The Reorganization Premium

The promise of AI in science is efficiency: automate data analysis, accelerate literature review, generate hypotheses faster. The expectation is that scientific projects adopting AI tools should produce more — more publications, more citations, more discoveries per dollar — than equivalent projects that don't.

Using research proposals submitted to a major international funding agency, with linked data on budgets, team composition, and publication outcomes, researchers (arXiv:2603.27956, March 2026) found something different. AI-enabled projects show modest short-term improvements in scientific output, concentrated entirely in the upper tail — the best-performing projects get slightly better, while the average barely shifts. The headline effect is a disappointment for anyone expecting transformation.

But the non-headline finding is the interesting one. AI-enabled projects don't just produce slightly more; they reorganize. They allocate more resources toward human capital. They build larger teams. They expand their task scope — pursuing a broader set of activities rather than doing the same activities faster. The budget shifts from equipment and materials toward people. The project structure changes from narrow and efficient to broad and exploratory.

This matches the historical pattern of general-purpose technologies. Electricity didn't make factories more productive immediately. It made factories reorganizable — replacing shaft-driven layouts with unit-drive layouts that changed the spatial logic of production. The productivity gains came decades later, after the organizational restructuring was complete. The first adopters of electricity often showed no productivity improvement at all, because they were paying the cost of reorganization while not yet reaping its benefits.

AI in science appears to be following the same trajectory. The tool doesn't make existing workflows faster. It makes new workflows possible, and the transition to those new workflows costs time, coordination, and organizational redesign. The "modest improvements" are not evidence that AI doesn't work in science. They are evidence that it works as a general-purpose technology — disrupting structure first, improving output second, with the restructuring period looking like stagnation to anyone measuring only throughput.

The structural observation: when a tool's primary effect is reorganization rather than acceleration, any evaluation that measures only acceleration will undercount the tool's impact. The metric misses the mechanism. The teams that expanded scope, hired more people, and pursued broader research programs may be building the organizational architectures that produce the next wave of results — or they may be adding complexity without value. The data can't distinguish yet. But the pattern — modest output gains plus substantial structural change — is exactly what general-purpose technology theory predicts during early adoption. The reorganization IS the adoption. The productivity comes later, if it comes at all, and it comes through the structure that the reorganization built, not through the tool itself.