AI in the Lab: AlphaFold2’s Impacts on Human-Produced Knowledge

This paper evaluates how Artificial Intelligence (AI) shapes scientific knowledge production by human researchers, taking Google’s AlphaFold2 (AF2) as a case. AF2 predicts protein structures with near-experimental accuracy at a fraction of the cost. Exploiting variation in AF2’s prediction confidence in a difference-in-differences design, I find that experimental output for well-predicted proteins falls by about 35%. This decline is more pronounced for proteins that are costly to experiment on or have limited downstream demand. The reallocation of effort is also heterogeneous across experimental methods, concentrated in widely accessible X-ray crystallography while leaving the capital-intensive cryo–electron microscopy (Cryo-EM) method largely unaffected. Meanwhile, AF2 has little effect on the quality of execution but appears to alter incentives: the academic reward for working on experimentally unattempted but well-predicted structures largely disappears, as reflected in a sharp drop in the probability of publication in top-tier journals. Overall, AF2 shifts experimental effort toward not-well-predicted areas where human comparative advantage persists.