Will AI automation widen science's productivity versus progress gap?

Kapoor and Narayanan argue that AI is likely to widen what they call the production-progress paradox. Publication has grown exponentially, "increasing 500 fold between 1900 and 2015," while "actual progress, by any available measure, has been constant or even slowing." The evidence is metascience: Park et al. find that disruptive work is "an ever-smaller fraction" of output; a Collison and Nielsen survey finds scientists rate Nobel-level advances from the 1910s to the 1980s about as important as recent ones; Bloom et al. find that rising researcher counts are offset by falling output per researcher in semiconductors, agriculture and medicine. The claim is about the system, not individuals: "Even if individual scientists benefit from adopting AI, it doesn’t mean science as a whole will benefit."

The mechanism runs through attention and incentives. A scientist's attention is finite, so as production rises, "it is too risky for authors of papers to depart from the canon," and novel work drowns in the noise. Publish-or-perish rules close the loop: "Production is easy to measure, and progress is hard to measure," so careers reward output and discourage the slow work that yields breakthroughs. Applied to AI, the authors expect automation to make it "even easier for scientists to chase meaningless productivity metrics," and expect AI to make individuals more creative while homogenizing the collective. They also report errors they say they have already found: traditional machine learning errors in over 600 papers across 30 fields, and, citing Roberts et al., no clinically useful tool among more than 400 COVID-19 diagnosis papers.

This is the inverse of the acceleration case elsewhere in the library. Can recursive self-improvement speed up the research process itself? treats automation as speeding outputs while research efficiency stays fixed, and Could automated AI research compress years of progress into months? treats a feedback loop as turning research into progress. The excerpt challenges the premise those notes share: that more output becomes more progress. Its evidence, falling output per researcher, is a diminishing-returns pattern of the kind the R&D note cites, here found across science. The nearest quantitative neighbor, Are AI feedback loops strong enough to sustain recursive self-improvement?, also separates narrow gains from broad ones, but for AI capability rather than science as a whole.

The excerpt does not establish the forecast. It stops before the authors' recommendations and before the line "The reality is closer to the opposite," so the reversal they say is possible is not argued here. The 500-fold and doubling-every-12-years figures carry no source in the excerpt, and the Park, Collison and Nielsen, and Bloom findings appear only as the authors' summaries. The authors hedge their own claim: AI "is likely to worsen the gap," "this may not be true in all scientific fields," and the outcome is "certainly not a foregone conclusion." At that strength, the excerpt supports one practical implication: output counts are a poor proxy for AI's effect on science, and the test is whether progress measures move. It does not support a general claim that AI will slow science.