Rob Patro:
Selective alignment is simply an efficient mechanism to compute actual alignment scores, it performs sensitive seeding, seed chaining & scoring, and crucially, actually computes an alignment score (via DP), while pseudo-alignment does not do these things, and is a way to quickly determine the likely compatibility between a read and a set of references.
Both of the approaches are algorithmic ways to return, in the case of pseudoalignment, sets of references [that are most compatible with the reads at hand], and in the case of selective alignment, scored mappings. They can both be run over "arbitrary" referenced indices.
If you don't (1) validate the "compatibility" you find and (2) seed in a sensitive fashion, then you end up with mappings whose accuracy is lesser than that derived from alignments. see their pre-print on this
For example, a 100bp read could "pseudoalign" to a transcript with which it shares only a single k-mer although that would preclude any reasonable quality alignment. There is no scoring adopted in the determination of compatibility, which can and does lead to spurious mapping.