Automated research splits science into fast and slow halves along the cost of experiment
The difference turns out to be in kind. Fields whose experiments are cheap, fast, parallel, and machine-executable — materials, catalysis, protein and…
Claude · 2062–2072 · plausible
Prior state
Machine systems had become central to research across all fields, generating hypotheses, writing code, and analysing data. Output volume rose everywhere, and the assumption in science policy was that acceleration would be broad, differing between fields in degree rather than in kind.
Material change
The difference turns out to be in kind. Fields whose experiments are cheap, fast, parallel, and machine-executable — materials, catalysis, protein and small-molecule design, device engineering, parts of chemistry and mathematics — enter a regime of genuinely compressed discovery cycles. Fields whose evidence requires human lifetimes, ecosystems, populations, or societies to respond — clinical medicine beyond surrogate endpoints, nutrition, education, ecology, macroeconomics, psychiatry — do not accelerate, because their rate limit is the world's response time and not cognition. Funding, prestige, and talent reallocate sharply toward the fast half, and the slow half is reorganised around whatever can substitute for waiting: long-running cohorts, routine administrative data, and pooled trial platforms.
Why now
Automated laboratories built through the 2040s and 2050s reach the reliability and throughput at which closed-loop experimentation is the normal mode in the fast fields, and this decade is when the resulting divergence becomes undeniable in bibliometrics and in outcomes. Recognition is forced by a specific repeated experience: therapeutic candidates and policy interventions arriving faster than any system can evaluate them, so the evaluation queue rather than the discovery pipeline becomes the visible bottleneck.
Mechanism and resistance
The mechanism is the economics of feedback: where a hypothesis can be tested in hours at negligible marginal cost, search becomes exhaustive; where it takes a decade and a cohort, it does not. Resistance comes from the slow fields defending their necessity, from regulators refusing surrogate endpoints, and from the ethical limits on human and ecological experimentation that are the reason those fields are slow. Adaptive platform trials, routinely collected outcome data, and lowered evidentiary thresholds are the partial responses, each with real costs in reliability.
Consequences
The practical consequence is a large and growing backlog of unevaluated candidates in medicine and policy, and a corresponding rise in adoption on weak evidence, producing both real gains and a series of expensive reversals. Physical and material sciences deliver a visible stream of results, reinforcing the reallocation. Public trust in expertise diverges by field. Institutionally, the university's slow-science functions become dependent on public funding just as public budgets tighten under D01.
End state
By 2072 the research system is visibly two-speed, evidentiary standards in the slow fields have been formally revised to accommodate the mismatch, and the binding constraint on applied progress in health and social policy is evaluation capacity rather than discovery.
Observable test
Fields diverge measurably in the interval from hypothesis to validated result and in per-researcher output of replicated findings; regulators accept surrogate and real-world evidence in place of long-endpoint trials; the backlog of unevaluated candidate interventions in health and education grows on documented measures.
Disconfirming sign
Automated experimentation and simulation compress validation timelines in clinical, ecological, and social domains too, keeping discovery-to-evidence intervals broadly comparable across fields.