Does ChatGPT Watermark Text Now? OpenAI's textGrain Explained
Table of Contents
Does ChatGPT watermark text now?
What is textGrain?
Who gets watermarked text, and who doesn't?
How well does the watermark survive editing?
What a text watermark does not tell you
How this differs from Claude and Gemini watermarks
FAQ
Yes, in the European Union. OpenAI's text watermark is called textGrain. On October 5, 2026, the company published Our approach to EU text provenance rules, saying it would add an invisible watermark to eligible ChatGPT and Codex text "over the coming weeks," only in the EU, and that it was "not making text watermarking a global default at launch."
By early November that wording had moved into the present tense. OpenAI's help article on provenance signals, checked on November 6, 2026, says: "In the EU, ChatGPT-generated text includes an invisible change to the randomness used in the model's word choices to comply with the EU AI Act." The detector that would confirm a mark is still not public. Access, OpenAI says, stays limited to approved researchers and expert organizations.
Image and audio verification is a different system. Those tools remain publicly available at openai.com/verify. That page checks files, not pasted text. If you upload a PNG or an MP3, you are not running a textGrain check.
Does ChatGPT watermark text now?
The short answer from OpenAI's October 5 post, read against the November help article, is this:
textGrain is OpenAI's name for the text watermark.
Eligible ChatGPT and Codex text in the EU is covered. The October 5 post framed a rollout "over the coming weeks"; the help article, as of November 6, 2026, describes EU ChatGPT text in the present tense.
Outside the EU, text watermarking is not a global default at launch.
API customers anywhere can opt in for select models. Watermarking "will remain off by default in the API."
The text watermark detector is not public at launch.
OpenAI's phrase "eligible" ChatGPT and Codex output is doing quiet work. The October 5 post does not publish a full model-by-model list of what counts as eligible on ChatGPT or Codex. Treat "eligible" as OpenAI's word, not as a promise that every EU reply on every plan is marked the same way. The post does say the EU rollout covers users "across all plans," which is a different claim from "every output type."
What is textGrain?
OpenAI's definition is short. textGrain "adds an invisible statistical signal to the model's word choices." A detector "looks for that signal to assess whether a passage contains an OpenAI watermark." The mark is not a hidden character, an invisible space, or odd punctuation. It lives in which words the model picked.
The help article goes a step further. At each step the model assigns likelihoods to possible next tokens. The watermarking system builds adjusted sets of those likelihoods according to a secret key, then the model draws from the selected set. Someone with the key can later ask whether the passage follows that pattern more often than chance would allow.
OpenAI says more detail sits in a technical report linked from the October 5 page, and that it "plan[s] to make the technology available in open source." It also says textGrain "matched or exceeded the performance of other approaches we tested, including SynthID for text." That is OpenAI's comparison under its own tests, not an independent benchmark.
On quality, OpenAI published watermarked and unwatermarked Astra benchmark pairs and said it does "not see meaningful performance differences with and without watermarking." The help article adds that prior ChatGPT tests showed no change in thumbs-down rates or other product measures from introducing watermarking, and that the speed cost is negligible.
Who gets watermarked text, and who doesn't?
| Channel | Region | Default |
|---|---|---|
| ChatGPT and Codex text | EU only | On for eligible output (Oct 5: rolling out; help article Nov 6: present tense) |
| ChatGPT and Codex text | Outside the EU | Not a global default at launch |
| API (select models) | Global | Off by default; customers can opt in |
| Image / audio verify tools | Global | Stay publicly available at openai.com/verify |
API customers turn watermarking on in project or organization settings under Text provenance, according to the help article. Enabling the watermark does not grant detector access. Those are separate programs.
One more limit from the help center matters for schools: OpenAI says academic-integrity applications, such as checking student essays for AI use, are currently outside the text-detector program's scope. Researchers studying provenance can apply. Classroom policing is not what the detector is for, at least for now.
