As of late 2026, OpenAI has started watermarking ChatGPT text for users in the EU, using a statistical pattern in word choice rather than hidden characters. Google has watermarked Gemini app text with SynthID since 2024. Neither watermark can be seen or removed by deleting characters. Stray characters in copied AI text, such as narrow no-break spaces, are formatting quirks, not watermarks.
What OpenAI does
For a long time, the answer was no. In August 2024 OpenAI said it had built a text watermark but was holding it back, because rewording or translating defeated it easily and it could unfairly affect non-native English speakers.
That changed in October 2026. OpenAI announced a statistical text watermark, rolling out first to ChatGPT and Codex users in the European Union, with an opt-in for API developers worldwide. As reported:
- It works by slightly shifting which words the model picks, using a secret key.
- It doesn't add any characters, visible or invisible.
- Detection is restricted, not a public tool.
- Light editing weakens it, and OpenAI says the absence of a watermark doesn't prove a human wrote the text.
So whether a given piece of ChatGPT text is watermarked depends on where and how it was made.
What Google does
Google has watermarked text generated in the Gemini app and on the web with SynthID Text since 2024. It also works through word choice, and Google released it as open source in 2024. Google's documentation says it survives small edits and light paraphrasing, but gets much weaker after thorough rewriting or translation. See SynthID explained.
What Anthropic does
Anthropic announced in August 2026 that text from Claude carries a statistical watermark worldwide, a variant of Google's SynthID Text. Like OpenAI's and Google's, it works through word choice, not hidden characters. Details in does Claude watermark text?
The hidden character myth
Many "ChatGPT watermark remover" sites claim AI text is marked with invisible characters. Here's what's actually true:
- Copied text from any chat interface can carry non-breaking spaces, zero-width spaces and curly quotes, mostly from display formatting.
- In 2025 some OpenAI models produced narrow no-break spaces (U+202F) in long answers. OpenAI reportedly said this wasn't a watermark, and the characters later stopped appearing.
- Unicode tag characters can hide a real message in text, and some tools use them to trace copies. These are rare in chatbot output.
Cleaning these characters is useful: it fixes broken formatting and removes anything hidden you didn't put there. But it has nothing to do with the statistical watermarks above.
What this means for you
| If you want to… | What works |
|---|---|
| Fix spacing and formatting in AI text | Paste it into the ChatGPT text cleaner |
| Check for hidden messages in text | The hidden message decoder |
| Know whether text is watermarked | Not possible without the company's detector |
| Make AI-assisted writing your own | Rewrite it in your words, and disclose AI use where it's expected |
| Say how you used AI | The AI disclosure statement generator |
The rules behind it
The EU AI Act's transparency rules (Article 50) apply from 2 August 2026. They require providers of generative AI to mark synthetic audio, images, video and text in a machine-readable way, and require disclosure in some cases, such as deepfakes and AI-written text published on matters of public interest without human editorial review. OpenAI, Google and Anthropic now all watermark at least some of their text output, and more companies are expected to follow.