OpenAI Watermarks ChatGPT and Codex: What Changes for EU Users

OpenAI will start adding invisible watermarks to some of its ChatGPT and Codex text outputs for European Union users, in line with transparency requirements of the EU AI Act. These watermarks, created by a technology called textGrain, subtly alter the way the AI models choose words to create a detectable pattern in the text. However, there are important limitations to the system.

The watermark will only be applied to eligible responses, not to every single one, and it can be significantly weakened if the text is edited, translated or paraphrased. Code snippets generated by Codex also get different treatment, as there are fewer possible choices for the next piece of code, limiting how much textGrain can alter the output.

Additionally, the system excludes code responses under 200 tokens, roughly 150 English words, from the watermarking process. Users will not have direct access to the detector initially, and it will only be available to approved researchers and qualified expert organizations. For EU users, having their own records and documentation of AI use remains crucial, as the invisible watermark cannot provide detailed explanations about why AI was used or how a text changed afterward.