This is not correct, Alphafold predicts crystal structures of proteins. It is a known issue in the field that these folding models do not generate ensembles of protein conformations like the ensembles generated by DE Shaw Research (who built a super computer to simulate proteins). And these folding models fail when we don't have crystal structures, so they are certainly not generalizing. A great paper on the subject: https://www.biorxiv.org/content/10.1101/2025.02.03.636309v1

Drug discovery is significantly more complicated than protein folding. Small molecules with similar chemistries can adopt novel binding poses, can end up binding to off targets (PXR, hERG, etc) or simply never make it to the protein due to solubility or permeability.

I will always defer to Pat Walters who seems like the most careful and sane person in this field: https://patwalters.github.io/