OpenAI to Deploy Invisible Watermarking on EU ChatGPT and Codex Text to Meet EU AI Act

OpenAI announced it will begin rolling out invisible text watermarking across eligible ChatGPT and Codex outputs for users in the European Union over the coming weeks, complying with Article 50 provenance and transparency mandates under the EU AI Act. The technology, based on OpenAI's textGrain architecture, embeds statistical token perturbations directly into generated passages to make synthetic text identifiable in a machine-readable format.
While European consumer interfaces will enforce watermarking by default, OpenAI confirmed that watermarking will remain strictly opt-in for global API developers, enabling enterprise customers to independently manage their regulatory and transparency requirements.
Access to OpenAI's proprietary text watermark detector will initially be restricted to vetted researchers and institutional evaluation partners on a case-by-case basis. OpenAI highlighted operational constraints in its technical disclosures, cautioning that watermarking efficacy diminishes significantly with heavy user editing, word substitution, or highly constrained technical and mathematical phrasing.