I don’t like when things shift automaticallly, it breaks my mental model What about niri?

They would need to be trained for open ended creative tasks, which is just hard in the current reinforcement learning paradigm.

Well the implementation would still need to be audited, but you could argue this is more objective and practical than overseeing manual counting.

The idea of using zero-knowledge proofs to avoid identifying voters has been proposed. Don't know if this is just theoretical at this point or if machines implement it, but that would solve the issue you mention.