Starting in the coming weeks, some of the words ChatGPT writes for people in Europe will carry a secret. OpenAI said on Monday, October 5, that it will add an invisible watermark to text generated by ChatGPT and Codex in the European Union. The reason is the EU AI Act, whose transparency rules took effect on August 2 and require AI companies to mark generated content in a way other systems can detect.
The short version
- The watermark rolls out to eligible ChatGPT and Codex users on all plans, but only in the EU
- Developers anywhere can switch it on through the API for select models, but it is off by default
- It works by nudging word choices, so it survives copy and paste
- OpenAI’s own test shows detection falling from about 92% to 66% after swapping 10% of words for synonyms
- Only approved researchers and expert organizations get the detector at first
How the watermark works
This is not a visible symbol or hidden character. OpenAI’s method, which it calls textGrain, works inside the model’s writing process. A language model picks each next word from a list of likely options. With a secret key, textGrain quietly sorts those options so that some choices are slightly more likely than others. Any single choice looks normal. Over hundreds of words, though, the pattern adds up, and a detector that knows the key can spot it using only the text.
OpenAI published a technical report on the method, co-written with researchers from the University of Pennsylvania and Yale. The company says the watermark does not identify the user and that it saw no meaningful change in its models’ performance with the watermark switched on. Because the signal lives in the wording itself, it travels with the text when someone copies and pastes it into another document.
Where it breaks
OpenAI is unusually frank about the weak spots. In one test, replacing 10% of the words with synonyms dropped detection from about 92% to 66%. Short passages, math answers and translated text are harder to detect too. That is not surprising. A watermark needs enough words to build a statistical pattern, and translation rewrites nearly every word.
This is why OpenAI is limiting the detector to approved researchers and expert organizations for now. A public tool that gives wrong answers can accuse the wrong person. The company also warns that a missing watermark “does not prove human authorship.” The text may be too short, too heavily edited, or written by a different company’s AI.
| Question | Answer |
|---|---|
| Who gets the watermark? | ChatGPT and Codex users in the EU, on all plans |
| What about the API? | Opt-in worldwide for select models, off by default |
| Does it identify me? | No, according to OpenAI |
| Who can run the detector? | Approved researchers and expert organizations |
| Hardest cases | Short text, math answers, translations, heavy edits |
Why now, and why only the EU
The EU AI Act is the push. Its transparency rules apply since August 2, and the big labs have signed up to the bloc’s code of practice on AI-generated content, including Anthropic, Google, Meta, Microsoft and OpenAI. We explained the wider rules in our guide to how the EU AI Act became enforceable and who is not ready.
OpenAI is not making text watermarking a global default at launch. That is a notable choice. The Wall Street Journal reported in 2024 that the company had a text watermark ready but held it back, partly out of fear that users would move to rivals that did not mark their output. Two months ago Anthropic took the opposite route and said it would watermark Claude’s text worldwide. That drew backlash from some Claude users, who argued they supplied the instructions, context and decisions while Claude was only a tool.
Google’s detector opens to everyone
The same week, Google gave the public its own answer. On Tuesday it launched a SynthID website where anyone can upload an image, video or audio clip and check whether it was generated with Google’s AI tools. It had been limited to select journalists and researchers since Google I/O last year.
- Images: JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF and GIF
- Video: MP4, MOV and WEBM
- Audio: WAV, MP3, OGG, FLAC, AAC and M4A
Google says people already make about 1 million verification requests a day, and SynthID checks are built into the Gemini app and Chrome. OpenAI, NVIDIA and Kakao also support SynthID, and Apple is said to be adding support soon. Microsoft and Meta run their own standards, and none of these tools is perfect. They often fail to recognize content from other companies’ models. If you want practical tips for spotting fakes without a detector, see our guide on how to spot AI-generated images when the old tricks stopped working.
What to keep in perspective
- A watermark is a hint, not proof. Detection depends on length and how much the text was edited
- No watermark does not mean human. OpenAI says so itself
- Standards do not match. Each company marks and checks content differently, so one universal checker does not exist yet
- Removal tools are spreading. Paraphrasing is the obvious workaround
What it means for you
- If you use ChatGPT in the EU, expect your output to carry the mark within weeks, even after you copy it elsewhere
- If you are a student or writer, heavy editing weakens the signal, but it never proves anything, so do not treat a detector result as a verdict
- If you build with the API, the opt-in switch lets you test it, and you should plan for the rules to tighten
- If you publish content, disclosure is the safer path than hoping nobody checks
Watermarks are a useful first layer of trust, but only a layer. The honest summary from OpenAI itself is that it can show a system generated or processed part of a passage, not how much human thinking went into it.
Sources and further reading
- TechCrunch: OpenAI will start watermarking ChatGPT’s text in the EU
- TechCrunch: Google’s new SynthID website can identify AI-generated media
- Cybernews: Text watermark removers surge after AI labeling law comes into effect in the EU
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
