OpenAI will watermark ChatGPT outputs by default in the EU

OpenAI announced it will automatically add watermarks to text generated by ChatGPT in the EU, while the feature is also available in other regions but off by default. The EU move is to comply with the EU AI Act, which takes effect in August and requires AI-generated content to be marked in a way detectable by other tools. OpenAI's watermarking method is proprietary, named textGrain, and it has published a technical paper explaining how it works. The approach is similar to other LLM watermarking tools: it embeds patterns in word choices that are imperceptible to human readers and do not noticeably affect output quality, and holders of the key can identify them with a dedicated detector. OpenAI says it will grant detector access to a limited number of researchers and institutions, with others able to join gradually through an application and approval process. The report also notes that existing standards such as SynthID and C2PA are relatively easy to bypass for people with basic knowledge.