My technical take from this story was that human-run labs turned out to be less reliable than current AI.

This is the sort of performance/quality gap that is not obviously seen (who in the general population understands lab protocols in detail?) and could be massively improved with AI.