Do ChatGPT, Gemini and Claude texts carry hidden watermarks?
Yes, some AI companies add watermarks to the text they generate. But this watermark is not a character hidden in the text; it is a statistical pattern in the choice of words. That is why only the company that added the watermark can check it. There is no public tool for checking it.
Which companies watermark text?
As of October 2026, the situation is as follows:
- Google: Text generated by Gemini has carried the SynthID text watermark since 2024.
- Anthropic: Text from Claude models released after August 2, 2026 is watermarked; the watermark is being added to older models step by step.
- OpenAI: In October 2026, it started adding a watermark to ChatGPT text, but for now only for users in the European Union.
How does this watermark work?
When an AI writes a text, at every step it chooses among many possible words. The watermark nudges these choices very slightly according to a secret key. A reader will not notice it, but someone who knows the key can find this pattern statistically in a long enough text.
Because the key stays with the company, only the company can do the check. For now, companies offer this check only to organizations they approve, such as researchers, regulators and fact-checking organizations. The check is less reliable for short texts and translated texts.
Are invisible characters a watermark?
In 2025, when invisible characters such as the “narrow no-break space” (U+202F) were found in some ChatGPT texts, some people claimed they were a hidden watermark. OpenAI explained that they were not a watermark but a side effect of how the model was trained. Real text watermarks do not add characters.
Invisible characters can still be a useful clue: they can show that the text was copied from an app or a website. Our text checking tool shows each of these characters where it appears in the text.
What do AI detectors do?
AI detectors such as GPTZero or Turnitin do not look for watermarks. They are statistical models that look at a text's style and estimate how likely it is to have been written with AI. They can be wrong and can flag human writing as AI. That is why they should not be used as evidence on their own.
Frequently asked questions
If I delete the invisible characters in a text, does the watermark go away?
No. A real text watermark is in the choice of words, not in the characters. Invisible characters have nothing to do with the watermark.
Can I find out for sure whether a text was written with AI?
No. Only the company that added a watermark can check it, and detectors can be wrong. Do not rely on a single tool when you make a judgment about a text.
How can I show that a text I wrote myself was not written with AI?
The strongest evidence is the writing process itself: your drafts, your notes and the version history of the writing app you used (e.g., the version history in Google Docs or Word). Keep them.