OpenAI plans invisible text watermarks for eligible EU outputs

OpenAI has announced plans to add invisible watermarks to eligible ChatGPT and Codex text in the European Union over the coming weeks. It is also allowing API customers globally to opt in for selected models, with watermarking off by default in the API.

This is not a logo placed on your writing. OpenAI’s textGrain system changes word-choice patterns to create a statistical signal that a detector can look for.

There are important limits. OpenAI says editing can weaken that signal, while detectors can miss watermarks or report them incorrectly. Initial detector access will be restricted to approved researchers and expert organisations.

For anyone who writes or edits online, the distinction is worth understanding: detecting a watermark does not establish whether an article is accurate, who owns it or how much human work went into it. Failing to detect one does not prove that the text was entirely human-written either.

I would not treat this as a simple pass-or-fail test for good writing. It is a provenance tool with limitations, and its usefulness will depend on how those limitations are handled.