watermarking then makes the model more likely to answer harmful requests that it would otherwise refuse.

To summarize: Watermarking fucks everything up by not only making LLM output less reliable in tool calling, but also making it less safe in general.

I didn’t think about it much before I read this research, but watermarking applies to all output. Not just text. That means when it generates code or JSON or similar technical, specific things it completely fucks it up a lot of the time.

The logical conclusion is the same that the world came to in the 90s: Watermarking is a bad idea for digital things.

I mean, come on: The water in the mark is digital!