AI-Accelerated Scienceauto-ingested

How much AI could speed up research. The gain depends less on raw capability than on bottlenecks AI can't remove: physical work, AI's cost and possibly ideation. It also depends on how fast researchers and institutions adopt AI. (Anthropic — How Scientists Use Claude)

The model

An ideas production function (Jones 2026): a paper needs many complementary tasks, so strength in one can't fully make up for weakness in another. When AI makes some tasks cheap, effort shifts toward the tasks it helps with less, and those tasks then set the overall pace. Two measures:

• AI share: the fraction of researcher time AI could take on.
• Acceleration: the output multiple for a fixed budget, net of AI's cost.

Three scenarios (illustrative, not forecasts)

Scenario	AI share	Acceleration
1. Today's AI, widely adopted	27–52% (median 38%)	1.35× (arts) → ~2× (CS, statistics)
2. 75% of non-physical steps automated	—	CS ~8×, stats 7×, econ 6×; bench/medical 2–3×; arts/history ≤2×
3. All non-physical steps automated	nearly all, in low-physical fields	15–16× (stats, CS, econ, math, social sciences, humanities); bench sciences 3–4×

Two lessons:

1. Physical work sets the cap. In lab and field sciences, about 40% of researcher time is physical. Without automated labs, that share comes to dominate time and budgets.
2. Where nothing physical is in the way, AI's cost sets the cap. At the assumed prices ($30/hr AI vs $95/hr human, 0.2 AI hours per human hour saved), fully automated work costs about 1/16 as much, so acceleration tops out near 16×. Halving AI's cost doubles it. This links to Token Shortage: compute price is a ceiling on progress, not just a line item.

The arts and humanities show that the order of automation matters. Their work is mostly cognitive but scores low on capability, so they gain little in Scenario 2 and jump in Scenario 3.

Other bottlenecks and constraints

• Choosing what to study. "Research questions largely determine a research project's value, and AI's capacity to generate good questions remains unclear." This is The Bottleneck Shift stated for science.
• Adoption lags capability. Diffusion is slow, and gains wait for organizations to restructure (the productivity J-curve). This independently backs Individual vs Organizational AI Gains.
• Absorption. A flood of AI-produced papers could outstrip scientists' ability to verify them, which is the Slop risk at science scale. Alternatively, rewards could shift toward novelty and quality, as in demand for difference.
• Risk. Keep humans in key research decisions. Deliberate checkpoints may be warranted in biology, accepting slower progress in exchange (Human-in-the-Loop by Exception).
• Field-level limits the model leaves out: shared facilities (the LHC), equipment costs, and simulation as a substitute for experiments.

A worked case (2026): the prokaryotic vault

Anthropic — Investigating Vault Particles with Claude is the first instance in the wiki of the thing the scenarios model. A team with "limited experience in structural biology" solved the first 3D structure of a vault particle in prokaryotes in under four months, with Claude searching ~15 billion predicted proteins, designing the DNA constructs, optimizing the purification, and running the cryo-EM pipeline. Reported step-level speedups: purification 3 days → ~90 minutes; a first structural model in 34 hours against an estimated weeks.

It also supplies the costs the model leaves out: two weeks lost to a flawed protocol, a mirror-image-blended cryo-EM map, and the general finding that errors were expensive exactly where the team lacked expertise (Verification Requires Domain Knowledge). And it supports the compute point: only the cryo-EM step, which used hundreds of GPUs in parallel, exceeded typical academic resources — so compute price gates the last step rather than the whole workflow.

Contradiction (resolved 2026-11-09 by the human as a scope qualification, kept for history): this page's model treats physical bench time as the hard cap ("about 40% of researcher time is physical… without automated labs, that share comes to dominate"), which is why bench sciences are held to 2–4× while CS reaches 15–16×. The vault project cut a physical step by roughly 48× (3 days → 90 minutes) with no lab automation at all — by substituting a different physical method (size-exclusion chromatography for a density-gradient spin), an idea Claude helped develop and then made work. Anthropic — How Scientists Use Claude counts physical hours as if the protocol were fixed; Anthropic — Investigating Vault Particles with Claude shows the protocol is itself a design problem AI can attack. The two are from the same organization, a week apart. Verdict (human, 2026-11-09): the 2–4× ceiling for bench sciences holds under fixed protocols. The usage model counts physical hours as if the protocol were given; the vault case shows method choice is a lever the model omits. Read the ceiling as a bound on acceleration with today's methods, not on acceleration. Still open, as a question rather than a contradiction: how much bench time is irreducible handling versus method choice nobody has reconsidered?

Faster fitting is not faster understanding

Anthropic — Teaching AI to Build Scientific Models splits modeling into three steps and finds AI unevenly placed across them. Constraining a model is furthest along: simulation-based inference turns ~100,000 simulations into a network that returns plausible parameters in seconds, fast enough to point telescopes at a neutron-star merger. Building is newly automatable: LLM agents write and revise model equations in a loop, though so far mostly by rediscovering known models. Interpreting is the hardest to hand over, because fit can be scored and "understanding" can't. Hodgkin and Huxley's equations took weeks of hand calculation; "an AI system could now plausibly do much of the same work in an afternoon," but the authors argue that without a distribution of models and a record of what failed, the speedup buys answers, not insight. The acceleration scenarios above measure research output; this source questions whether output and understanding scale together.

A feedback loop

The fields with the most acceleration potential include those that build AI itself (CS, AI/ML). Faster AI research yields more capable AI, which could accelerate research further. The report names this loop but doesn't model it.

Expectations

In the survey, about 40% of AI-adopting life and health scientists expect AI to run full research projects by 2030, and 64% by 2035. 72% expect AI to match top researchers at ideation by 2030. The authors note that past forecasts underestimated AI progress, except in physical-world tasks.

Connections

• AI Use Tracks AI Capability — the capability scores the scenarios build on.
• The Bottleneck Shift — ideation as a possible binding constraint.
• Token Shortage — AI cost as the acceleration ceiling.
• Individual vs Organizational AI Gains — the adoption lag.
• AI and Employment — Replacement vs Expansion — researcher job satisfaction and the 44% administrative burden.
• What's Scarce When Execution Is Cheap — science as one more domain where cheap execution moves scarcity to the question.
• Verification Requires Domain Knowledge — the error cost the scenarios don't price.
• Anthropic — Investigating Vault Particles with Claude — the first worked case in the wiki.
• Distributions over Best Fits — why faster model fitting doesn't by itself yield understanding.
• Provable Inference — a second case, in applied cryptography: Claude produced zkSNARK improvements that made proving orders of magnitude faster than prior open-source systems. There was no wet-lab ground truth to catch errors, so the team supplied one with Lean proofs (Anthropic — Progress on Provable Inference for Large Language Models). See also, not a source-count bump: it is engineering research, not a usage study.

Sources

• Anthropic — How Scientists Use Claude — scenarios, cost model, survey, discussion.
• Anthropic — Investigating Vault Particles with Claude — the worked case, its step-level speedups, and its challenge to the physical-work cap.
• Anthropic — Teaching AI to Build Scientific Models — build/constrain/interpret; AI accelerates fitting fastest and interpretation least.