The ChatGPT watermark is real, but it is narrower than most people assume. On October 5, 2026, OpenAI announced textGrain, an invisible statistical watermark woven into the word choices of ChatGPT and Codex text, rolling out over several weeks to users in the European Union only. Outside the EU, ChatGPT text is not watermarked by default, there are no hidden characters to find, and the detector is not public: only approved researchers can use it.
Below: how textGrain works, what OpenAI's own tests show, how it compares with Claude and Gemini, and what you can check today.
Does ChatGPT have a watermark? The short answer
It depends on where you are and what kind of output you mean. OpenAI's help center lays out the current picture by content type (OpenAI Help Center):
| Output | Signal OpenAI uses | Who can check it |
|---|---|---|
| Text (ChatGPT, Codex) | textGrain watermark, EU only | Approved researchers and expert organizations |
| Text (OpenAI API) | textGrain, opt in per project or organization, any region | Same restricted detector |
| Images | Content Credentials plus a SynthID watermark | Anyone, via OpenAI's verify tool or Google's SynthID Detector |
| Audio | SynthID watermark | Anyone, via OpenAI's verify tool |
In words: in the EU, ChatGPT and Codex text carries a textGrain watermark once the rollout reaches your account, on every plan. Elsewhere it does not, because OpenAI chose not to make it a global default at launch (OpenAI). In the API it stays off unless a developer switches it on. Images and audio already carry watermarks that anyone can check.
Why OpenAI added it now: the EU AI Act
The trigger is Article 50 of the EU AI Act. Its transparency obligations have applied since August 2, 2026, and they require providers of generative AI to mark audio, image, video and text output in a machine readable way so that it is detectable as AI generated (European Commission).
Providers can show compliance by signing the voluntary Code of Practice on Transparency of AI Generated Content, published in final form on June 10, 2026 and signed by about 190 organizations by the end of July. OpenAI says its text watermarking follows its commitments under that code.
One detail explains a lot: the code does not require watermarks on outputs shorter than 200 tokens, roughly 150 English words, or on code snippets (OpenAI Help Center). Short replies and code are exactly where text watermarks work worst.
How the textGrain watermark works
A language model writes one token at a time, a token being a word or a piece of one. At each step it has a ranked list of plausible next tokens and picks one with a bit of randomness. Normally that randomness comes from an ordinary random number generator.
textGrain swaps the source of that randomness. According to OpenAI's help center, the system builds several adjusted versions of the model's next token probabilities, each favoring different choices according to a secret key, and balances them so that on average they match the original probabilities. The key picks which version to use, and the model then makes a fresh random choice from it. Over a whole passage, those choices form a pattern that a detector holding the same key can test for: does this text follow the secret pattern more often than chance would allow?
The technical report frames this as an optimal transport problem with Gumbel based costs and an adjustable strength setting, and notes that the detector needs only the text and the secret key (textGrain technical report).
Three consequences follow from that design:
- Nothing is added to the text. No hidden characters, invisible spaces, odd punctuation or extra tokens.
- It needs room to choose. When an answer must be exact, such as a math result, code or verbatim text, there is little room for the pattern.
- It needs length. One sentence holds too few choices for a confident test.
What OpenAI's own tests show
OpenAI published unusually candid numbers alongside the launch (OpenAI). All of them are at a target false positive rate of 1 percent, meaning the detector is tuned so that roughly one in a hundred unwatermarked passages could still be flagged:
- Length matters. For flexible content such as psychology answers, the detector found the watermark in about 80 percent of 200 token passages and about 95 percent of 400 token passages. Detection was substantially lower for mathematics, where word choice is constrained.
- Light editing hurts. On 400 token passages, swapping 10 percent of the words for synonyms cut detection from about 92 percent to 66 percent. Swapping 25 percent cut it to 17 percent.
- Language matters. Across all 24 official EU languages, detection on translated prompts ranged from 69.0 percent for Spanish down to 42.2 percent for Romanian before OpenAI raised the watermark strength for weaker languages (OpenAI Help Center).
- Quality looks unaffected. On the benchmarks OpenAI uses for its Astra model, scores with and without watermarking were within normal run to run noise, and earlier ChatGPT tests showed no change in thumbs down rates.
OpenAI says textGrain matched or exceeded the other approaches it tested, including SynthID for text. Treat that as a vendor evaluation, not an independent audit; OpenAI itself warns that ideal condition results do not guarantee reliable everyday detection.
Who can use the ChatGPT watermark detector?
Right now, almost nobody outside a research program. OpenAI opened applications on October 5 for approved researchers and expert organizations, granted case by case, and names early partners at Cornell University, ETH Zurich and INSAIT, and the Kempelen Institute of Intelligent Technologies. The tool reports only whether it finds an OpenAI watermark; it does not identify the user or reveal prompts or conversations.
The help center lists what is out of scope for now: general business uses such as fraud checks or resale, and academic integrity work such as checking student essays or theses.
What you can check today:
- Images and audio from OpenAI: the OpenAI verification tool and the free Content Provenance API look for SynthID watermarks and OpenAI's Content Credentials.
