OpenAI Introduces Invisible Watermark for ChatGPT Text and Code in Europe

OpenAI is marking ChatGPT's text and code outputs with an invisible watermark across the European Union. However, testing shows that altering just one in four words to synonyms can drop detection rates to as low as 17%.
This move aligns with the EU AI Act, which requires machine-readable labels on AI-generated content. OpenAI states its technology is still immature and has not publicly released detectors for general use.
The system, called textGrain, leaves hidden statistical patterns in word choices rather than adding visible markers or special punctuation. Detection rates rise with text length: 66% for around 150 words, increasing to 92% at 300 words. Simple edits can weaken the watermark significantly—replacing one out of ten words drops detection from 92% to 66%, while changing one in four words reduces it to 17%.
OpenAI does not provide public tools for detection but offers access to approved researchers and organizations, emphasizing that results won't reveal user data or conversation content. Detection failures are common with short, edited, or translated text, as well as with competing AI tools.
For images and audio, OpenAI provides a public tool called SynthID for watermark checks. Outside Europe, ChatGPT watermarks are disabled by default but can be enabled via API settings for global customers. This implementation follows the EU's August start of transparency enforcement under its AI Act, with potential implications for crypto-related compliance as AI misuse in crime has risen 40% year-on-year according to TRM Labs.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 챗GPT 워터마크 유럽 도입…오픈AI 감지기 준비