OpenAI rolls out invisible watermarks for ChatGPT in the EU
ChatGPT and Codex users in the European Union will soon find their AI-generated text invisibly marked. OpenAI is embedding its textGrain watermark to meet the EU AI Act. This change only affects the EU. Other regions are not included.
OpenAI says these watermarks have limits. They do not show who wrote the text. They do not prove the text is true. The watermark is hidden from users but can be picked up by machines. Regulators and platforms can spot AI-generated content without changing how it reads. OpenAI explained in an official announcement that the rollout will take several weeks. Not everyone will see it right away.
The textGrain watermark works by embedding a statistical signal in the model's word choices, making it invisible to readers but detectable by specialized tools.
Why OpenAI is adding textGrain in Europe
Regulators pushed for this move. The EU AI Act forced big AI companies to add new transparency tools. OpenAI says textGrain matches or beats other options like Google DeepMind's SynthID. Anthropic started using SynthID last August for the same reason. OpenAI is not making textGrain global. The legal rule comes from Article 50 of the EU AI Act. This rule says that, starting August 2, 2026, AI-generated or changed content must have a machine-readable label. The goal is to make detection possible. Details are in a report by Indian Express.
Not everyone is happy. Some users call the watermark pointless. Others doubt it works well. OpenAI admits textGrain "does not guarantee reliable detection". It cannot prove who owns the text or how much a human helped. It does not check facts. OpenAI's own tests found no big difference in output quality with or without the watermark. At first, only approved researchers and expert groups will get the detection tool. Short texts or edits can make detection harder.
The watermarking feature is not enabled by default for API users worldwide; it is initially limited to eligible plans in the EU, while API customers can opt in for specific models if desired.
What this means for users and the industry
For now, only EU users will get watermarked ChatGPT and Codex text. The rest of the world is not affected. OpenAI is holding back a global launch. Other AI companies are also scrambling to meet new rules, even if the tech is not perfect. OpenAI says textGrain only shows its system was used. It does not reveal who wrote the text, how much was edited, or if the facts are right.
This is not OpenAI's first big change under outside pressure. As reported earlier, OpenAI recently canceled the GPT-6.1 Astra launch after safety problems. Now, textGrain watermarks are another sign that OpenAI will make changes when needed. Sometimes, the company has no choice.
For OpenAI, this is about following the rules. It is not a technical leap. The watermark does not prove who wrote the text or if it is true. Detection is shaky. EU users get more invisible checks, but little real benefit. For the industry, it is clear. Lawmakers set the pace. Compliance wins over innovation.
