Does ChatGPT Have a Watermark?
No, not in the way the question is usually meant. OpenAI has never shipped a text watermark. What people find in ChatGPT output is hidden characters, which is a different problem with a different fix.
Check your own text
0 chars
Your text is processed in memory and never stored. See the full list of characters this removes.
The precise answer
There are two completely different things called an "AI watermark", and conflating them is why this question is so hard to get a straight answer to.
A statistical watermark biases which words the model chooses, so that a passage carries a signal detectable with the right key. Nothing is added to the text; the signal is the word choice itself. OpenAI has never shipped one of these for text, despite reportedly having built one.
A character mark is an actual codepoint sitting in the text: invisible, but really there, and removable with a find-and-replace. This is what people find in ChatGPT output, and it was never described by OpenAI as a watermark.
What happened in April 2025
Users noticed that o3 and o4-mini were emitting the narrow no-break space U+202F where an ordinary space belonged. It was consistent enough to look deliberate, and the watermark theory spread widely.
OpenAI attributed it to a quirk of large-scale reinforcement learning rather than an intentional marking system. The behaviour stopped appearing within days, which is not how a deliberate watermark programme behaves.
Who does watermark text
Two companies do, and both are open about it.
| Company | What they ship | Since | Removable by cleaning characters? |
|---|---|---|---|
| SynthID-Text in Gemini | 2024 | No, statistical | |
| Anthropic | SynthID-Text in Claude | 11 August 2026 | No, statistical |
| OpenAI | Nothing for text | None | N/A |
Anthropic's wording is unambiguous: "Nothing is added to the text and there are no hidden characters."
Why this is driven by regulation, not secrecy
The EU AI Act's transparency rules for AI-generated content applied from 2 August 2026, and the Code of Practice on Transparency of AI-Generated Content was signed by the major model providers. Marking output is a compliance obligation on providers, which is why Anthropic announced it publicly rather than deploying it quietly.
The obligation sits with providers and, in some cases, with deployers publishing on matters of public interest. It is not a rule aimed at an individual tidying up their own draft.
So should you do anything?
If your text has hidden characters, clean them. Not to hide anything, but because they break search, code and imports. If you are publishing AI-assisted content in the EU on a matter of public interest, look at your disclosure obligations as well.
FAQ
Frequently asked questions
- Does ChatGPT watermark its text in 2026?
- No. OpenAI has signed the EU Code of Practice on Transparency of AI-Generated Content and has said it intends to extend provenance signals, but as of September 2026 it has not shipped a text watermark.
- Then why do people keep finding a ChatGPT watermark?
- Because they are finding hidden characters, which are real. Those come from the model's output, from web pages and from word processors. They are not a marking system, and they are removable.
- Did OpenAI build a text watermark?
- It has been reported that OpenAI developed a text watermarking method with high claimed accuracy and chose not to release it, citing concerns including the effect on non-native English speakers and the ease of circumvention.
- Does Claude watermark text?
- Yes. From 11 August 2026, using Google DeepMind's SynthID-Text approach, across every surface including the API. It is statistical, so no character-cleaning tool can remove it.
Last reviewed .