Orthogonality thesis

The thesis that, in principle, very different final goals can coexist with very high intelligence. Skill at finding means does not itself determine which ends an agent pursues.

Nick Bostrom develops the argument in The Superintelligent Will. It concerns possible agents, rather than establishing which goals a particular training process will produce. It also does not say alignment is impossible.

The practical question for AI alignment is what makes a capable system pursue human-compatible aims. Instrumental convergence considers strategies that could be useful across different aims.