ChatGPT text generated for users in the European Union will now carry automatic watermarks. OpenAI announced the change, which applies by default in the EU but remains optional elsewhere.
The company says the move is necessary to comply with the EU AI Act, which took effect in August. The law requires AI-produced content to be marked so that other tools can detect it. OpenAI notes that no completely effective and reliable method for this exists yet.
Existing standards, including SynthID and the C2PA project, are described as relatively easy to circumvent for anyone with basic know-how. The source suggests the same is likely true for OpenAI’s watermark.
OpenAI’s proprietary method, called textGrain, works by inserting patterns into word choices. These patterns are not clear to human readers and do not meaningfully change output quality, but a specialized detector can find them if the user has a key. A technical paper explains how the system works.
Access to the detector will be given to a limited number of researchers and organizations. OpenAI will also provide a request-for-approval process for others to be added over time.