One fish, two fish, but not the whole sea: Alignment reduces language models' conceptual diversity