OpenAI is adding invisible watermarks to ChatGPT and Codex for users in the EU

Sayan
By Sayan
3 Min Read
Disclosure: This website may contain affiliate links, which means I may earn a commission if you click on the link and make a purchase. I only recommend products or services that I personally use and believe will add value to my readers. Your support is appreciated!


OpenAI has begun adding an invisible fingerprint to content created by ChatGPT and Codex in the European Union. The rollout is in response to requirements under the EU AI Act.

We’re expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology.

Our tools already help verify whether an image or audio file was created with our…

— OpenAI (@OpenAI) October 5, 2026

How will OpenAI’s invisible text watermark work?

OpenAI calls the AI-watermarking system ‘textGrain’. It adds a hidden statistical signal through the model’s word choices, which a detector can search for later. OpenAI says textGrain matched or exceeded approaches it tested, including Google’s SynthID for text. However, the company admits detection can fail sometimes.

In testing, the detector found watermarks in about 80% of 200-token psychology passages. That climbed to about 95% with 400-token passages. OpenAI says editing can significantly weaken the signal. For instance, replacing 10% of words with synonyms dropped detection from about 92% to 66%, whereas replacing 25% of the AI-generated text reduced it to 17%.

Why OpenAI says its watermark is not foolproof

The watermark is not a magic “written by ChatGPT” stamp. OpenAI says it provides only a limited signal about its systems’ role in creating text. It cannot measure human contribution, establish ownership, identify the user, or verify whether text is accurate.

Missing a watermark also does not prove human authorship entirely. A watermark can show that AI was involved, but not necessarily how much it contributed. OpenAI is hardly the only company wrestling with this problem. Claude’s approach to AI watermarking has already raised questions about how reliable and practical these systems will be.

And naturally, people are already looking for ways around it. There is already an open-source GitHub project to remove AI fingerprints, including statistical text watermarks.

OpenAI will add watermarks to eligible ChatGPT and Codex output across all EU plans over the coming weeks. However, it will not make watermarking a global default at launch. API customers worldwide can opt into watermarking for selected models starting now. Approved researchers and expert organizations can also apply for detector access, but it will remain restricted initially.

RELATED COVERAGE:



Source link

Share This Article
Leave a Comment

Leave a Reply

Your email address will not be published. Required fields are marked *