In that situation, what function do the upvotes serve in the first place? If the potential audience already knows they’re going to read and enjoy more content from the same source, do they need to see upvotes to tell them what they already know?
(Remember that without effective permanent karma, upvotes only serve to call attention to particular posts or comments in the short term.)
If an account is upvoted because it’s posting high-quality content, we’d expect those votes to come from a variety of accounts that don’t otherwise have a tendency to vote for the same things.
Suppose you do regression analysis on voting patterns to identify the unknown parameters determining how accounts vote. These will mostly correlate with things like interests, political views, geography, etc.—and with bot groups—but the biggest parameter affecting votes will presumably correlate with a consensus view of the general quality of the content.
But accounts won’t get penalized if their votes can be predicted by this parameter: precisely because it’s the most common parameter, it can be ignored when identifying voting blocs.
There are legitimate reasons for creating a “low-usage” server to host your personal account, so you have full control over federating etc.
If we start assuming all small instances are spam by default, we’ll end up like email now—where it’s practically impossible for small sites to run their own mail servers without getting a corporate stamp of approval from Google.
Here’s an idea: adjust the weights of votes by how predictable they are.
If account A always upvotes account B, those upvotes don’t count as much—not just because A is potentially a bot, but because A’s upvotes don’t tell us anything new.
If account C upvotes a post by account B, but there was no a priori reason to expect it to based on C’s past history, that upvote is more significant.
This could take into account not just the direct interactions between two accounts, but how other accounts interact with each of them, whether they’re part of larger groups that tend to vote similarly, etc.
Instead of trying to detect and block it, just disincentivize it.
Most AI spam on social media tries to exploit various systems intended to predict “good” content on the basis of a user’s past content by tracking reputation/karma/etc. Bots build up karma by posting a massive amount of innocuous (but usually insipid) content, then leverage that karma to increase the visibility of malicious content. Both halves of this process result in worse content than if the karma system didn’t exist in the first place.