OpenAI has begun deploying invisible text watermarking for ChatGPT and Codex outputs in the European Union — starting this week and expanding over the coming weeks. The change is a direct response to the EU AI Act’s transparency requirements, which took effect on August 2, 2026, and mandate that AI-generated content be marked in a machine-detectable way.
The watermark isn’t a visible tag or footer. It’s embedded in the language itself: OpenAI’s models subtly bias word selections during generation using a secret cryptographic key. Hundreds of these micro-adjustments form a statistical pattern, undetectable to readers but recoverable by a detector armed with the same key. The company calls the technique textGrain and published a technical report co-authored with researchers from the University of Pennsylvania and Yale.
Because the signal lives in the words, it persists when text is copied, pasted, or shared across platforms. OpenAI says the watermark does not identify individual users and introduces no measurable performance drop in model output quality.
The rollout is geographically limited. Only ChatGPT and Codex users in the EU will receive watermarked output by default. Developers using OpenAI’s API anywhere in the world can opt in today for select models. Including gpt-4o and codex-2026, but it remains off by default globally. OpenAI explicitly stated it is not making text watermarking a global default at launch.
Detection has limits. OpenAI’s internal tests show replacing just 10% of words with synonyms cuts detection accuracy from ~92% to 66%. Short passages, mathematical answers, and translated text are also harder to flag reliably. For that reason, the company is restricting initial access to its detector to approved researchers and expert organizations. No public tool is available yet.
OpenAI also issued two important caveats. First, a missing watermark doesn’t prove human authorship, it could mean the text was too short, heavily edited, or generated by another AI provider. Second, the watermark only signals that an OpenAI system generated or processed part of the text. Not how much human input, editing, or creative judgment went into the final version.
The move follows Anthropic’s August 2026 announcement that it would apply text watermarking to Claude outputs worldwide, a decision that triggered user backlash over attribution and agency. Some Claude users argued they provided the prompts, context, and editorial decisions, while Claude served only as a tool.
OpenAI had built a working watermark years earlier but delayed deployment, reportedly fearing competitive attrition to rivals without such markers. A concern flagged in a 2024 Wall Street Journal report.
Anthropic, Google, Meta, Microsoft, and OpenAI are all signatories to the EU’s voluntary Code of Practice on AI-generated content. But implementation varies: Anthropic went global. OpenAI is starting with the EU only, and with clear technical and policy guardrails around detection scope and responsibility.
Next, OpenAI says it will monitor feedback from researchers and refine textGrain’s robustness before considering broader deployment. No timeline for global rollout or public detector access has been announced.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.