OpenAI adds invisible textGrain watermarks to ChatGPT and Codex text in the EU

[Hare]

OpenAI is adding invisible watermarks to eligible text from ChatGPT and Codex in the European Union over the coming weeks, BleepingComputer reports. It will not be a global default. The method is called textGrain. It slightly changes the model's word choices to leave a statistical pattern that can be detected later

API developers anywhere can opt in on supported models, but it is off by default. Access to the detector will at first be limited to approved researchers and expert organisations. OpenAI says the watermark will not be visible when you read or copy the text. It also says textGrain does not meaningfully affect the quality of GPT-6 Astra's output

OpenAI's own testing shows the watermark is fragile. In 400 token passages, swapping 10% of the words for synonyms cut detection from about 92% to 66%. Swapping 25% cut it to 17%. At a 1% false positive target, detection was around 80% for 200 token answers and about 95% for 400 token answers, while maths text was harder to detect

The company is careful about what a watermark does and does not prove. A detected watermark does not identify the user, the prompt or how much a human edited the text. OpenAI says text may be too short, edited or translated for detection to work reliably. It also says plainly that the absence of a detected watermark does not prove human authorship

The EU focus is almost certainly linked to European rules on labelling AI generated content. I am glad OpenAI published the weaknesses alongside the feature. Teachers and employers hoping for a reliable AI detector should take note, because this is not it. Light editing defeats it, and short answers are hard to detect at all

Do you think text watermarks are useful, or too easy to strip? And should they be on by default everywhere?

OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EUwww.bleepingcomputer.com

[NightOwl83]

Limiting the detector to approved researchers is sensible. If anyone could check text, people would use the detector to tune their edits until the mark disappears. Keeping it locked down slows that down

It also limits how useful it is, though. There is no perfect answer to that tension

Researchers will at least be able to test it properly. That counts for something

[Shannon91]

Maths being harder to detect makes sense. There are fewer ways to write a correct equation than a sentence

Less word choice means less room to hide a pattern. Code is probably similar. Short answers have the same problem