I agree that it's quite uncertain what is needed for safety in strongly superhuman systems, and/but I think behaviorist methods are *even less* likely to be adequate for superintelligent systems.
Robustly aligning human-level and weakly superhuman AI seems like it's a super important step to get right in service of eventual superintelligent alignment, though, because then we might be able to rely on them to help with that problem. And a weakly superhuman alignment researcher that has hope of tackling the problem of superintelligence alignment, in my opinion, is not formed by Skinnerian behavioral conditioning!