OpenAI Adds a Watermark to ChatGPT Text

OpenAI is adding an invisible watermark, textGrain, to text generated by ChatGPT and Codex, its chatbot and code-writing tool. The move is meant to satisfy the EU's AI Act, which requires AI-generated text to carry a machine-readable marker. In the EU, the watermark will become mandatory for ChatGPT and Codex. OpenAI's API lets developers build their own products on top of its models. There, the watermark stays optional everywhere in the world. OpenAI tuned its detector to a 1% false-positive rate, so it rarely flags ordinary text by mistake. At that setting, it found the watermark in about 95% of 400-token excerpts on psychology topics, roughly 300 words each. After a quarter of the words in that text were swapped out, detection accuracy fell to 17%.

For developers building products on the API, this barely matters. The feature is off by default, and turning it on takes extra effort. Even when it is on, it stops working as soon as a quarter of the text has been edited. At that point, it is no better than a coin flip. Anthropic, the company behind the Claude chatbot, does it differently. Its watermark is mandatory everywhere, through every way of accessing Claude, not just where the law requires it. For OpenAI, the EU rule looks like a box-checking exercise for compliance. It does not look like a reliable tracking tool.