Text written in ChatGPT by European users will soon carry a signal that almost nobody outside a short list of approved institutions can read. OpenAI announced on October 5 that it will add an invisible watermark, which it calls textGrain, to ChatGPT and Codex responses for eligible EU users on every plan over the coming weeks. For publishers who worry about AI-content detection, the company’s own numbers say how much that mark can prove.
The mechanism is statistical. According to OpenAI’s announcement, textGrain nudges the model’s word choices, and a detector looks for that pattern. API customers worldwide can opt in for select models starting now, though the setting stays off by default. Search Engine Journal noted that the post leaves two questions open: which API models qualify, and if EU ChatGPT users get a way to turn the mark off.
Length and subject matter change the odds. At a 1% false positive target, OpenAI reports that the detector flagged about 80% of 200-token samples and about 95% of 400-token samples on content such as psychology. Math fared much worse, because fixed answers leave little room to vary wording. Short, constrained text is the hardest case.
Edits do more damage. In a test on 400-token passages, swapping 10% of words for synonyms cut detection from about 92% to 66%. Swapping 25% cut it to 17%. Most published content passes through a human editor, and a heavy edit is exactly the situation where the signal fades.
OpenAI is also blunt about the limits of what a result means. The company says a watermark does not gauge the human share of a passage, does not establish who owns or answers for a text, does not identify the user, and does not verify accuracy. A missing watermark does not prove a human wrote the passage either, since text may be too short, edited, translated, or produced by another vendor’s tool.
That caution explains the access policy. The detector is not public. OpenAI is accepting applications from approved researchers and expert organizations, with access granted case by case, and ties that restriction to the chance of both undetected marks and wrongly flagged text. Image and audio verification tools remain open to everyone.
Search Engine Journal adds the regulatory backdrop and the competitor comparison. It reports that the transparency duties in the EU AI Act’s Article 50 took effect on August 2, 2026, with a later deadline for systems already on the market. It also reports that Anthropic already watermarks output from its supported Claude models in every region, using an adaptation of Google DeepMind’s SynthID-Text, while OpenAI confines its rollout to the EU.
None of the sources says Google reads, uses, or plans to use OpenAI’s watermark, and this article does not suggest otherwise. Google has repeatedly said it evaluates quality rather than authorship. Its recent guidance goes the other direction from detection: our earlier report covers Google’s guidance calling manual review of AI-generated content critical, a human check that no watermark replaces.
The practical read is narrow. A closed detector, a signal that weakens under ordinary editing, and a published statement that absence proves nothing make textGrain a poor basis for labeling or policing content. Anyone hoping to certify a page as machine-written or human-written will find the mark too fragile, and too hard to check, to carry that weight.
Keep fact-checking and editing as the controls your team relies on, and leave watermark detection out of agency agreements and freelancer clauses until independent access and measured error rates exist.
Based on OpenAI’s “Our approach to EU text provenance rules” (October 5, 2026) and Search Engine Journal’s reporting by Matt G. Southern (October 5, 2026).
