Moreover, if there is financial gain, the incentives in the current system are to grab an idea and keep the details of its realisation under tight wraps as far as possible: Further development is done outside of public academic circles and results are protected by trade secrets or patents.
Yes, the results of the research will eventually benefit society - when offered as a commercial service or product - but only in an opaque, "black box" kind of way as the details are not for public discussion. Even that offer is primarily driven by the desire to make money and consolidate power an only secondary to advance society.
I'm starting to believe that this may still be sufficient to fulfill the 80 of the 80/20 rule. Properly directed, powerful pattern recognition can do a lot when tied into regular old programming logic.
Regular programs work well in low uncertainty environments, and some well trained ML networks can create that low uncertainty environment out of a relatively high uncertainty environment.
The problem is to find the real problems to solve hiding under the obvious solutions.
As far as AI risks, I am more concerned of the risk of human error in believing they understand things enough to deliver systems that can really mimic intelligence when in reality I've seen nothing intelligent in anything that claims AI in it, high speed weighted pattern recognition, yes, intelligence, no.