ChatGPT's Secret Text Watermark Starts Rolling Out in Europe

NEW YORK — The words coming out of ChatGPT are about to carry a hidden signature, at least in Europe. On October 5, 2026, OpenAI began rolling out an invisible text watermark on writing produced by ChatGPT and Codex, embedding a statistical signal into the models' word choices that readers will never see but detection software can find. The change reaches users in the European Union on every plan over the coming weeks, while developers anywhere in the world can switch on the same text watermark through OpenAI's API, according to PYMNTS.

The technology behind the move is called textGrain. Instead of stamping a visible label onto generated writing, it subtly biases which words the model picks, creating a mathematical pattern across a passage that a detector can later spot. It works on the same basic principle as SynthID, the watermarking system from Google that OpenAI already uses for images and audio. Internal testing shows the text watermark holding up well against other detection methods, and OpenAI plans to open source the underlying technology, as reported by AI Daily Post.

Why Brussels made OpenAI add the text watermark

The rollout is a direct answer to the European Union's AI Act. Since August 2, 2026, the law's transparency provisions have required companies whose systems generate text, images, audio, or video to mark their output in a machine-readable format so it can be recognized as artificially generated. The rules also place a duty on deployers: anyone publishing AI-generated text to inform the public on matters of public interest must disclose that fact, unless the content went through human review or editorial control.

Brussels has been tightening the screws all year. The European Commission published detailed guidelines for the transparency obligations in July 2026, and a voluntary Code of Practice on transparency of AI-generated content had gathered roughly 190 signatories by the end of that month, including Anthropic, Google, Meta, Microsoft, and OpenAI. A limited grace period covers models placed on the market before August 2, but those systems must comply by December 2, 2026, which makes OpenAI's October move the first big compliance step from a major lab. The stakes of ignoring the rules are steep: penalties for the transparency provisions reach up to 15 million euros or 3 percent of a company's global annual turnover.

Who gets the text watermark and who doesn't

The rollout comes in three tiers. API customers worldwide can enable the text watermark on select models starting immediately, but the setting stays off unless they turn it on. ChatGPT and Codex users inside the European Union get the text watermark automatically, regardless of which plan they pay for, over the coming weeks. And the detection tool itself is not going public: OpenAI is opening applications so approved researchers and expert organizations can use it under the EU's Code of Practice.

Notably, OpenAI is not making watermarking the global default. Users outside the EU will see no change in their ChatGPT or Codex text under this plan, and developers must actively opt in. The feature will also flow through cloud partners such as Microsoft Azure in the coming weeks. The contrast with rivals is sharp: Anthropic watermarks the output of every Claude model everywhere with no way to turn it off, and Google applies its SynthID system across its own products.

Why the text watermark probably won't catch everyone

For all the engineering, OpenAI is unusually frank about what the text watermark cannot do. Detection can fail on short passages, heavily constrained writing, translated text, or anything that has been substantially edited after generation. The company calls text watermarking an early technology with significant limitations, and a positive detection says only that an OpenAI system likely generated or processed part of a text — not who prompted it, how much a person contributed, whether the content is accurate, or who owns it. And the absence of a watermark proves nothing at all about whether a person wrote a passage.

Skeptics go further. According to New Scientist, experts consider text watermarking fundamentally unreliable, because text carries far less data than images or video to hide a signal in. Simple tricks — paraphrasing, translation, or just trimming a passage — can strip the pattern before a detector ever sees it. That means the technology is unlikely to settle the most emotionally charged use case, academic cheating, even as schools and universities face mounting pressure to identify machine-written work.

What the hidden signal means for students and creators

For students, the most immediate effect may be psychological. A generation that grew up treating AI detectors with suspicion now faces a text watermark backed by law rather than a startup's black box — but the same limits apply. Teachers in the EU may soon have a tool that can flag machine-assisted writing, yet any accusation built on a watermark still rests on a signal OpenAI itself admits can miss and can mislead.

For creators and job seekers, the implications run in the other direction. As platforms learn to detect machine-generated text, AI-polished cover letters, captions, and product descriptions could start tripping automated filters, whether or not the underlying work is genuine. And for publishers, the EU's deployer rules add a compliance layer: text meant to inform the public on matters of public interest must be disclosed as AI-generated when it hasn't passed through human review.

The bigger picture is that provenance is becoming infrastructure. OpenAI already marks its images and audio; text was the missing piece, and regulators moved first. Over the next two months, every major AI provider must meet the December 2 compliance deadline, which means more invisible marks, more detectors, and more arguments about what they prove. The invisible ink is here — the question is whether anyone will trust what it says.

On the AI News beat, this is part of a wider push to trace AI-generated media: see our earlier look at the RAF Fairford AI photo hoax, where an invisible SynthID watermark helped fact-checkers trace a viral fake image back to AI software.