AI · Oct 21, 2026
OpenAI on Google Cloud lets companies spend Google commitments on GPT-6, and that changes procurementOpenAI’s EU provenance plan publishes where its textGrain detector breaks down
· Updated · Signal Desk newsroom
Analysis of reporting by OpenAI. How we work

OpenAI detailed how textGrain will meet the EU AI Act’s text-marking rule, including failure data: at a 1% false-positive target, detection reached about 80% on 200-token passages and fell to 17% when a quarter of words were swapped. It says those limits are why the detector is not public.
Key takeaways
- textGrain adds an invisible statistical signal to word choices; a keyed detector looks for it.
- At a 1% false-positive target: about 80% detection on 200 tokens, about 95% on 400 tokens for prose like psychology.
- Mathematics, with constrained word choice, performed substantially worse.
- 10% synonym swaps cut detection from about 92% to 66%; 25% swaps cut it to 17%.
- OpenAI says Astra benchmarks showed no meaningful difference with watermarking and pledges an open-source release.
What OpenAI laid out
OpenAI described how it will meet the EU AI Act’s requirement that generated text be machine-identifiable. Its textGrain system adds an invisible statistical signal to a model’s word choices, and a detector looks for that signal. Watermarks are coming to eligible ChatGPT and Codex output in the EU over the coming weeks, API customers worldwide can opt in, and detector access opens only to approved researchers and expert organizations (rollout coverage).
For images and audio, OpenAI already uses Content Credentials, is C2PA conformant, embeds SynthID watermarks and offers public verification through openai.com/verify and a Content Provenance API. It describes its approach as layered because no single technique suffices, and says it will revisit text provenance as technology, standards and regulation evolve.
The failure data
The post stands out for publishing where the system fails. At a target false-positive rate of 1%, OpenAI says its detector caught watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology, and did substantially worse on mathematics, where word choice is constrained. Replacing 10% of words with synonyms dropped detection from about 92% to 66%; replacing 25% dropped it to 17%.
Labs rarely publish the conditions under which safety features stop working. OpenAI draws the appropriate conclusion that these limits are why the detector is not going to the public. A tool that loses about a third of its hits after light synonym swaps, misses most text after heavier edits and accepts some false alarms is a research instrument rather than a lie detector. The figures also indicate who will be caught: long, unedited prose from a supported model is detectable, while short answers, math, translations, paraphrased drafts and text run through another company’s tool mostly are not. The least sophisticated users are most likely to be flagged, which is a fairness problem if institutions treat detection as enforcement.
What a watermark cannot say
OpenAI says a watermark does not identify the user, measure human contribution, establish ownership or verify accuracy, and that no detection does not prove a human wrote something. It also cannot say whether using AI was appropriate, a judgment that remains with people. OpenAI says it is exploring how to distinguish AI assistance from AI authorship.
OpenAI reports no meaningful benchmark differences on Astra with and without watermarking, which, if it holds, removes a major practical objection. The pledge to release textGrain as open source matters because provenance works best when shared across companies. The likely future is a background signal platforms and researchers use to study synthetic text in aggregate; the risk is a statistical hint being used as proof against an individual. More in our OpenAI coverage, regulation coverage and research coverage.
Who feels it
- Educators and employers
- Light editing defeats detection, so results should not stand alone as evidence.
- Developers
- API watermarking is opt-in, making EU compliance an explicit product decision.
- Provenance researchers
- Published error rates and detector access give them something to test.
- Other labs
- A detailed benchmark sets a disclosure standard rivals may be asked to match.
What to watch
- Updates to the technical report and the open-source release.
- Independent replication of detection and false-positive figures.
- Conditions for expanding detector access.
- Progress on separating AI assistance from authorship.
FAQ
How does textGrain work?
It adds an invisible statistical signal to word choices that a keyed detector can find.
How reliable is detection?
About 80% on 200-token and 95% on 400-token prose passages at a 1% false-positive target, worse on math.
How easily is it defeated?
Swapping 25% of words for synonyms dropped detection to 17%, OpenAI says.
Will textGrain be open source?
OpenAI has pledged to release it as open source.
Companies: OpenAI
More on OpenAI
AI · Oct 14, 2026
White Circle says newer OpenAI and Google image models are easier to jailbreak than their predecessorsAI · Oct 19, 2026
Microsoft ships Guardian Agents, saying frontier models ‘can’t be trusted’AI · Oct 13, 2026
After months off the frontier, Google ships Gemini 4 Argon with benchmark claims and aggressive pricingAI · Sep 22, 2026
Opus 5.5 hands risky cyber prompts to an older Claude model, and that routing is the real launch story