What if AI watermarks become machine-to-machine triggers?

Maybe I’m overthinking this, but Claude’s watermarking made me think about where this could go in 5 years.

Imagine most text, code and software is generated by LLMs and carries invisible machine-readable patterns. Today those patterns are meant for provenance, but theoretically a future model could be trained to recognize one as a trigger and behave differently when it sees it.

The watermark itself isn’t a backdoor. But once machines start leaving signals mainly other machines can read, you’re creating a new layer of communication and a new attack surface.

And if LLMs keep getting more capable and autonomous, I’m not sure we can assume we’ll always fully understand or control how those signals are used.

Maybe slightly dystopian, but I think the implications go way beyond simply detecting AI-generated content.