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

OpenAI plans to deploy an invisible watermarking system for AI-generated text in the EU, embedding subtle statistical patterns via altered word choices in ChatGPT and Codex outputs. This watermarking, named textGrain, allows later detection of AI-generated content but is not foolproof—normal text edits can degrade detection rates substantially. The watermark is not globally enabled by default but is available as an opt-in feature for API users worldwide. OpenAI restricts access to watermark detection tools initially to approved researchers and organizations. The watermark does not disclose user or prompt information and does not reliably prove human authorship if absent. Benchmark tests show detection rates vary by text length and subject matter, with lower detection in shorter or heavily edited texts.

The watermarking technology enables detection of AI-generated text with moderate reliability, particularly for longer, unedited passages. However, its effectiveness decreases significantly with text edits or shorter text lengths. The watermark does not compromise user privacy or reveal identifying information. It does not prevent or mitigate security vulnerabilities but serves as a tool for content provenance and AI text identification. There are no known exploits or direct security threats associated with this watermarking technology.

No remediation or mitigation is required as this is a watermarking feature implemented by OpenAI to identify AI-generated text. The technology does not introduce a security vulnerability or threat. Organizations should be aware that watermark detection may be unreliable if the text is edited or translated. Access to watermark detection is controlled by OpenAI and limited to approved parties initially.