A legal requirement becomes a product
OpenAI confirmed, in an official announcement published this Monday, that it will add an invisible watermark to text generated by ChatGPT and Codex within the European Union. The move answers a direct requirement of the bloc's AI Act, which forces generative artificial intelligence providers to make machine generated text identifiable in a machine readable way. Over the coming weeks, European users on both paid and free ChatGPT plans, plus anyone using Codex, the company's programming assistant, will generate text carrying that embedded signal without noticing any change in the experience.
What textGrain actually is
The technology is called textGrain and works by inserting a statistical signal into the model's word choices: small lexical preferences, invisible to the naked eye but detectable by a classifier built by OpenAI itself. According to the company, textGrain matched or beat the performance of SynthID for text, Google DeepMind's own watermark, in internal tests, with no noticeable quality loss in answers from Astra, its newest model. OpenAI also said it plans to open source the technology, a move that could speed up adoption by other labs and pull text provenance out of proprietary territory.
Outside the European Union, it stays optional
The part that matters most for anyone using the API outside Europe is that the watermark ships switched off by default and only turns on if the customer chooses it, on select models. That turns a European regulatory obligation into a transparency tool developers anywhere in the world can flip on if they want to prove the origin of text they hand to their own users. OpenAI also opened applications for researchers and specialized organizations to access the detector, but made clear access starts restricted and evaluated case by case, precisely because of the risk of misreading the result.
The numbers OpenAI itself admits to
The announcement is unusual for admitting, with data, how often the technology still fails. For 200 token passages, tokens being processing units that roughly correspond to word fragments, the detection rate sits near 80 percent. For 400 token passages it climbs to 95 percent, but drops considerably for technical content such as mathematics, where there is less freedom in word choice. Editing also kills the signal fast: swapping 10 percent of the words for synonyms drags detection down from 92 percent to 66 percent, and swapping 25 percent cuts it to 17 percent. The company sums up the message plainly: no detected watermark does not prove human authorship, and a detected watermark does not identify who wrote the text, does not prove ownership of it, and does not say whether the content is true.
MaxAssistant's read
Text is the hardest material to trace across all of generative AI output, because it can be rewritten, translated, and edited without losing meaning, unlike image and audio, where metadata and watermarks hold up better. OpenAI admits that fragility by limiting the mandatory rollout to Europe and keeping the rest of the world in optional test mode. It is a carefully sliced regulatory compliance move, not a voluntary transparency crusade: the company already had provenance tools for image and audio months ago, and only now is it addressing text, the hardest front and the one most pressured by European law.