The watermark is not a visible symbol or hidden characters. Instead, it is a statistical signal woven into the words themselves, subtly shaping the model's word choices so that a detector can later spot the pattern. Because the signal lives inside the text, it travels with the content when it is copied and pasted. OpenAI said the watermark does not identify the user and that internal testing showed no meaningful change in model performance with it switched on.
Why the EU — and why now
The rollout is a direct response to the EU AI Act's transparency obligations, which took effect on August 2, 2026 and require AI companies to mark AI-generated content in a way that other systems can identify. OpenAI said the watermark will reach eligible ChatGPT and Codex users on all plans in the EU over the coming weeks, and there will be no way to switch it off.
Outside the EU, developers using OpenAI's API anywhere in the world can turn watermarking on for select models starting today, but it is off by default. OpenAI said it is not making text watermarking a global default at launch. The company is the last of the major AI labs to take this step: Google DeepMind has watermarked its outputs for years and Anthropic launched a similar marker in August, while Google recently announced users could remove watermarks from images, videos and songs in countries where they are not required.
How textGrain works
OpenAI calls its system textGrain. Alongside the announcement, the company published a technical report co-written with researchers from the University of Pennsylvania and Yale that walks through the method: a secret key is used to sort next-word predictions, and hundreds of these tiny nudges together let a detector identify AI-generated content using only the text and the key.
Detection access will not be public at launch. OpenAI said approved researchers and expert organizations will be granted access on a case-by-case basis, citing the risk of false positives.
The company stressed what the watermark cannot do: it cannot prove who used the system, how much a person contributed, who owns the text, or whether the content is accurate — and a missing watermark does not prove a human wrote something.
The watermark can be weakened
OpenAI's own tests suggest the watermark is far from unbreakable. In one test, replacing just 10 percent of words with synonyms dropped detection from about 92 percent to 66 percent, and replacing 25 percent cut it to 17 percent. Detection reached about 80 percent on 200-token passages and about 95 percent on 400-token ones, while short passages, math answers and translated text are harder to detect.
The admission is a notable shift from the company's earlier stance. OpenAI developed text-watermarking technology years ago and reportedly judged it internally to be highly effective, yet chose not to ship it. European regulation has now forced the issue — a sign that, as one outlet put it, regulation works.
By Hamza Faisal
