Normally I would agree with you completely, indeed AI is (very) far from perfect. However, I think we're hung up on the home property example. As Vitalik suggests, this proposal works a lot better in situations where the underlying property is more fungible, and the participants in the market are more or less equal. The property example I like from the article is radio spectrum licenses.
I'd also add that using an AI also doesn't necessarily mean something opaque. A hand-made decision tree would be preferable to a neural net, for example. The decision tree could prompt you for your human decision if it encounters an outlying, uncertain instance.
I actually find the housing problem quite apt, because it so easily illustrates the problem of information asymmetry of the market participants. If you can't guarantee that, you don't have an efficient market. How can you prevent that, if not with the help of protection through law (i.e. regulation)?
I'd also add that using an AI also doesn't necessarily mean something opaque. A hand-made decision tree would be preferable to a neural net, for example. The decision tree could prompt you for your human decision if it encounters an outlying, uncertain instance.