OpenAI Sets EU Text Watermark Policy

OpenAI announced its strategy for complying with the EU AI Act's text provenance requirements on October 5. The plan is tiered: starting immediately, API customers worldwide can opt in to text watermarking; within weeks, eligible ChatGPT and Codex outputs in the European Union will have invisible watermarks added by default.

While OpenAI has previously released public tools for detecting generated images and audio, text watermarking remains technically limited. Official evaluations show that at a 1% false positive rate, detection accuracy drops from ~95% for 400-token passages to ~66% if 10% of words are replaced with synonyms.

Consequently, OpenAI is restricting access to its text watermark detector to approved researchers and expert organizations rather than making it publicly available. This move establishes a compliance baseline for large language models in regulated markets while acknowledging current vulnerabilities in anti-tampering robustness.