ChatGPT text watermarking (textGrain) review: honest scoring on the invisible mark, editing robustness, EU-only scope, detector access, and what it means.
As a transparency measure, OpenAI’s textGrain is a sensible, standards-driven move: invisible, copy-paste-durable, honest about its limits, and offered to API developers as an opt-in. As a reliable "is this AI" test it is deliberately not that — it is EU-first for ChatGPT and Codex, the detector is gated to approved researchers, and the mark erodes fast under editing. Good provenance intent; a poor lie detector for a working creator.
On October 5, 2026, OpenAI said it will start adding an invisible watermark to text generated by ChatGPT and Codex, beginning with users in the European Union, and publish a technical report for the method it calls textGrain. API developers worldwide can opt in for select models the same day; it is off by default. This review scores that watermarking system — not the writing quality of the model behind it.
I run a competing content engine, so here is the bias disclosure up front, and because of it I will credit what textGrain does well and flag only limits that genuinely exist. The honest headline: this is a thoughtful, restrained transparency feature built to answer the EU AI Act, and the most useful thing a creator can do is understand exactly what it can and cannot tell you before building anything on top of it.
Three facts shape the verdict. First, the mark is invisible and leaves a response readable — as a user experience it is close to free. Second, it is scoped: EU-first for the consumer apps, global only as an opt-in API toggle, with the detector limited to approved researchers at launch. Third, it is fragile by OpenAI’s own numbers — replacing roughly 10% of words cut detection from about 92% to 66%. Everything below is scored against the system’s described state as of 2026-10-05; confirm the detection tooling and coverage on OpenAI’s own pages before relying on it.
ChatGPT text watermarking, branded textGrain, is OpenAI’s system for marking text its models generate. It does not insert a visible symbol; it subtly shapes the model’s word choices as it writes, using a secret key to sort next-word predictions across many small nudges, leaving a statistical pattern a detector can recognize. Because the signal lives in the words rather than in file metadata, it survives copy-paste. The technical report was co-authored with researchers from the University of Pennsylvania and Yale, and OpenAI says it plans to release textGrain as open source. The marking reaches eligible ChatGPT and Codex users on all plans in the EU over the weeks after launch, and is available as an opt-in setting to API developers anywhere (off by default). OpenAI opened applications for its text-watermark detector, initially limited to approved researchers and expert organizations. It is a transparency and provenance feature, not a content tool: it does not generate, format, or publish anything — it labels what the model already wrote.
The clearest fit is anyone who needs to demonstrate AI transparency — teams meeting the EU AI Act’s Article 50 obligations, platforms building trust signals, and developers who want to opt their own API output into a machine-readable mark. For an individual creator, it is mostly a background fact for now: if you are in the EU, your ChatGPT and Codex text will carry a signal, and the useful response is a consistent disclosure habit rather than a detector you operate — because you cannot operate it yet. Where it fits poorly is as a lie detector. The gated detector, the EU-first scope, and the sharp decay under editing mean it will disappoint anyone hoping to reliably prove a given passage is or is not AI-written.
| Dimension | Score | Why |
|---|---|---|
| Transparency intent | 4.3 / 5 | A responsible, standards-driven move with a published technical report and a stated plan to open-source the method. |
| User-experience impact | 4.5 / 5 | The mark is invisible and leaves the response readable — as a UX cost it is close to zero for the reader. |
| Scope & coverage | 3.3 / 5 | EU-first for ChatGPT and Codex consumer apps; global only as an opt-in API toggle, which leaves most output worldwide unmarked for now. |
| Regulatory fit (EU AI Act) | 4.3 / 5 | Directly answers the Article 50 transparency rules that took effect August 2, 2026. |
| Developer control (API) | 4.0 / 5 | Offered to API developers as a clean opt-in for select models, so builders can add provenance where they need it. |
| Robustness to editing | 2.8 / 5 | Replacing ~10% of words dropped detection from ~92% to 66% and 25% to 17% — a light rewrite erodes the signal. |
| Detection access & reliability | 2.8 / 5 | The detector is limited to approved researchers at launch, with false positives and negatives on short, math, or translated text. |
| Clarity of what the mark means | 3.6 / 5 | OpenAI is clear that the mark does not prove authorship, ownership, or accuracy — correct but easy to misread. |
There is no separate price. Watermarking is built into the covered models: for EU ChatGPT and Codex users it applies on all plans, from free through paid, and on the API it is a no-cost opt-in for select models. You do not buy it, and OpenAI has said it will publish the method openly. On cost alone, a free, invisible transparency layer is well priced by definition.
