OpenAI Starts Adding Invisible textGrain Watermarks to ChatGPT and Codex Text in the EU

OpenAI has started adding an invisible watermark to text produced by ChatGPT and Codex for users in the European Union. The signal is hidden in word choice, so readers will see nothing different, but approved researchers will be able to test whether a passage is likely to have come from an OpenAI model. The company announced the change on October 5, 2026 (BleepingComputer, The Next Web).

What OpenAI is rolling out

According to OpenAI's post, the technique, called textGrain, is a statistical signal that nudges which words the model picks. It is not a visible mark and adds no hidden characters. TechCrunch reports that the change responds to transparency rules in the EU AI Act, which require providers of generative AI to mark generated content in a machine-readable, detectable way.

The watermark will reach eligible ChatGPT and Codex users in the EU on all plans over the coming weeks. Developers using the API anywhere can opt in now for selected models, but the setting is off by default. Detector access is limited to approved researchers and expert organisations, so ordinary users cannot paste a passage into a public checker. OpenAI says those limits reflect its wish to have outside experts "evaluate reliability and responsible uses" first.

How well it works

Test (OpenAI's evaluation)	Detection rate
200-token English psychology answers (ELI5), 1% false-positive target	About 80%
400-token answers, same test	About 95%
400-token passages, unedited baseline	About 92%
400-token passages, 10% of words swapped for synonyms	About 66%
400-token passages, 25% of words swapped	About 17%

A token is a small chunk of text, usually a word or part of one, so these are passages of a paragraph or two. Maths answers and other tightly constrained outputs are harder to mark because there is little freedom in the wording (Cybernews). OpenAI also says "the absence of a detected watermark does not prove human authorship."

Those numbers matter for anyone hoping to use watermarks as proof. A lightly edited essay can pass as unmarked, while a long, untouched answer is likely to be flagged. The system is a statistical clue, not a verdict.

Privacy and quality claims

OpenAI says the watermark does not identify the person who wrote the prompt, and it reports no meaningful change in model quality. Those are OpenAI's claims, and because the detector is restricted, independent researchers who are not approved will find them harder to check.

The watermark sits alongside other provenance tools OpenAI points to, including Content Credentials, Google's SynthID and a verification page at openai.com/verify, which cover images and other media rather than text.

The wider context

TechCrunch says Anthropic announced similar watermarking for Claude in August 2026 and that it drew backlash from users; that detail rests on one outlet. If accurate, it suggests the EU rules are pushing more than one company in the same direction. Users outside the EU will not get the consumer watermark for now, while the API option lets companies elsewhere switch it on.

What it means for students, teachers and developers

For students and teachers, a watermark is not a reliable accusation. A student whose own writing is flagged cannot see the evidence, and one who heavily edits an AI draft may remove the signal, so any school relying on such a tool needs a human process around it.

For developers, the opt-in offers a way to mark generated text in their own products, for example automated customer replies. Indian users are not part of the consumer rollout; Indian developers can opt in through the API for selected models.

What to watch next

Watch for the exact date EU users begin receiving watermarked text, which researchers are approved to use the detector, independent test results, and whether regulators accept statistical watermarks as meeting the AI Act's requirement.

Related on Gadgets365

• OpenAI brings GPT-6 to ChatGPT with Intelligent UI
• FTC opens its first US probe into AI agent risks
• OpenAI safety lead David Robinson resigns