Presume that the AI uses[judgemental bootstrapping](http://www.forecastingprinciples.com/intro_pdf/06-bootstrapping.pdf) and empirically (that is to say statistically) determines the key features that highly predictive algorithms have. This then collapses to ontology/domain problems right? That is, we have a bunch of rules or rules for constructing rules in various domains. The main problem then seems to be how to reason about domains, which has kicked the ball further up the abstraction ladder. We’ve seen machine learning models build their own ontologies though via dimensionality reduction.