OpenAI plans to add an invisible watermark to text generated by ChatGPT and Codex in the EU over the coming weeks. The signal is intended to help detectors identify OpenAI-generated text, but it is not a reliable authorship verdict.
OpenAI Watch analysis
What happened
Anadolu Ajansı reports that OpenAI calls the technique textGrain: it embeds a statistical signal in the words selected by its models. OpenAI is also opening applications for access to a detector, initially for approved researchers and expert organisations.
The report links the rollout to the EU AI Act’s requirement for generative AI providers to make generated text identifiable in a machine-readable format. OpenAI cautions that the technology is early-stage: short passages are harder to detect, while editing, paraphrasing or translation can weaken the signal.
Why it matters
A watermark could give researchers and other approved users one more way to check whether text may have come from OpenAI’s systems. But the boundaries matter: a detected watermark does not prove a passage was entirely AI-generated, and no watermark does not prove a person wrote it. This is a clue, not a digital lie detector.
Our read
The useful part is that OpenAI is trying to make generated text detectable without adding a visible label to every sentence. The important part is the caveat: a watermark can support an assessment, but it cannot settle authorship on its own. Treating it as proof would give a fragile signal far more authority than it can carry.
What to watch
- When the EU rollout begins and which ChatGPT and Codex outputs it covers.
- Who can access the detector, and what standards applicants must meet.
- How reliably textGrain works after ordinary editing, translation or paraphrasing.
Discussion spark: Should AI watermarks count as meaningful evidence of a text’s origin, or are their limits too great for people to rely on them?
Sources and evidence
- OpenAI to add invisible watermarks to ChatGPT text in EU – Anadolu Ajansı (6 October 2026, 14:05 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.