Does ChatGPT watermark its text?

textGrain nudges which words the model chooses, leaving a pattern a reader can't see but a detector can pick up. OpenAI reports no meaningful change in how its models perform with it switched on. Where it applies:

• ChatGPT and Codex: eligible text for users in the EU, rolling out over the weeks after the announcement.
• The API: developers anywhere can opt in for their own apps. It stays off by default for now.
• Detection: OpenAI is taking applications for its detector, at first only from approved researchers and expert organisations.

The EU part follows the EU AI Act, which from August 2026 requires AI providers serving the EU to mark what their systems generate. Anthropic, which applies its watermark worldwide, explains its own in the Claude watermark.

OpenAI published unusually concrete numbers. On passages of 400 tokens, about 300 words, its detector found the watermark this often:

• Unedited92%
• 10% of the words swapped for synonyms66%
• 25% of the words swapped17%

Detection is also weaker on short passages, in maths, where there's less freedom in word choice, and on translated text. The more of the wording changes, the less of the pattern is left.

textGrain isn't OpenAI's first text watermark. In 2024 the company said it had built one that was accurate and held up against local edits such as paraphrasing, but was easy to get around by translating the text, rewording it with another model, or asking the model to put a special character between every word and then deleting them. It also worried a watermark could stigmatise AI as a writing aid for people who aren't native English speakers. It didn't release that one.

In April 2025 people noticed narrow no-break spaces, U+202F, in text from OpenAI's o3 and o4-mini models. They look exactly like ordinary spaces, and some took them for a watermark. OpenAI said they weren't one, but a quirk of how the models were trained.

Characters like these are easy to clean out, and glyphwash turns them into plain spaces along with other unusual spaces. See invisible characters in AI text for the rest.

Since textGrain lives in word choices, deleting characters won't touch it: the wording has to change. glyphwash takes out hidden characters, then rewrites every passage with an open-weight model that adds no watermark of its own, and checks that the meaning, names and numbers survived. Each result shows how much of the original wording is left, which OpenAI's own numbers say is what matters.

If you write with ChatGPT, you can do this without leaving it: add glyphwash as a custom MCP server in ChatGPT's Plugins, then ask ChatGPT to run a text through glyphwash. The washed text comes back in the chat, with its checks. Setting it up in ChatGPT.