The main challenge behind AI on a decentralized Blockchain is the consensus computational overhead required for catching malicious parties (this is an uneducated assumption). The ICP offers ‘.raw.’ type of queries that makes a canister call that does not require consensus from nodes, thus making it less resource hungry and less secure. (Again uneducated guess)
It is currently possible to implement zk logic that proves that a canister call has ran from an honest node has ran the AI model correctly and has produced the desired output?