts with researchers. OpenAI says the method matched or exceeded Google's SynthID-Text in its tests but cautions that paraphrasing can degrade detection.
On this page
What textGrain actually does
OpenAI announced on 5 October that eligible ChatGPT and Codex text output in the EU now carries textGrain, an invisible statistical watermark embedded in the model's word choices, with a separate detector that looks for the signal. The technical report describes an entropy-calibrated design: the watermark is applied where the model's sampling choices leave room for it, so the signal concentrates in low-certainty positions and stays out of the way when the next word is essentially forced.
The scope is narrower than the word watermarking suggests. Only EU-bound consumer output is covered by default, because the trigger is regulatory; API customers anywhere in the world can switch it on voluntarily, and detection is not public. OpenAI says access to the detector starts with researchers, which keeps the tool out of the hands of platforms that might otherwise use it to ban AI-written homework at scale, and out of the hands of attackers probing how the signal behaves.
The Article 50 clock, and who moved first
The forcing function is Article 50 of the EU AI Act, which requires providers of generative systems to mark synthetic output in a machine-readable way. Models launched in Europe on or after 2 August 2026 had to comply immediately; older models carry a grace period to 2 December. Anthropic moved first, making Claude's watermarking mandatory for all users in August, and Google had already embedded SynthID signals in Gemini text and opened SynthID-Text detection.
OpenAI's own history here is worth remembering: the company built a text watermarking prototype years ago and withheld it, reportedly over worries about affected users and circumvention. Regulation closed that debate by making the question one of compliance rather than preference, and OpenAI's post says textGrain matched or exceeded the alternatives it tested, including Google's SynthID-Text.
Detection is a research privilege, for now
The honest part of the announcement is the caveat: OpenAI cautions that laboratory performance does not guarantee reliable detection in the wild, because paraphrasing, translation, and light editing can degrade the statistical signal, and third-party tools that strip provenance marks already exist for other vendors' outputs. A watermark that survives casual rewriting is a research problem, not a solved one, which is presumably why the detector stays with researchers while the marking ships to everyone.
For people who use ChatGPT in Europe, nothing about their output changes visually or contractually; the change is that their text now carries a statistical trace a detector can look for. For the ecosystem, the shift is that the three largest frontier providers now mark EU text by default, and the open questions move to the detector: who gets access, under what governance, and what happens the first time a watermark detector's verdict matters in a court or a classroom.