EU-bound ChatGPT and Codex text to carry invisible OpenAI watermarks

OpenAI will add invisible watermarks to eligible ChatGPT and Codex text outputs in the European Union to comply with transparency requirements under the EU AI Act, the company said.

The system, called textGrain, will be rolled out in the EU over the coming weeks. It subtly changes the statistical pattern of word choices in generated text, using a secret key, so that specialised tools can identify it as AI-produced without altering how it looks to readers.

The watermark does not consist of visible symbols or hidden characters, and does not identify the person who generated the text or establish who owns it. OpenAI said it also does not show how much of a document was written by AI or later edited by a human.

The company acknowledged that detection is less reliable when text is short, heavily constrained or substantially rewritten, and that editing only part of a document can make the watermark harder to spot.

The EU rollout applies to eligible ChatGPT and Codex outputs rather than serving as a global default. API customers worldwide can opt in to text watermarking for selected models, although the feature is switched off by default for them.

The move follows the EU AI Act's transparency rules for providers of generative-AI systems, which require machine-generated content to be identifiable in a way that is technically feasible and does not materially affect readability or usability.

OpenAI has restricted access to its watermark detector to approved researchers while it evaluates the system's performance and potential for misuse, and said the detector could produce unreliable results in some circumstances, particularly when text has been substantially edited.

The development could affect publishers, educators and employers who use AI-detection tools to assess written material. A watermark may provide an additional provenance signal, but it should not be treated as conclusive proof that a particular person used ChatGPT or that a whole document was generated by AI. OpenAI's announcement covers text only, and its separate provenance systems for images and other media use different technologies.

Sources: OpenAI, MediaNama & BusinessToday