Does ChatGPT watermark its text?
What OpenAI has researched, what it has actually deployed, and what remains detectable in GPT output.
Which providers watermark · 2 min read
OpenAI has researched text watermarking for years and discussed it openly. As of this writing it has not deployed one in ChatGPT.
That now makes OpenAI the outlier: Google has shipped SynthID-Text in Gemini, and Anthropic began marking Claude's text output in August 2026 under EU AI Act Article 50. OpenAI is the remaining major provider without a deployed text watermark.
What OpenAI has and hasn't done
OpenAI built and evaluated a text watermarking method internally, and reporting in 2024 described it as working well in testing while being held back over concerns about how it would affect users — non-native English speakers in particular — and how easily it could be defeated by paraphrasing.
Separately, OpenAI did ship and then retire a classifier — a tool that guessed whether text was AI-generated by analysing the text itself. It was withdrawn in July 2023 for low accuracy. That was a detector, not a watermark, and the distinction matters: a classifier guesses from the outside, a watermark is a signal deliberately planted at generation.
So the current state is: no deployed watermark, and no first-party detector either.
What remains findable in GPT output
Plenty, but none of it is a watermark:
- Em dashes, heavily. Same as Claude.
- Curly quotes and apostrophes rather than straight ASCII.
- The U+2026 ellipsis character.
- Occasional non-breaking spaces and exotic whitespace.
- A recognisable vocabulary layer: "delve", "landscape", "realm", "seamless", "pivotal", "it's worth noting", "in today's fast-paced world".
These are removable in the exact sense for the characters, and only rewritable for the vocabulary.
Why the paraphrasing problem killed it
Any token-level watermark shares a weakness: the signal lives in the specific words chosen. Rewrite the passage and the signal degrades. Translate it and back, and it degrades further. That means a watermark reliably catches unedited copy-paste and reliably misses anyone who edits, which is exactly backwards from a compliance point of view — it penalises the careless rather than catching the deliberate.
This is also why watermark detectors report a confidence score rather than a yes or no, and why that score falls apart on short passages. Anthropic's documentation for Claude's mark names the same limitations: heavy editing, paraphrasing, translation and very short passages.
Practical takeaway
If your concern is "will something detect that ChatGPT wrote this", the honest answer is that no first-party signal exists to find. What exists are third-party classifiers making statistical guesses, and they are wrong often enough in both directions to be a poor basis for any decision.
If your concern is "does my document contain hidden characters", that is a different and answerable question. Run a free scan and you will get exact counts.
Related: does Anthropic watermark Claude's text? · EU AI Act Article 50 · what SynthID does