OpenAI has announced that it will start inserting an invisible — imperceptible to the human eye but detectable by machines — watermark into all text generated by its AI models. The change explicitly covers source code produced by the models as well.
According to the company, the watermark is intended to ensure compliance with European Union AI regulations, which require model-generated text to include a machine-readable identifier indicating that the content was produced by an AI.
OpenAI says the watermark will be optional at first worldwide: users can enable or disable it starting today. Over the coming weeks, however, content generated by ChatGPT and Codex within the European Union will receive the watermark automatically, and in that region it will no longer be possible to turn it off.
The company also stated that researchers and various organisations will later be granted access to a tool capable of detecting the invisible watermark. OpenAI explained its phased rollout and restricted early access by noting that the detection system may initially produce false positives — incorrectly labeling human-written text as AI-generated — and therefore it will provide verification tools on a case-by-case basis at the start.
Other industry players have already implemented similar approaches: Anthropic introduced a machine-readable label in August, which some users attempted to circumvent, while Google DeepMind has been watermarking its generated content for years. Google, however, announced in August that where watermarking is not mandatory, users can disable watermarks on images, videos and music.
In the United States, federal-level AI regulation has made limited progress so far. A Rainey Center survey indicates that nearly eight in ten respondents support rules requiring watermarks for AI-generated content — including 82 percent of Democrats and 75 percent of Republicans — yet substantive federal regulation has not yet been enacted.
