OpenAI’s EU Watermark Push Tests How Much AI Text Can Really Be Traced
OpenAI’s EU Watermark Push Tests How Much AI Text Can Really Be Traced
OpenAI has begun moving toward text watermarking in the European Union, where the AI Act’s transparency requirements took effect on August 2. Over the coming weeks, eligible ChatGPT and Codex users on every plan in the bloc will receive outputs carrying an invisible, machine-readable signal.
The company is not putting a visible label on prose. Its system, called textGrain, subtly steers word choices so that a detector with the relevant key can identify a pattern across a passage. OpenAI says the signal survives copying and pasting, does not identify the user, and produced no meaningful performance hit in testing.
The rollout is deliberately limited. OpenAI says it is “not making text watermarking a global default at launch,” describing the EU-only deployment as a chance to learn from real-world feedback. Developers worldwide can opt in for select API models, while approved researchers and expert organizations can apply for detector access.
Yet the company’s own evidence underscores the limits of the experiment. In one test, replacing 10% of words with synonyms cut detection from roughly 92% to 66%; short passages, mathematical answers and translated text are also harder to identify. OpenAI therefore warns that textGrain “does not guarantee reliable detection.”
That caveat matters beyond technical performance. A watermark can suggest that an OpenAI system generated or processed part of a passage, but cannot determine whether it is accurate, who owns it, or how much human judgment shaped it. Crucially, the company says, an absent mark “does not prove human authorship.”
OpenAI’s approach contrasts with Anthropic’s worldwide Claude watermarking push, which drew objections from users who argued that their prompts and decisions remained central to the finished work. For now, OpenAI is betting Europe will provide the first real test of whether an invisible provenance signal can earn trust without overstating what it proves.
Write a comment