Not really surprising, since AI (like any empirical model) is a kind of averager - it inevitably averages out a bunch of the noise in the training data. Some of that "noise" is likely associated with chaotic dynamicss and so shouldn't be discarded.