I see others also mentioned the voting ring detector. I think they're correct that something needs adjustment, but I don't know how it works exactly so it's hard to suggest thresholds. The patterns I am observing is that the LLM generated clones will get instantly upvoted multiple times I assume by other warmed up LLM accounts. Back in the days of running phpBB I would mitigate some of that with ranks and I assume there must be some parallels to that with account age and comments. Perhaps some math to lower the weight of upvotes by newer accounts, or the newer account being the upvote target so they can still gain some traction but slower. Perhaps one of the factors could be if the submission is for one of the public git repo sites AND it's a newer account it could have a divisor in the voting formula if it is not already.

The risk I see is that if the goal is growth but the low quality submissions are drowning out other submissions that could negate growth as some of the newer accounts that legit try to be part of a community would just drift away as they are in the poisoned well. I too struggle to think of a way to separate them out of the noise without creating a system that would just be gamed by the LLM's. On one hand if the system requires the regulars to "vouch" so to speak it will create little elitist bubbles whereas too much tweaking to algorithms will just be detected and gamed by LLM's.

Out of curiosity, are the LLM posts coming from residential and mobile addresses or from AI data-centers themselves? If it's not already that could be yet another weighting factor. And/or AI user-agents as a weight. There are many bot signals that could be weighting or division factors. Bots are easy to spot from the server.