What Is textGrain?
OpenAI’s invisible watermark for ChatGPT and Codex text in the EU — and what you can do about it
On October 5, 2026, OpenAI announced textGrain: a statistical watermark woven into the word choices of ChatGPT and Codex output generated in the European Union, rolled out to every plan over the following weeks. It is not a hidden character, not a tracking token, and not a style score — it is a keyed pattern only OpenAI’s detector can verify. This page explains how textGrain works, who it affects, and how to handle text that carries it: Quick Clean strips the invisible characters ChatGPT text always carries (free, in your browser); Deep Rewrite restructures wording to weaken the statistical signal (best-effort).
Clean characters ≠ clean text — the textGrain watermark hides in .
Line breaks, spaces, and tabs are formatting, not watermarks — we never count or remove them.
Default. Strong change, readable output.
Free to try — no sign-up needed.
Deep Rewrite supports up to 8,000 characters
Quick Clean removes invisible characters — free & unlimited
Deep Rewrite rewrites text to break statistical watermarks (textGrain, EU ChatGPT output)
One quick check before we spend your credit
See it in action
Try a sample to see how Deep Rewrite changes wording and AI-pattern signals — before using your own text.
Example of how wording can change. Deep Rewrite is best-effort and results vary.
Our own heuristic scan — not a vendor detector.
Characters removed ✓ — the textGrain is still there.
Show where the hidden characters were
Rewriting with Deep Rewrite… usually ~10s
Higher change = more of the watermark signal diluted.
AI style lives in sentence rhythm — switch to and rewrite again.
How textGrain Works
A language model writes one token at a time. At each step it holds a list of plausible next words, each with a probability, and samples one. textGrain leaves that list alone but changes how the model picks from it: a secret key, combined with the preceding text, produces pseudorandom numbers that nudge the model toward the candidates those numbers favour. Each single choice still looks completely natural — OpenAI reports no meaningful difference in output quality across its benchmarks — but across a few hundred words the nudges accumulate into a statistical signature. A detector holding the same key and settings can test whether the expected pattern is present, and report how confident it is.
Two properties follow. First, there is nothing to delete: the watermark is the wording itself, so no character-level cleaner can see it or touch it. Second, detection is length-dependent and edit-sensitive. In OpenAI’s own evaluation at a 1% false-positive target, the detector found textGrain in about 80% of 200-token psychology passages and about 95% of 400-token ones — and substantially fewer for constrained content like mathematics, where word choice has less freedom. Editing weakened the signal quickly:
| Edit applied to a 400-token watermarked passage | Detection rate (1% false-positive target) |
|---|---|
| None — original output | ~92% |
| 10% of words replaced with synonyms | ~66% |
| 25% of words replaced with synonyms | ~17% |
OpenAI is equally clear about what a textGrain result does not tell you. A detection does not measure human contribution, establish ownership, identify the user, or verify accuracy — and the absence of a detected watermark does not prove human authorship, because short, edited, translated, or pre-watermark text may carry no detectable signal. That is why OpenAI limits detector access to approved researchers and expert organizations rather than publishing a public checker.
Who textGrain Affects
The rollout is regional by design. If you use ChatGPT or Codex from within the European Union, the mark is being switched on for every plan over the weeks after October 5, 2026, to comply with Article 50 of the EU AI Act, which requires generated text to be identifiable in a machine-readable way. If you build with OpenAI’s API, watermarking stays off unless you opt in for supported models — a choice OpenAI offers globally, with cloud-partner availability following. And because the pattern travels with the words, text generated in the EU keeps its signature when pasted into a document in New York or Tokyo. Everyone else — US and other non-EU ChatGPT users whose text did not come through an opting-in API customer — is not watermarked at launch, and OpenAI has said the EU-first approach is deliberate, not a global default.
Can textGrain Be Removed?
Partially, and honestly only with a rewrite. Since the signal lives in word choice, changing the wording is the only lever — and OpenAI’s own numbers show how sensitive the watermark is to exactly that. Replacing a quarter of the words with synonyms cut detection from 92% to 17% in its evaluation. Our Deep Rewrite mode goes further than synonym swaps: it sends your text through an LLM that paraphrases and restructures it, which dilutes the keyed pattern across essentially every sentence. Treat the result as best-effort, not a guarantee — the official detector is not public, so no third party can verify a text is clean, and future detector improvements are unknowable.
One layer textGrain does not change: the invisible characters. ChatGPT’s chat interface has long left Unicode debris — narrow no-break spaces, zero-width characters — in copied output, and EU watermarking adds nothing at the character level. Quick Clean still fixes that layer in one click, entirely in your browser, with a verifiable zero on re-scan. The two layers are independent: a text can carry the formatting debris without textGrain, textGrain without any hidden characters, or both.
textGrain vs Claude’s Watermark vs Hidden Characters
| Hidden characters | Claude watermark | textGrain | |
|---|---|---|---|
| Vendor | Chat interface debris (any tool) | Anthropic | OpenAI |
| Where it lives | Invisible Unicode code points | Word choices (SynthID-Text-style, own key) | Word choices (own scheme) |
| Scope at launch | Wherever text is copied from | All Claude models since Aug 2, 2026 rollout | EU ChatGPT/Codex output; API opt-in |
| Public detector | Not needed — deterministic | Private preview (elig. organizations) | Approved researchers only |
| Removal layer | Quick Clean — verifiable | Deep Rewrite — best-effort | Deep Rewrite — best-effort |
Why People Look Up textGrain
EU-based writers and teams
Your ChatGPT drafts now carry a statistical mark that follows the text across borders. Deep Rewrite weakens the signal before the text travels — best-effort, with no false promises.
Publishers and editors
You received a submission and want to know what textGrain can and cannot prove. Short answer: less than headlines suggest — it signals OpenAI involvement, not authorship, quality, or intent.
Builders on the OpenAI API
Watermarking is opt-in for API output. If you turned it on for compliance reasons, this tool explains what your users’ text now carries and how editing affects detectability.
Anyone cleaning ChatGPT text
Invisible characters breaking your CSV, diff, or CMS have nothing to do with textGrain — Quick Clean still solves that in one click, free, with nothing uploaded.