ChatGPT text is watermarked only in the EU for now
Yes, but only in the EU. On October 5, 2026 OpenAI said ChatGPT and Codex text for eligible EU users will carry an invisible watermark, rolled out over the following weeks across every plan. Everywhere else, ChatGPT stays unmarked unless an API customer opts in, and OpenAI says it is not making watermarking a global default at launch.
Where OpenAI's text watermark applies
Every ChatGPT and Codex plan in the EU gets the watermark
Every eligible ChatGPT and Codex user in the EU gets the mark, on free and paid plans alike. The switch is flipped gradually over a few weeks rather than on a single day.
OpenAI describes the regional start as a way to learn from real use before going further. It gave no date or condition for a wider rollout.
Outside the EU, nothing changes in ChatGPT
If you use ChatGPT from the US, the UK or anywhere else outside the EU, this announcement does not put a watermark in your text. OpenAI states plainly that watermarking is not a global default at launch.
For anyone hoping to catch AI-written work, that leaves a large gap. Most English-language ChatGPT output in the world carries no mark at all, so a detector has nothing to find in it.
API developers anywhere can opt in
API customers in any country can switch on text watermarking for selected models, and it stays off until they do. The option went live on the day of the announcement. OpenAI is also working with cloud partners so that OpenAI models served on their platforms can carry the mark in the coming weeks.
"Over the coming weeks, we will introduce text watermarking to eligible ChatGPT and Codex users across all plans in the EU only. We are not making text watermarking a global default at launch. This regional approach gives us room to learn from real-world use and feedback."（Our next steps）／"Starting today, API customers globally will be able to opt in to text watermarking for select models. Text watermarking will remain off by default in the API."（Our next steps）— OpenAI announcementView official source →
OpenAI's textGrain hides the AI watermark in word choice
textGrain changes which word the model picks next, using a secret key that only OpenAI holds. A detector with the same key checks whether that bias shows up across a passage.
Language models write one token at a time (a token is a word or a piece of a word). When several next words would fit equally well, the model normally picks among them at random. textGrain replaces that randomness with values derived from the secret key and the text that came just before, so the choices lean in a direction the key can later recognize.
Five OpenAI researchers and four academics from the University of Pennsylvania and Yale co-wrote the technical report. OpenAI plans to release textGrain as open source.
No hidden characters, so copy and paste keeps the mark
The mark lives in the words themselves, so copying the text copies the mark. OpenAI's help page says textGrain inserts no hidden characters and no watermark-only tokens, and the text looks exactly as it would without it.
Pasting the answer into Word, Google Docs or an email therefore carries the mark along. There is nothing invisible to strip out, and nothing a reader could spot by looking.
Short answers, code and math leave little room for it
The watermark needs room to choose between words, and short replies, code and exact answers have very little. A reply of a few sentences rarely gives the detector enough to go on, and in code there are usually only one or two sensible next tokens.
A math result or a passage the model was told to copy word for word leaves almost no freedom either. The EU transparency Code of Practice allows for this, exempting outputs under 200 tokens (roughly 150 English words) and code snippets from the watermark requirement.
Benchmarks show no quality cost from watermarking
Nudging word choice raises an obvious worry about answer quality. OpenAI ran eight benchmarks on Astra, its latest model, with and without the watermark, and found no meaningful difference. Five scores went up with the mark on and three went down.
Astra (max) benchmark scores without and with the watermark
Left bar in each pair is without the watermark, right bar is with it. Values are percentages, except the Artificial Analysis Intelligence Index, which is a score. Source: OpenAI announcement.
The largest gap, on Terminal-Bench Science 0.1, favored the watermarked model. In day-to-day use you are unlikely to notice any difference in the answers.
SynthID and the Claude watermark work the same way
Biasing word choice with a key is not unique to OpenAI. Google DeepMind's SynthID for text works on the same principle, and so does the watermark Anthropic described for Claude on August 14, 2026, covered in "Claude text watermark: what a hit actually proves".
OpenAI says textGrain matched or beat every other approach it tested, SynthID included. It adds that good results under ideal conditions do not guarantee reliable detection in ordinary use, and the announcement gives no head-to-head figures.
"No. Text watermarking changes the statistical pattern of word choices; it does not insert hidden characters or add watermark-only tokens."（Does text watermarking add hidden characters, extra tokens, or visible changes?）／"To account for these limitations, the EU AI Act Code of Practice on the Transparency of AI-Generated Content does not require watermarks in outputs shorter than 200 tokens—about 150 words in English—or in code snippets."（Do text watermarks apply to short answers and code?）— OpenAI Help CenterView official source →
"Across the benchmarks we use to assess Astra, our latest frontier model, we do not see meaningful performance differences with and without watermarking."（Impact of watermarking on output quality）／"In our evaluations, textGrain matched or exceeded the performance of other approaches we tested, including SynthID for text. Even so, strong performance under ideal conditions does not guarantee reliable detection in everyday use."（How our watermarking works and performs）— OpenAI announcementView official source →
ChatGPT watermark detection weakens on short or edited text
Detection is fairly strong on long, untouched English text and drops quickly once the text is short, edited or in another language. All of OpenAI's figures assume a 1% false-positive rate, meaning that one unmarked passage in a hundred is wrongly flagged as marked.
