Applying symbolic logic, hermeneutics, retroduction, and criticism are higher abstract levels of creative problem solving, via inference to the likely explanation with respect to priors, a.k.a. the scientific method. These kinds of reasoning simply aren't practically doable by math alone yet, not to mention deep learning tools in the field.
The next wave of AI methods will need to go beyond statistics to come full circle, because classification is not reasoning. What else is there? All the GOFAI stuff, prolog, expert systems, knowledge graphs, fuzzy/modal logics. These things were always great at reasoning but the last winter happened because they were bad at dealing with uncertainty. Now the stats methods have solved that problem but they suck at reasoning. It is time for a synthesis.
As a startup I don't want a data science PhD to tinker with an optimal CNN design for a year to eke out that extra 1 or 2 percent of precision and recall. I want an ensemble of specialized models, ideally pre-trained, grafted together in a logically consistent way that is configurable and introspect-able via text/data at runtime.
The economy is misconfigured! That said, humanities people should probably learn how to code as a research skill.