Comment by xnorswap
We don't have the infrastructure for it, but models could digitally sign all generated messages with a key assigned to the model that generated that message.
That would prove the message came directly from the LLM output.
That at least would be more difficult to game than a captcha which could be MITM'd.
Hosted models could do that (provided we trust the providers). Open source models could embed watermarks.
It doesn’t really matter, though: you can ask a model to rewrite your text in its own words.