- Images, video and audio across several companies: Google's SynthID Detector, which recognizes content from Google and partners including OpenAI. Our SynthID Detector guide walks through it.
- Text in general: only classifier based tools such as Pangram, GPTZero and Turnitin, which analyze writing style rather than reading a watermark. See our picks for the best free AI detector and the research on whether AI detectors are accurate.
ChatGPT vs Claude vs Gemini text watermarks
OpenAI is not alone. The same EU rules apply to every major provider, and each has made different choices.
| Item | ChatGPT (OpenAI) | Claude (Anthropic) | Gemini (Google) |
|---|---|---|---|
| Method | textGrain, OpenAI's own | A version of SynthID-Text | SynthID-Text |
| Where applied | EU only in ChatGPT and Codex; API opt in globally | Worldwide, on supported models | Gemini output |
| Public text detector | No; researchers only | No; private preview for eligible organizations | No public text check in the SynthID portal |
Anthropic says its watermark applies wherever Claude is offered because it does not yet have a durable way to limit it by region, and that models launched before August 2, 2026 are being added during the transition period. Its detection API is in private preview for groups such as regulators, media, fact checkers, researchers and educational organizations, plus enterprises with their own compliance duties (Anthropic; Claude Help Center). Google's SynthID-Text, the basis for Claude's approach, was described in a 2024 Nature paper that tested it on live Gemini traffic (Nature).
The key point: each scheme has its own secret key. A detector for one company's watermark cannot see another's, so a negative result from any single checker says nothing about text from other tools.
What a watermark can and cannot prove
Both OpenAI and Anthropic are careful to limit what a positive result means. Pulling their statements together:
- It shows involvement, not authorship. A detected watermark is evidence that the company's model likely generated or processed part of the text. It cannot tell "the model wrote this" from "the model heavily edited this," and it does not measure how much a person contributed.
- It does not identify anyone. Neither company's watermark carries information about the user, organization, account or conversation.
- It says nothing about truth or ownership. It does not verify accuracy, decide who owns the text, or establish legal responsibility.
- No watermark is not proof of human writing. The text may be too short, edited, translated, from an older or unsupported model, from outside the EU, or from another company entirely.
That last point matters most in schools and newsrooms. A missing watermark is weak evidence. A present one is stronger, but it still answers a narrow question. For how watermarks fit alongside classifiers and provenance metadata, see what is AI detection. For marks that prove who owns a model rather than a piece of content, see AI model fingerprinting.
The hidden character myth and "watermark removers"
Search suggestions for the ChatGPT watermark are full of "remover" and "unicode detector" tools. Many of them scan for invisible Unicode characters, such as zero width spaces, and strip them out. That has nothing to do with textGrain. OpenAI and Anthropic both say their text watermarks add no hidden characters at all; the signal is the wording itself (OpenAI Help Center; Anthropic).
So "no hidden characters found" tells you nothing about whether text came from ChatGPT, and stray invisible characters often come from copying web pages or documents. Be wary of any site claiming to detect the ChatGPT watermark: the only working detector is not public.
What this means for you
- Students and writers in the EU: ChatGPT text may carry a watermark, but no teacher can check it with OpenAI's tool today.
- Teachers and editors: do not wait for a public watermark checker. Use classifier detectors carefully, alongside drafts and conversation. A 2025 ACL paper found that five frequent ChatGPT users, voting as a group, misclassified only 1 of 300 articles (arXiv).
- Developers: API watermarking is off by default; enable it under Organization settings, Data controls, Text provenance, or per project.
- Researchers and fact checkers: apply for detector access with OpenAI and Anthropic.
FAQ
Does ChatGPT have a watermark on text?
Only in the EU, and only as an invisible statistical pattern in word choices. OpenAI is rolling out textGrain to eligible ChatGPT and Codex users there on all plans. Outside the EU, ChatGPT text is not watermarked unless a developer opts in through the API.
Can I check if text has a ChatGPT watermark?
Not with a public tool. OpenAI's text watermark detector is limited to approved researchers and expert organizations, and academic integrity checks are excluded for now. Classifier detectors estimate AI writing by style, which is a different test.
Does copying and pasting remove the ChatGPT watermark?
No. The watermark is the wording itself, so it travels with copied text. Heavy paraphrasing or translation can make it undetectable; in OpenAI's tests, replacing a quarter of the words with synonyms cut detection to about 17 percent.
Does Claude have a watermark like ChatGPT?
Yes. Anthropic uses a version of Google DeepMind's SynthID-Text and applies it worldwide on supported Claude models, not only in the EU. Detection is in private preview for eligible organizations.
Does ChatGPT watermark images?
Yes. Images from ChatGPT, Codex and the OpenAI API carry both Content Credentials and a SynthID watermark, and anyone can check them with OpenAI's verification tool or Google's SynthID Detector.
Does the watermark make ChatGPT's answers worse?
OpenAI reports no meaningful difference on its benchmarks and no change in thumbs down rates in earlier ChatGPT tests. Anthropic reports the same for Claude, citing Google's earlier Gemini study.
