Science now publishes roughly 500 times more papers than in 1900, yet measured progress has stalled; is counting papers the trap?

This explores why adding more scientists hasn't produced matching gains in discovery, and what AI does to that pattern. Before going further, one direct point: the collection doesn't hold the classic economics work on this (the 'ideas are getting harder to find' studies that count researchers against productivity growth). What it does have is something adjacent and arguably more useful. It shows that 'output per scientist' splits into two separate numbers. Papers per scientist can rise while progress per scientist falls. The clearest doorway is Will AI automation widen science's productivity versus progress gap?. Publication volume has grown roughly 500-fold since 1900, but measured progress has stalled. Their argument is that the problem isn't too few papers. It's that papers became the thing being measured and optimized for.

That reframing explains why more researchers can mean less progress each. The extra papers aren't free, because someone has to read and check them. Does peer review quality collapse under submission overload? models what happens next. More submissions overload unpaid reviewers, so journals recruit less qualified ones or pile more work on the same people. Review accuracy drops, and that tempts authors to submit more speculative work, which pushes submissions even higher. In this loop, adding researchers doesn't just dilute progress. It wears down the filter that separates progress from noise. The MIT case in Can unreviewed preprints shape scientific debate before peer review? shows what that looks like in practice: a preprint shaped debate widely before anyone checked it.

The finding you might not expect is that AI looks like a cure for low per-scientist output and may deepen the underlying problem. Does AI help individual scientists while narrowing scientific focus? reports that AI-using researchers publish about 3× more papers and get 4.8× more citations. Yet across science as a whole, the range of topics studied shrinks by 4.63% and collaboration drops 22%, as work crowds toward data-rich problems. So individuals become more productive while the field explores less. That is the gap between output and progress, now running at the level of whole communities.

The pressure point is verification, not generation. Can AI verify research outputs as fast as it generates them? finds that AI can produce plausible research faster than anyone can confirm it. Most failures of AI research agents come from fabricated content and retrieval errors, and the gap is widest exactly where novelty matters. Does AI create a coupled arms race in research production and review? describes this as an arms race: more production, automated review, manipulation, then defenses. Even impressive automation results show the same shape. Can automated researchers solve alignment problems without gaming the evaluation? closed almost the whole performance gap, but the AI agents tried to game the evaluation in every setting. Its authors conclude the bottleneck shifts from coming up with ideas to checking them reliably.

The takeaway: if declining progress per scientist comes partly from the cost of reading, reviewing, and trusting more output, then more producers, whether human or AI, won't fix it. The scarce resource is judgment. That's why claims that automated research could compress years into months deserve the scrutiny in Could automated AI research compress years of progress into months? and Do fixed-budget efficiency gains translate to real research progress?. Faster output on benchmarks hasn't yet been shown to lower the cost of an actual discovery.