OpenAI to Watermark ChatGPT Text in the EU
OpenAI is shipping a compliance feature while its own detection numbers show the watermark survives copy-paste but not light editing, which puts the burden of proof back on the EU's labeling rule rather than on the tool.
Reporting from 1 source: GIGAZINE.
OpenAI will add invisible watermarks to text generated by ChatGPT and Codex to comply with the EU AI Act. The feature rolls out within weeks for eligible plans in the EU, and API developers can enable it on some models, though it is off by default. Watermarking works by subtly varying word choices. OpenAI's detection rate drops from 92 percent to 66 percent when 10 percent of words are swapped for synonyms, and to 17 percent at 25 percent.
The watermark lives in word choice, not metadata. OpenAI says the pattern is invisible to readers and persists when text is copied and pasted, because it is built into the words themselves. The company also published a report on textGrain, a technology for determining whether that watermark is present.
Detection is limited. OpenAI reports that swapping synonyms erodes it, and that short sentences, math answers, and translated text are also hard to detect. Access to the detector will start with approved researchers and specialized institutions. OpenAI warns that a missing watermark does not prove human authorship, and a present one cannot measure how much editing or judgment went into the text.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.
Sources
- GIGAZINE OpenAIがChatGPTのテキストに透かしを入れる取り組みを開始