A 200-token passage is caught about 80% of the time
On loosely worded topics such as psychology, OpenAI's detector flagged roughly four in five 200-token passages, rising to about 95% when the passages doubled to 400 tokens. Even where wording is flexible, then, one short passage in five got past it.
Topics with less freedom in word choice did worse. OpenAI names mathematics as a case where detection rates were substantially lower, without giving the number.
Swapping 10% of words drops detection to 66%
Light editing does real damage to the signal. In OpenAI's test on 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing a quarter of them cut it to 17%.
Those passages were English answers to questions from the ELI5 dataset. Anyone who rewrites a ChatGPT draft in their own words is therefore likely to remove most of the mark without trying to.
Detection ranges from 69.0% in Spanish to 42.2% in Romanian
Across the 24 official EU languages, detection ran from 69.0% for Spanish down to 42.2% for Romanian at the same 1% false-positive rate. OpenAI built that test by translating 500 English prompts into the other 23 languages.
The watermark has a strength setting, and OpenAI turned it up for languages that scored below 60%. The help page shows the resulting gains only as shaded portions of a chart, so 42.2% is the lowest rate it states in words.
"Shorter or more constrained text is harder to detect. At a target false positive rate of 1%, our detector identified watermarks in about 80% of 200-token passages, compared with about 95% of 400-token passages, for content such as psychology. Detection rates were substantially lower for content such as mathematics, where there is less flexibility in word choice."／"Editing can weaken the watermark. In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced it to 17%."（How our watermarking works and performs）— OpenAI announcementView official source →
"At 1% false-positive rate, Spanish has the highest detection rate (69.0%), while Romanian has the lowest (42.2%)."（How does text-watermark detection vary by language?）— OpenAI Help CenterView official source →
AI writing detectors cannot read the ChatGPT watermark
Only OpenAI's own detector can read textGrain, and that detector is not available to the public. Commercial AI writing detectors work in a different way and have no access to the key, so a teacher or editor running one learns nothing about the watermark.
The textGrain detector stays private over false positives
OpenAI is keeping the detector private because it can miss watermarks and flag unmarked text. Approved researchers and expert organizations can apply for access, and each request is reviewed under the EU Code of Practice.
When the detector does run, it reports only whether it found an OpenAI watermark. It does not name the user, and it does not reveal the prompt or the conversation. The public openai.com/verify page checks images and audio, not text.
Watermark detection is not what AI writing detectors do
Tools such as Pangram, the kind usually meant by a ChatGPT detector, run a classifier over finished text instead of looking for a key. They look for patterns typical of AI writing, such as word choice, and give a probability.
A watermark works the other way around. The signal is placed in the text while it is generated and can only be confirmed with the matching key. The EU AI Act asks for that embedded kind of signal, which is why OpenAI built textGrain rather than relying on classifiers.
A missing watermark does not prove a human wrote it
OpenAI says directly that the absence of a detected watermark is not proof of human authorship. Text can be too short, edited or translated for detection to work, and anything written before watermarking began, by an unsupported model or with another company's tool, will not carry the mark.
"Approved researchers and expert organizations can apply starting today. In accordance with the Code of Practice, access will be initially granted on a case-by-case basis to support evaluation and improvement of text provenance. The tool will report whether it detects an OpenAI watermark, without identifying the user or revealing their prompts or conversations. Given the risk of missed watermarks and false positives, we are not making it publicly available at launch."（Our next steps）／"The absence of a detected watermark does not prove human authorship. Text generated with OpenAI tools may be too short, edited, or translated for detection to work reliably. It may also come from an unsupported model, predate watermarking, or have been generated by another company's tools."（What a text watermark doesn't tell you）— OpenAI announcementView official source →
"Third party detection tools like Pangram typically use classifiers to identify AI-generated text from patterns such as word choice. That approach analyzes the text after generation rather than detecting an embedded signal. The EU AI Act specifically requires a signal to be embedded in generated text."（How do OpenAI’s text watermarks differ from third party AI detection tools?）— OpenAI Help CenterView official source →
The EU AI Act made OpenAI ship a watermark it shelved
The watermark exists because EU law now requires generative AI output to carry a machine-readable mark. Two years earlier, OpenAI had built a text watermark and decided not to ship it.
Timeline of OpenAI's text watermark
| Date | Event |
|---|---|
| August 4, 2024 | OpenAI says it has a text watermark but holds it back |
| August 2, 2026 | Transparency rules in Article 50 of the EU AI Act start to apply |
| August 14, 2026 | Anthropic publishes how Claude's text watermark works |
| October 5, 2026 | OpenAI announces textGrain for the EU |
| December 2, 2026 | Deadline under Article 50(2) for AI systems already on the market before August 2 |
Article 50(2) applies from August 2, 2026
Since August 2, 2026, Article 50(2) of the EU AI Act has obliged any company whose AI produces text, images, audio or video to tag that output with a mark software can read, so the content can be recognized as machine-made. Products launched before then get a grace period running to December 2, 2026, according to the European Commission's AI Act Service Desk.
