some llms can identify bot generated text. if you use a smarter llm it may block dumber llms.
bots can be reported with kind=1984 if they are not properly tagging their posts/profiles
there could be 'prove you are not a robot' over a NIP or something.
sheer number of bytes / sec may give them away.
messages looking similar may give them away.
proof of work could slow them down.