Should they be doing both? I can't imagine shipping code that I don't know how it works, but it passes tests...
Otherwise, you'll get pseudoscience - just mix random chemicals and see how patient reacts, rinse and repeat until something works...
We call that machine learning
Yes, but with different aims. The learning about ‘how a treatment works’ is useful for developing new potential treatments and new hypotheses, but the mechanism itself is not necessary nor sufficient in its potential treatment efficacy.
What people really care about is side effects. That's one of the reasons modern pharmas are so fixated in testing old drugs for new disorders (the other reason is they might be able to get a new patent case in point when Eli Lilly made pink fluoxetine for girls with a new indication for PMS [0]).
Similarly, the best way to identify side effects is by looking at the molecule and guessing (static analysis), testing with cells, animals or other non-human biological systems (I guess this is alpha testing as we rarely think of the testers as human...), and then in human studies (early access).
Even then we roll out incrementally (phase 2), before widely testing in phase 3 and collecting feedback.
If that were the criteria, then we would not have medicine at all, because biological systems are so incredibly complicated, nobody understands even a tiny fraction of how a particular chemical will affect them.
Psych drugs have difficulties because they aren't used for their biochemical effects, they're used for their psychological effects, and psychological effects of biochemical processes are poorly understood.
So even if you're prescribed a statin for the biochemical effect on cholesterol you might get some of the purported psychiatric effects.