Legal edition
AI-written text gets a recognisable pattern. That pattern can disappear if someone edits the text.
A watermark helps identify AI-written text, but does not prove who wrote it
OpenAI adds an invisible mark to text from ChatGPT and Codex. If the mark is missing, that does not prove a person wrote the text.
OpenAI adds an invisible watermark to text from ChatGPT and Codex. This applies to users in the EU. The technique is called textGrain. It will become available over the coming weeks to eligible users on all plans. A detector can use the watermark to identify AI-written text. A detector is a program that checks the text for the mark. OpenAI says the mark stays in place when you copy and paste the text.
If you need to assess where a text came from, the watermark may help. It is a clue, not proof. OpenAI says detection becomes much less reliable if someone changes the wording. For now, the detector is only available to approved researchers and organisations.
OpenAI is doing this to comply with the EU’s transparency rules. These are rules about being open when AI is used. They have applied since 2 August 2026. For systems that were already on the market, the next deadline is 2 December 2026. In Europe, it is therefore increasingly possible to trace whether a text came from AI. Facts, sources and context still need checking as well.
TextGrain slightly steers the model's choice of words. The model is the computer program behind ChatGPT and Codex. This creates a pattern that a detector can recognise, even after copying and pasting. Readers cannot see the pattern. OpenAI says the detector recognises it almost every time if nobody has changed the text. If someone changes a quarter of the words, detection drops to below 20 per cent.
The watermark helps identify text as AI-written. It does not prove who wrote the text or how it was used. The detector is not yet available to everyone. Editing can largely disrupt the pattern. So a detector result alone is not a basis for a legal conclusion. If the mark is missing, that still does not prove a person wrote the text.
Background for subscribersA watermark can help identify text from ChatGPT and Codex. It does not prove who wrote the text, though. Editing can disrupt the pattern, and the detector is not yet available to everyone. Other AI news this week raises a similar question: what applies already, what starts later, and what is only a signal? Keeping those apart can help when you advise an organisation.
HR AI gets a delay, but rules on openness and bans still apply
The AI Omnibus has applied since 27 July. It is a change to the European AI rules. It moves the start date for obligations on high-risk AI from 2 August 2026 to 2 December 2027. High-risk AI is AI that the law says could have major effects on people. This includes systems for hiring and assessing staff. Transparency duties and bans on certain uses still apply. Transparency duties are rules that require you to be open about using AI. A new ban also comes into force on 2 December 2026. It covers AI that creates intimate images without consent. A delay does not mean an organisation has nothing to do. Rules that already apply remain in force. AI for hiring is still classed as high-risk. Organisations have more time to meet the strictest requirements. They do not need to stop preparing. It may help to check now which rules apply to the system your organisation uses. The change moves only the start date for obligations on high-risk AI. Other rules, such as those on transparency and banned uses, still apply. Several obligations may therefore apply at the same time, each with its own date. So the delay does not pause all obligations.
Background for subscribersThis is a taster.
Subscribers read the whole Legal edition: every story, each with its background in plain language.
Subscribe →Not ready for that? Read the free letter first →
AI and tech, made simple.
This is the Legal edition. Subscribers get the background to every story, in plain language.
Subscribe
inklaretaal · Amsterdam · © 2026