Michael H

Posted on Oct 16 Originally published at agent5.news

Does ChatGPT Watermark Text? What Changed in October 2026

#ai #policy #chatgpt #aiwatermarking

Originally published on agent5.news.

Does ChatGPT watermark text? Since October 2026, it depends on where you live. On October 5, 2026, OpenAI said it will add an invisible watermark to ChatGPT and Codex text for users in the European Union, rolling out over the following weeks. Elsewhere it stays off unless a developer turns it on. This guide explains what the ChatGPT watermark is, who gets it, how it works, whether it can be detected or removed, and how Claude and Gemini compare.

Does ChatGPT watermark text now?

Yes, but by default only in the European Union, where OpenAI is introducing it over a period of weeks. OpenAI's announcement of October 5, 2026 set out three steps:

ChatGPT and Codex in the EU. Eligible users on all plans get the watermark over the coming weeks. Codex is OpenAI's coding tool.
The API, worldwide. An API is the interface developers use to build OpenAI's models into their own apps. API customers anywhere can switch the watermark on for selected models. It is off by default.
A detector for a small group. A tool that checks text for the watermark exists. Only researchers and expert organizations can apply, and OpenAI approves them one by one.

OpenAI says it is not making the text watermark a global default at launch. For now, text from ChatGPT outside the EU therefore carries no watermark.

Some details are not public. OpenAI has given no exact start dates, has not said which users count as eligible, and does not explain how it decides that a user is in the EU. Its announcement and help page also describe no switch that lets ChatGPT users in the EU turn the watermark off. The only setting they describe is for API customers.

How does the ChatGPT text watermark work?

The watermark is a hidden statistical pattern in the model's word choices. It is not a visible mark. OpenAI's name for the method is textGrain.

A language model writes one token at a time. A token is a word or a piece of a word. At each step several tokens are possible, and the model picks one partly at random. textGrain uses a secret key to steer those random picks very slightly. Over a long passage, the small shifts add up to a pattern. A detector that holds the same key can then measure whether a text matches the pattern too often to be a coincidence.

OpenAI is clear about what the watermark is not:

Nothing is inserted into the text: no hidden characters, no invisible spaces, no odd punctuation.
It does not identify the user. OpenAI says it carries no link to a person, an account, a prompt or a conversation.
It does not show how much of a text a human wrote, who owns it or whether it is accurate.

OpenAI also says answer quality does not suffer. It ran the benchmarks (standard tests) it uses for Astra, which it describes as its latest frontier (most advanced) model, with and without the watermark. On DeepSWE v1.1 the score was 72.80% without the watermark and 71.68% with it. On Terminal-Bench 4.0 it was 53.90% without and 56.06% with. OpenAI says such gaps are within the normal noise between test runs.

OpenAI published a technical report on textGrain, written with researchers from the University of Pennsylvania and Yale University. It plans to release the method as open source, which means others could inspect and reuse the code. For the general idea, see our explainer on how AI watermarks work.

What does EU AI Act Article 50 require?

Article 50 of the EU AI Act is the reason the ChatGPT watermark is EU only for now. It says providers of AI systems that generate text, images, audio or video must mark the output in a machine-readable format, so that it can be detected as artificially generated. Machine-readable means software can read the signal even when a person cannot see it.

The main points:

Article 50 has applied since August 2, 2026.
Systems placed on the market before that date have until December 2, 2026 to meet the marking rule.
The law asks for marking that is effective, robust and reliable as far as current technology allows.
Fines can reach 15 million euros or 3% of worldwide annual turnover, according to the European Commission.

Large AI companies rely on a voluntary EU code of practice, published on June 10, 2026, to show they comply. About 190 organizations had signed it by the end of July 2026, the European Commission says, and TechCrunch lists Anthropic, Google, Meta, Microsoft and OpenAI among the companies that committed to it. For text, the code requires an invisible watermark, except for very short text, which it defines as under 200 tokens. OpenAI says that is about 150 words in English. How much say companies should have over rules like these is a wider debate, covered in our guide to what AI regulatory capture means.

OpenAI was cautious about this technology before the law applied. In an update dated August 4, 2024, it said it had developed a text watermarking method but was still weighing risks, including how easily it could be bypassed.

Can you detect the ChatGPT watermark?

Not as an ordinary user. People searching for how to detect the ChatGPT watermark will not find a public tool, because OpenAI has not released one.

Who can apply. Researchers and expert organizations. OpenAI reviews each application separately.
Who is left out. General business uses, and academic integrity checks such as testing student essays or theses for AI use.
OpenAI's public verify page. It checks images and audio, not text.
Other AI detectors. Tools such as Pangram judge the style of a text. They do not read the embedded signal, which requires OpenAI's key.

The EU code permits this restriction for now. It lets signatories limit text detectors to verified expert users while the technology is unreliable, and says that limit should be temporary.

