OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act. The Act’s transparency rules, which came into force on 2 August, require AI companies to mark AI-generated content in a way that other systems can detect. The watermark will be activated in the coming weeks for eligible ChatGPT and Codex users in the EU only.
What changes for developers
Developers using OpenAI’s API anywhere in the world can enable the watermark for selected models starting today, but it is off by default. This means an Israeli organisation building an AI system must actively decide whether to enable watermarking, especially if its clients operate in Europe or are subject to European regulation.
This decision is not purely technical. It affects how your client can demonstrate transparency to their regulator, and how other systems will identify the content your system generates.
The limits of watermarks
In tests, replacing 10% of words with synonyms dropped detection from around 92% to 66%. This means the watermark is not full proof, and it does not prevent editing or modification of the content.
Any system that relies on detecting AI content must understand these limits. If you are building a system that needs to identify AI-generated content, the watermark is one component only, not a complete solution. If you are building a system that generates content, watermarking may affect how your clients use that content.
What to do this week
If you are using OpenAI’s API to generate text, check whether your clients operate in Europe or are subject to European regulation. If so, review OpenAI’s documentation on enabling the watermark, and decide whether to activate it.
If you are building a system that needs to detect AI content, do not rely on the watermark alone. Plan additional layers of control and verification. If you need to meet EU regulatory requirements, consider reviewing AI security and governance solutions suited to regulated environments.
Sources
Frequently asked questions
Does OpenAI's watermark work on content translated into Hebrew?
OpenAI has not published technical details on how the watermark behaves in translation. In tests, light editing significantly reduced detection, so translation is likely to affect the ability to identify the watermark.
Am I required to enable watermarking if I use OpenAI's API in Israel?
No, watermarking is off by default in the API. You must enable it actively. The decision depends on your clients and their regulatory requirements, especially if they operate in Europe.