# ChatGPT now watermarks text in the EU. "Just have another model rewrite it" fails in three ways, and I measured each — retained record

> Retained reddit record from reddit:ChatGPT.

Canonical: https://highsignal.app/data/records/b7663caf6fc9889f

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ChatGPT)  published 2026-10-09 15:20 ingested 2026-10-10 02:34

[original source ↗](https://www.reddit.com/r/ChatGPT/comments/1x1o5t2/chatgpt_now_watermarks_text_in_the_eu_just_have/)

## retained text



```
OpenAI said this week that ChatGPT is beginning to mark the text it produces for users based in the EU. Claude has done this since August in all places, and Gemini's app has done it since 2024. And no detectors are public, which is quite frustrating. In each thread on this topic, someone says "just have another model rewrite it", and I wanted to learn what that really does, so I tested it. My first language is not English, and most of what I write publicly passes through a model, so this issue is mine as well.

The first problem is that the model doing the rewriting may put a mark on the text once more. If you ask Claude, you swap one mark for a different one, because Claude applies its mark at the level of the model, in the app and also in the API. Gemini's app adds a mark. The kinds of model that carry no known mark are those from China: DeepSeek, GLM and Qwen, and anything you run on your own.

The second problem is that when you ask a model to "rewrite this", it often changes very little. In August I handed six model systems the same six documents, along with the same simple instruction: they changed from 40% to 73% of the words. When I tried the free models available on OpenRouter, they did worse: one gave me back my text with 2% changed, and a free Qwen with reasoning turned off changed one third. To give a sense of scale, the OpenAI post says that swapping out one fourth of the words for synonyms lowered their detection from about 92% to 17%, and for 10% of the words, only to 66%. One third of the words is not much.

The third problem is the one that matters to me, because it is silent, is losing facts. Four judge model systems, each from a different vendor, tallied gross errors in meaning in those rewrites without knowing which model had made them. MiniMax made one error across six documents while it changed 40% of the words; Qwen3.7 Plus changed 73% , but produced almost three errors per document; a small Qwen had four. Another idea - back translation, going to German and back was worse than a simple rewrite using the same model, and going through Chinese was worse still, with over four errors per document. A reviewer who read six texts produced by my own rewriter found that "should" became "must" and "covers 17.5%" became "covers up to 17.5%". Little words, big outcomes in a thesis or contract. Always read yourself what AI has paraphrased.

So if you take this path, 

* choose a model that has no mark, 
* ask it to rebuild each sentence instead of improving the text (the weak free models paid no attention even to that), 
* tell it to keep each term, name and number, 
* and after that check two things: what share of the wording actually changed, and whether your facts survived. 

For the easier check I built a fully free and open source comparison tool that shows what share of five-word sequences (which is a good proxy metric to check if watermark survives, according to my research) are new, which facts are missing , which sequences didn't change; it is a rough substitute, since the actual detectors are private, and it is the only measure available to you. The tool, the complete write-up with every table has a link in the comments, and the open-source rewriter I use is linked there too; it works with your own key and the free models.

PS. why watermarking pisses me off so much [https://www.reddit.com/r/ChatGPT/comments/1x1o5t2/comment/pev8s5q/](https://www.reddit.com/r/ChatGPT/comments/1x1o5t2/comment/pev8s5q/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button)
```



document fetched 2026-10-10 00:20 · canonical https://reddit.com/r/ChatGPT/comments/1x1o5t2/chatgpt_now_watermarks_text_in_the_eu_just_have
