The AlphaFold story: a technique for predicting protein structures at experimental precision (for many proteins, anyway!) and assigning confidence to scores was developed, hailed as a revolution in structural biology (many declared that the problem was solved) and then subjected to more scrutiny as it was used for more things. AlphaFold structures are often being found by many groups to not be quite as good as experimental results.
This study does a more direct comparison, arguing that AF2 may be retrospectively worse than experiments but may still be very useful prospectively, to help future decision-making. They compared experimental hit-rates and affinities from large library docking against AF2 models vs the same screens targeting experimental structures of the same receptors. Even though they got different molecules out of docking onto AF2 structurees, the hit rates were similar, so the results aren't actually subpar.
This preprint is getting a lot of buzz in places, even including Science writeups (worth checking out for a better summary than I give).