OpenAI extends text watermarking to EU ChatGPT and Codex as AI Act compliance step

What OpenAI announced

OpenAI is widening its content provenance work to cover AI-generated text in the European Union, and it presents the change as part of its response to the EU AI Act. According to a dev.to report published on October 5, 2026, the rollout is deliberately staged.

API customers anywhere in the world who qualify can switch on text watermarking for a subset of models right away. In the EU, ChatGPT and Codex outputs that qualify are set to gain watermarks over the next several weeks. Access to the detector side of the system is narrower still: OpenAI will initially restrict detector use to approved researchers while it keeps studying how well the signal survives real-world use.

How the watermarking fits together

Watermarking embeds a detectable signal in model output without visibly marking the text itself. OpenAI does not describe it as a full solution to content authenticity. As the dev.to report explains, text is simply the newest layer in a provenance stack that already handles images and audio through Content Credentials, C2PA conformance, SynthID image watermarks, and verification tooling. The company's Content Provenance API and verification tools apply to non-text media today, which leaves the text rollout comparatively limited.

What watermarks can and cannot tell you

OpenAI is candid about the technology's limits. Detection reliability, the company says, depends on factors including how long the text is, whether it has been edited or translated, and what kind of content it is. That matters because text is trivially easy to modify: passages get shortened, rewritten, translated, blended with human writing, or mined for isolated ideas, and any of those steps can weaken the signal.

The caveats cut both ways. A positive detection does not establish who wrote a piece, what the intent was, or whether a human contributed. A missing detection does not prove a person wrote the text either. For teams designing review pipelines around AI-assisted copy, a watermark supplies context about where text came from, not a ruling on originality, accuracy, rights, or editorial responsibility.

What developers and content teams should do

The practical first move, per the report, is mapping where AI-generated text already enters your workflows, whether that is marketing drafts, support replies, product descriptions, internal documentation, or coding assistance, and deciding who needs to understand what a provenance signal means.

For API users, that means weighing whether opting in suits the use case, then documenting how watermarked outputs are handled through editing, translation, approval, and publication. For ChatGPT and Codex users in the EU, it means tracking eligibility and rollout details rather than assuming every output is treated identically. Because detector access is limited to researchers for now, no broadly available public verification step exists yet, and watermark checks should not act as a standalone approval gate.

Why it matters

The EU AI Act is pushing machine-readable provenance from an academic debate into day-to-day compliance planning. Any organisation that generates, publishes, or processes AI-assisted text for EU audiences now has a concrete reason to put provenance tooling into workflow design rather than on a distant roadmap.

The gap between the two halves of the system is what developers should watch most closely. Generation-side watermarking is arriving now, but detection remains researcher-only and fragile under ordinary editing. The workable posture is to treat watermark results as supporting evidence alongside human review and clear internal records of where AI was used. Teams that read a watermark as proof, in either direction, will get answers the technology cannot actually give.