OpenAI's New ChatGPT Watermark Breaks If You Edit One Word in Four

Not sure what EU regulators really expect here. There's no way to watermark plaintext that can't be circumvented in a few seconds with even very simple methods. Anybody with a Claude/Codex subscription can build their own "unwatermark text" webapp in five minutes flat. (And if Claude/Codex refuse, that's what Mimo and DeepSeek are for...)

The risk is that the quality of output, in terms of word choice and style, will decline.