The real "cost" is a workflow one, not a dollar one. For EU creators the mark is free and silent, but it means provenance is now part of your content whether you plan for it or not, which argues for a deliberate disclosure habit. For developers, the opt-in is genuinely useful — you can stamp your own API output for compliance at no incremental charge.
The honest read: as a no-cost, low-friction transparency signal, it is priced fairly. Just do not mistake "free and built in" for "a reliable detector you control" — the limits here are in the scope, the robustness, and the gated access, not the billing.
| Use case | Fit | Why |
|---|---|---|
| Meeting EU AI Act Article 50 transparency obligations | Strong | This is exactly what the system was built for, and the EU-first rollout is aimed squarely at it. |
| Adding transparency without hurting the reading experience | Strong | The mark is invisible and leaves meaning and readability untouched. |
| A developer opting their own API output into a provenance mark | Strong | The API toggle for select models makes this a clean, deliberate choice. |
| Catching lightly copied-and-pasted AI text | OK | The signal survives copy-paste and light edits, so unmodified text usually stays detectable. |
| A creator self-checking whether their own text is marked | Weak | The detector is limited to approved researchers and experts at launch, so you cannot run it yourself yet. |
| Detecting heavily rewritten or translated AI text | Weak | A ~25% synonym swap cut detection to 17%, and translated or short text is harder still. |
| Proving a piece of text is human-written | Weak | OpenAI states an absent mark is not evidence of human authorship, and the detector yields false results. |
| A standalone "is this AI" gate for moderation outside the EU | Weak | The consumer mark is EU-first and the API mark is off by default, so most output simply is not marked. |
If you are weighing textGrain as a creator, the useful frame is that a provenance signal and a publishing workflow are different things. The watermark labels what a model wrote; it does not turn that draft into finished, on-brand posts across platforms. That is [Kompozy](/)'s job — and Kompozy is not a detector or an evasion tool. It generates copy with OpenAI alongside Claude, so text produced through a watermark-enabled OpenAI model can carry the same signal; the point is what happens after generation, not stripping the mark.
Where the two intersect is disclosure. Kompozy fans one source into 18 formats — persona video, clipped shorts, carousels, quote cards, blogs, newsletters — and its per-post review gate is where you set a consistent AI-disclosure line that then rides every asset, so provenance is handled once instead of per tool. One scope note worth getting right: textGrain is a text watermark, so it does not attach to Kompozy’s images, which come from gpt-image and Gemini composited through HyperFrames. Read together, the watermark and a publishing engine are complementary — one marks that AI was involved, the other makes AI-assisted content distinctive, disclosed, and actually published.
Only partially. The textGrain signal survives copy-paste and light edits, but erodes quickly under rewriting — OpenAI says replacing about 10% of words cut detection from ~92% to 66%, and 25% to 17% — and short, math, or translated text is harder to detect. An absent mark is not proof a human wrote something, so it should never be the sole basis for a moderation or academic-integrity decision.
That an OpenAI model with watermarking enabled generated the text — not who authored it, who owns it, or whether it is accurate. OpenAI is explicit that the mark does not measure human contribution and that a missing mark is not evidence of human authorship. Read it as "a watermarked model wrote this," not "no human was involved."
Not yet. For ChatGPT and Codex the mark is being added for EU users first, on all plans, over the weeks after the October 5, 2026 announcement. Developers using OpenAI’s API can opt in for select models anywhere in the world, but it is off by default, so most output outside the EU carries no mark unless a developer turns it on.
Not at launch. OpenAI opened applications for its text-watermark detector but is initially limiting access to approved researchers and expert organizations while it evaluates and improves the technology. The practical takeaway for creators is to plan a disclosure habit rather than to try to audit the mark.
It can, on the text side. Kompozy generates copy with OpenAI and Claude, so text produced through a watermark-enabled OpenAI model can carry the signal. textGrain is a text watermark, so it does not attach to Kompozy’s images, which come from gpt-image and Gemini via HyperFrames. The practical move is consistent disclosure, not trying to remove marks.
The EU AI Act’s Article 50 transparency rules took effect on August 2, 2026 and require providers of generative AI to mark machine-generated content so other systems can identify it. textGrain is OpenAI’s answer, launched EU-first for the consumer apps with a global opt-in for API developers.
No. It is built into the covered models — free for EU ChatGPT and Codex users on all plans, and a no-cost opt-in for select API models. OpenAI has also said it plans to release textGrain as open source.
Editing weakens it: a roughly 10% synonym swap dropped detection to 66% and a 25% swap to 17%, and translation or trimming to a short passage degrades it further. But because the mark certifies that a model wrote the text and is applied for transparency, the durable strategy is clean disclosure and distinctive work, not evasion.
See ChatGPT Text Watermarking (textGrain) vs Kompozy comparison → · Get Started →