OpenAI adds invisible text watermarks to ChatGPT and Codex in EU

By Splitfeed EditorialLast updated Oct 5, 2026, 20:50 UTC

On October 5, 2026, OpenAI announced it would embed invisible watermarks into text outputs from ChatGPT and Codex for users located in the European Union. The rollout covers both products simultaneously, making Codex, OpenAI's dedicated coding tool, subject to the same disclosure requirement as the better-known ChatGPT assistant. The move is a direct response to the EU AI Act, which mandates that providers disclose when content has been generated by artificial intelligence. OpenAI has itself acknowledged a significant limitation in the implementation: editing the generated text can degrade or eliminate the watermark, making it harder to detect after the fact.

The broader stakes here go beyond a single company's compliance checkbox. The EU AI Act represents the most comprehensive binding legal framework for AI systems yet enacted, and how companies satisfy its disclosure requirements will set practical precedents for enforcement across the bloc. Invisible text watermarking is one of several technical approaches being explored industry-wide, but the durability problem OpenAI has flagged raises a genuine policy question: if a watermark can be removed through ordinary editing, does it meaningfully fulfill the spirit of a disclosure rule, or does it merely satisfy its letter? That tension, between technical compliance and actual transparency, is likely to draw scrutiny from regulators and researchers alike.

None of the three camps had published specific reactions to this announcement at the time of writing. The pro-AI camp would typically frame a move like this as evidence that the industry can self-regulate and adapt to legal requirements without stifling innovation, pointing to the inclusion of a developer tool like Codex as a sign of broad good-faith effort. The anti-AI camp would likely treat the durability flaw as confirmation that voluntary or easily circumvented technical measures are insufficient, and argue that the regulation needs stronger enforcement teeth. The middle-ground camp would probably welcome the compliance step while calling for independent audits and clearer standards on what robustness a watermark must have to count as genuine disclosure.

The argument will sharpen as EU regulators begin assessing whether implementations like OpenAI's meet the AI Act's requirements in practice. Any formal guidance from the European AI Office on watermark durability standards, or any enforcement action taken against a provider whose watermarks proved removable, would go a long way toward settling whether this approach is a viable compliance path or a stopgap.