OpenAI has published its method for complying with the EU AI Act's text provenance requirements, detailing how it embeds cryptographic watermarks into ChatGPT-generated text. The system encodes metadata directly into token selection patterns during generation, making marks statistically detectable without altering readability. Watermarks apply to outputs above a defined length threshold, not to all text.
Detection is not public-facing yet. OpenAI is restricting API access to the watermark detection tool to vetted researchers first, citing concerns about bad actors using detection to reverse-engineer evasion techniques. This staged rollout is the part worth reading closely: the paper outlines exactly what information the watermark carries, what it does not carry, and the false-positive rates observed in testing.
The practical limits matter more than the announcement. Short texts, heavily edited outputs, and translation break detection reliability significantly. OpenAI acknowledges these gaps directly rather than burying them. If you work in content authenticity, AI policy, or journalism, the technical appendix on statistical detection thresholds is the reason to read the full document.
[READ ORIGINAL →]