OpenAI says its phased rollout reflects both the legal requirements and the limits of the technology. The wider set of August rules is covered in "EU AI Act 2026: what applies in August, what waits" and enforcement in "EU AI Act Enforcement Begins: Article 50 Labels".
"Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system ／ are marked in a machine-readable format and detectable as artificially generated or manipulated."（Article 50, paragraph 2）— EU AI ActView official source →
"Transparency rules (Article 50) start to apply."（02 Aug 2026）／"Transitional deadline for certain providers of AI systems, including GPAI systems, generating synthetic content, that were already placed on the market before 02 Aug 2026, to comply with Article 50(2)."（02 Dec 2026）— European Commission AI Act Service DeskView official source →
OpenAI held back its text watermark in 2024
In an update on August 4, 2024, OpenAI said it already had a text watermarking method but was not shipping it. It gave two reasons for holding back: the mark was easy to defeat, and it could unfairly single out some users.
The method held up against light paraphrasing but not against translation, rewording by another model, or a trick that inserts and then deletes a character between every word. OpenAI also worried it could stigmatize AI as a writing aid for non-native English speakers. Translation and rewriting still strip the mark today, and OpenAI has answered by limiting the rollout to the EU and the detector to researchers.
"Our teams have developed a text watermarking method that we continue to consider as we research alternatives."／"it is less robust against globalized tampering; like using translation systems, rewording with another generative model, or asking the model to insert a special character in between every word and then deleting that character - making it trivial to circumvention by bad actors."／"For example, it could stigmatize use of AI as a useful writing tool for non-native English speakers."（Text watermarking）— OpenAIView official source →
OpenAI API customers turn on text watermarking in settings
API customers can switch the watermark on in their settings, and OpenAI then adds it during generation with no extra step per response. It can be set for a whole organization or overridden for one project.
Set it for the whole organization or override it per project
An organization-wide setting applies to every project unless a project overrides it. The switch itself is the same in both places.
- For the whole organization, open Organization settings, then Data controls, then Text provenance. For one project, open that project's Settings, then Text provenance.
- Turn on Allow text watermarking.
- Select the models that should carry the mark, then select Save.
The settings page lists the models currently supported, and OpenAI says coverage will extend to all legacy models over the coming weeks. Turning the watermark on does not give you access to the detector, which stays limited to approved research and academic organizations.
The watermark does not replace visible AI labels
A watermark is meant for software, so it does not replace a visible AI label. OpenAI's help page says watermarks and C2PA metadata do not replace visible labels, banners or other notices that the law may require. Whether you also need a visible disclosure depends on how and where you publish the text, and OpenAI leaves that call to your own legal team.
"Turn on Allow text watermarking, select your models, then select Save."／"Organization default: Organization settings → Data controls → Text provenance."／"Project override: Project Settings → Text provenance."（I’m an API customer, how do I turn on text watermarks?）／"No. Watermarks and C2PA metadata provide information that software can read. They do not replace visible labels, banners, or other notices that may be required."（Do provenance signals replace visible AI labels or other disclosures?）— OpenAI Help CenterView official source →
FAQ
- Is my ChatGPT text watermarked if I live in the US or UK?
- No. Automatic watermarking covers eligible ChatGPT and Codex users in the EU only, and OpenAI says it is not making it a global default. Outside the EU, text is watermarked only when an API customer switches the feature on for selected models.
- Can I check a document for the ChatGPT watermark myself?
- Not yet. The text detector is limited to approved researchers and expert organizations, and OpenAI is keeping it private because of missed watermarks and false positives. The public openai.com/verify page checks images and audio, not text.
- Does the watermark reveal who wrote the text or what they asked?
- No. The detector reports only whether it finds an OpenAI watermark. It does not identify the user, and it does not reveal the prompt or the conversation that produced the text.
- Do AI writing detectors use the ChatGPT watermark?
- No. Tools such as Pangram run a classifier over finished text and look for patterns like word choice. They do not hold OpenAI's secret key, so they cannot read the embedded signal, and their verdicts are unrelated to the watermark.
Related articles
OpenAI, ServiceNow expand deal with ChatGPT pluginOpenAI and ServiceNow expanded their deal on Oct 28, 2026. A ServiceNow MCP plugin for ChatGPT is live now; Codex, security and voice tools follow in 2027.
ChatGPT and Codex usage limits merge in DecemberFrom December, Plus, Pro and Business plans will share one token-based allowance across ChatGPT and Codex. Instant chat won't count, and no start date is set.
ChatGPT now works with iCloud Mail and CalendarOpenAI added iCloud Mail and Calendar plugins to ChatGPT on Oct 27 for all plans. You connect with an Apple app-specific password that outlives a disconnect.
ChatGPT limits weapons information for teensOpenAI said on Oct 30, 2026 that ChatGPT will limit weapons information for teens, end chats fixated on violence, and report on safety every six months.