FULL STORY

textGrain: OpenAI Rolls Out Text Watermarks for EU Compliance

OpenAI extended its provenance watermarking to text with textGrain to comply with the EU AI Act. The rollout sparked debate over the quality cost of text watermarks.

2026-10-05 ~ 2026-10-06 · 2 episodes · 17 posts

Episode 1 · OpenAI Rolls Out textGrain Text Watermarking, Starting in the EU (2026-10-05, 15 posts)

Around October 6, OpenAI announced an expansion of its content provenance solution, extending watermarking to text for the first time in order to meet regulatory requirements under the EU AI Act. This marks the company's first systematic public disclosure of its roadmap for text watermarking, signaling that mainstream AI-generated text will gradually carry verifiable provenance markers.

What's confirmed

  • Effective immediately, global API customers can opt in to text watermarking for select models
  • In the coming weeks, eligible text output by ChatGPT and Codex in the EU will automatically carry invisible watermarks
  • The technology is called textGrain
  • The watermark works by embedding an invisible statistical signal as text is generated; a detector can identify it, with no visible markers or special characters added
  • OpenAI stressed in its notes: the watermark answers one question only—whether the text may have been generated by an OpenAI model; it does not affect model performance

Why it matters

  • This is the first large-scale deployment of a verifiable text provenance mechanism by a mainstream AI vendor, marking the point at which the EU AI Act's compliance requirements for generative content begin to carry real teeth
  • The watermark is invisible and does not degrade output quality, while providing detectable provenance credentials for regulators, publishers, and other use cases

Episode 2 · OpenAI Adds Text Watermarks for EU Compliance, Sparking Quality Concerns (2026-10-06, 2 posts)

To comply with the EU AI Act, OpenAI will add watermarks to text and image outputs, following Anthropic's earlier global rollout. Reddit users debate whether the technique—injecting statistical signals by not always picking the optimal token—degrades output quality.