Cloud partners are still catching up. OpenAI says it is working with cloud and distribution providers so provenance signals land on eligible outputs there too, and that availability may vary by partner. If you get ChatGPT-class models through a third-party host, do not assume the EU ChatGPT app rules apply unchanged.
How well does the watermark survive editing?
OpenAI's own evaluations, dated with the October 5, 2026 post, are the numbers to use. They are lab results under OpenAI's conditions, not a field study and not a remover guide.
At a 1% false-positive target, detection was about 80% on 200-token passages and about 95% on 400-token passages for psychology-style content.
Rates were "substantially lower" for mathematics, where word choice is less flexible.
On 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%. Replacing 25% cut it to 17%.
The help article adds that short passages and code are harder to mark, and that the EU Code of Practice does not require watermarks in outputs shorter than 200 tokens (about 150 words in English) or in code snippets. Detection rates also vary across the 24 official EU languages; OpenAI publishes language-by-language figures on the help page.
None of that is an instruction for stripping a watermark. It is OpenAI saying the signal weakens when the wording changes a lot, which is exactly why the company is keeping the detector off the public internet for now. Given the risk of missed watermarks and false positives, OpenAI says, it is "not making it publicly available at launch."
What a text watermark does not tell you
OpenAI lists five limits. They are worth keeping as a set, because a detection result is easy to overread.
1. It does not measure human contribution. A mark can indicate that an OpenAI system generated or processed part of a passage. It cannot say how much judgment, editing, or drafting a person put in.
2. It does not establish ownership or responsibility. It does not settle who owns the text, whether use was lawful, whether disclosure was required, or who is accountable for it.
3. It does not identify the user. It does not associate a person, organization, account, prompt, or conversation with the text.
4. It does not verify accuracy. A watermarked passage can still be wrong, misleading, or out of context.
5. Absence of a watermark does not prove human authorship. Text may be too short, edited, translated, from an unsupported model, older than the rollout, or generated by another company's tools.
That last point is the answer people actually need. "No watermark found" is not the same as "a person wrote this." OpenAI says so in plain language, and the help article repeats that a missing signal can mean the watermark degraded, the export path was unsupported, or another model wrote the text.
How this differs from Claude and Gemini watermarks
Claude's text watermark is a separate system. Anthropic applies it more broadly across models and surfaces, and detection also sits behind limited access rather than a public checker. Our November 2 post, Claude Text Watermark Comes to Opus 4.7, Opus 4.6 and Sonnet 4.6 on November 3, covers which Claude models carry the mark and what Anthropic says about editing. This piece does not retread those mechanics.
Google's SynthID for text is the comparison OpenAI names in its own evaluations. DeepMind's public SynthID Detector is built around image, video, and audio. Do not treat that public tool as a paste-in text checker for Gemini or ChatGPT.
If you are comparing detector scores across tools this fall, remember the models behind those scores moved too. Our post on fall 2026 AI detector model changes explains why a before-and-after number is often two instruments, not one.
FAQ
Does ChatGPT watermark text now?
In the EU, OpenAI's help article (checked November 6, 2026) says ChatGPT-generated text includes textGrain. The October 5 announcement said the EU ChatGPT and Codex rollout would land over the coming weeks from that date, and that it is not a global default at launch. API watermarking is opt-in and off by default.
What is textGrain?
OpenAI's name for its text watermark. It embeds a statistical signal in word choices that a detector can look for.
Can the public check for a ChatGPT text watermark?
Not at launch. OpenAI says detector access is limited to approved researchers and expert organizations. openai.com/verify checks images and audio, not text.
If a watermark isn't found, is the text human?
No. OpenAI says absence of a detected watermark does not prove human authorship.
Does editing remove the watermark?
Editing weakens it. In OpenAI's evaluation, replacing 10% of words cut detection from about 92% to 66%; replacing 25% cut it to 17%. That is OpenAI's figure under its test conditions, not a remover guarantee.