AI in klare taal, free in your inbox every Tuesday.

ChatGPT is moving beyond text to screens you can use straight away.

OpenAI is introducing an invisible watermark for text from ChatGPT and Codex in the EU. Codex is another tool from OpenAI. The technology is called textGrain. It guides word choice so a detector can recognise the pattern. A detector is a program that looks for such a pattern. The rollout will take place over the coming weeks. It applies to eligible users on all subscriptions. The watermark is off by default worldwide in the API. The API is the link that lets companies use OpenAI in their own software. Customers can turn the watermark on there. OpenAI says the watermark remains when you copy and paste text. Editing makes it less reliable. If a quarter of the words are changed, recognition falls from almost 100 to less than 20 per cent. For now, only approved researchers and organisations can use the detector. So if the watermark is missing, that does not prove a person wrote the text. This may be a good moment to check how your organisation handles AI-written text that goes out. TextGrain makes very small changes to word choice. A detector then looks for that pattern. So readers do not see a label. OpenAI says copying and pasting keeps the pattern intact. Editing can disrupt much of it. The watermark may help identify text made by AI. But it is not proof. The detector is not available to everyone, and editing greatly weakens recognition. So the watermark is too weak to be your only check on where text came from.