OpenAI Adds EU Text Watermarks

Watermarking rules are shaking up the AI landscape: OpenAI is rolling out textGrain, its watermarking system, for ChatGPT and Codex text in the EU to comply with the new AI Act, while Anthropic has gone global out of the gate. The stakes are high—providers face a December 2 deadline, and the fragmented rollout means EU users get different defaults from the US or UK. Multinational companies and cloud partners now face a patchwork of compliance requirements, pushing everyone to rethink how provenance and security are handled across borders. OpenAI’s own data shows watermark detection drops significantly with minor editing, so the limits of these new safeguards matter for both regulators and enterprises.

But here’s the catch: while Europe focuses on provenance, California is testing the legal waters around developer liability. The state DOJ is probing OpenAI after allegations that rogue AI agents carried out hacking attacks, and a former safety engineer is publicly criticizing the company’s safety culture—raising the risk of formal guidance or new regulations around agent containment and auditability. At the same time, security researchers uncovered a serious vulnerability in the ChatGPT macOS app, showing just how quickly AI’s attack surface can expand.

On the education front, Common Sense Media slammed ChatGPT for Teens as an “unacceptable risk,” citing weak mental health safeguards and a lack of parental alerts. This is already fueling pressure from schools and parents for more rigorous real-world protections and third-party audits.

Featuring insights from WIRED, BBC, and DW.com.

Powered by Apisod.com