Interesting perspective although at this moment I do not quite see how that would work.

For me, the whole idea of "smart contracts" is that one restricts severly the range of outcomes. They build trust between parties because both see the code and can "map" in advance the entire range of outcomes.

If a smart contract on-chain calls an AI module off-chain then isn't "mapping out the range of outcomes" becoming a lot more difficult or nearly impossible (inherent to the "I" of AI ...) ?

To be discussed further ...