OpenAI is adding an invisible watermark called textGrain to ChatGPT outputs, claiming it survives ordinary use. Internal data indicate the watermark signal weakens over time. The article's author tested how easily the watermark can be broken.