Even with access, results are uncertain. A false positive is a case where the detector reports a watermark that is not there. With the false positive rate set at 1%, OpenAI's detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages on a subject such as psychology. Rates were much lower for mathematics, where there is less freedom in wording. In a test across the 24 official EU languages at the same setting, detection ranged from 42.2% for Romanian to 69.0% for Spanish. OpenAI says it can raise the watermark strength for weaker languages and did so for those below 60%.

OpenAI writes that a missing watermark "does not prove human authorship." The text may be too short, edited, translated, or produced by another company's tool.

Can you remove the ChatGPT watermark?

Copying and pasting does not remove it, but rewriting can. The pattern sits in the wording itself, so it travels with the text. Change enough words and it fades.

OpenAI tested this on 400-token passages:

Unedited text: about 92% detected.
10% of words replaced with synonyms: 66% detected.
25% of words replaced: 17% detected.

OpenAI's help page adds that substantial paraphrasing or translation can make the watermark undetectable. Three cautions apply:

Style-based detectors do not depend on the watermark at all.
The watermark and disclosure rules are separate. OpenAI says the watermark does not replace visible labels or notices that may be required.
Companies that signed the EU code agreed not to offer tools whose purpose is to get around the markings.

Is there a Codex watermark on code?

Codex is part of the announcement, but code is where the watermark is weakest. OpenAI's post covers Codex text output. Its help page says code is harder to watermark than ordinary prose, because there are fewer plausible choices for what comes next.

The EU rules also ask for less here. OpenAI says the EU code of practice does not require watermarks in code snippets or in outputs shorter than 200 tokens. The European Commission's FAQ lists source code among the outputs that fall outside the marking duty.

Anthropic describes the same limit for Claude. Code that has to be exact gets little watermarking, while free choices such as code comments can carry it. OpenAI's pages do not say which parts of a Codex answer carry the watermark, such as the explanation or the code itself.

What does it mean for students, writers and developers?

In short, no public tool can check for the watermark, and a result would not prove who wrote a text. Here is what that means for each group.

Students. A school or university cannot use OpenAI's text detector to check coursework for now, because academic integrity uses are outside its access program. A detection result would also not show how much of an essay a student wrote. Anthropic takes a different line and lists educational organizations among those eligible for its Claude detection preview.

Writers and publishers. A watermark does not settle who owns a text. Proofreading is a special case: Anthropic notes that when a model only corrects your writing, nearly all the words stay yours, leaving little to watermark. Article 50 also has a rule for deployers, meaning people or organizations that use an AI system in a professional role. If they publish AI-generated text to inform the public on matters of public interest, they must disclose that, unless the text went through human review or editorial control and someone holds editorial responsibility. Purely personal, non-professional use is outside the AI Act's scope.

Developers. In the API the watermark is off until you enable it. The organization default is under Organization settings, Data controls, Text provenance. A project can override it under Project Settings, Text provenance. There you turn on Allow text watermarking and pick the models. Enabling it does not give you detector access.

A safety note: Ars Technica reported on September 17, 2026 on a study by a researcher at Lasso Security. In it, Google's open-source SynthID-Text changed how six open models responded to harmful requests, especially under prompt injection, where crafted text tries to override a model's instructions. The study did not test Claude's own setup or OpenAI's textGrain. The report's advice to developers: test how your models and AI agents (models that act through tools) behave with watermarking switched on.

How do Claude and Gemini compare with ChatGPT?

All three watermark text in some form. They differ on where it applies and who may check it.

ChatGPT (OpenAI). Its own method, textGrain. On by default only in the EU, optional in the API. Detector limited to approved researchers and expert organizations.
Claude (Anthropic). A version of SynthID-Text, a method Google DeepMind published in 2024. Anthropic set out its approach on August 14, 2026. It applies worldwide, because Anthropic says it has no durable way yet to limit it by region. It covers supported models in the Claude apps, the API and Claude Code, and older models are being added. A detection API is in private preview for eligible organizations such as regulators, media, fact-checkers and researchers.
Gemini (Google). Google says SynthID watermarks text from the Gemini app and web experience. On October 7, 2026, Google opened its SynthID Detector to everyone for images, video and audio. That announcement does not mention text.

On editing, Anthropic says "Light editing probably won't remove the watermark completely," while a full rewrite will. OpenAI says textGrain matched or beat SynthID for text in its own tests. OpenAI's and Anthropic's detectors each look only for their own company's watermark, so no single tool checks text from all three.

What happens next

The EU rollout, from October 2026. OpenAI said on October 5, 2026 that the watermark would reach eligible ChatGPT and Codex users in the EU over the coming weeks. It gave no exact dates. It also plans to update its technical report and extend watermarking to cloud partners in that period.
December 2, 2026. By this date, AI systems that were on the market before August 2, 2026 must meet the Article 50 marking rule. The grace period for older systems ends then. Anthropic's help page says it is still adding watermarks to earlier Claude models.
February 2, 2027. Companies that signed the EU code committed to have an interoperability solution in place by this date, meaning a way for different companies' watermark detectors to work together. One option in the code is a public marker that tells a checker which company's detector to use. OpenAI has also said, without dates, that it plans to widen detector access and release textGrain as open source.

We follow these dates in our running AI policy news.

DEV Community