OpenAI begins rolling out invisible text watermarking for ChatGPT and Codex in the European Union to meet EU AI Act text provenance requirements

Decision brief

What changed

OpenAI has begun a staged rollout of an invisible, machine-readable watermark in ChatGPT and Codex text outputs for users in the European Union, framing it as part of compliance with EU text provenance rules. Initial detection access is starting with researchers rather than broad public availability, and OpenAI says watermarking does not guarantee reliable detectability in every case.

Why it matters

For leaders operating in the EU, this changes the compliance and governance baseline for AI-generated text: some OpenAI-generated content may now carry provenance signals that regulators, researchers, or counterparties could test for even when end users cannot. It also creates operational implications for product, legal, and communications teams because watermark coverage, detector access, and reliability appear limited and staged, so firms should not assume universal verification or use it as a sole control.

Affected roles

CEO COO CTO CISO CMO

Evidence

The Verge and OpenAI’s own blog both report that OpenAI is introducing invisible text watermarking for ChatGPT and Codex in the EU, with a staged rollout tied to EU provenance requirements and researcher-first access. Trending Topics EU is directionally consistent on the EU scope and staged deployment, and adds reported detector performance figures for roughly 200- and 400-token outputs, but those details are not independently corroborated in the other cited coverage.

What remains uncertain

Open questions include which models, plans, and text contexts are covered; how often the watermark survives editing, translation, or formatting changes; and when detection will be available beyond researchers. It is also unclear from the coverage how regulators or enterprise customers should validate compliance in practice, so any assumption that this solves provenance or disclosure obligations end-to-end is unverified.

Monitor next

Watch for OpenAI to publish broader detector access terms or technical verification guidance specifying coverage, performance thresholds, and availability beyond researchers in the EU.

Analytical support, not advice — assumptions and open questions stated above.

Source coverage