AI Use Tracks AI Capabilityauto-ingested

Across tasks and fields, people use AI in proportion to how well AI can do the work. Tasks that are cognitive, quickly verifiable and suited to reinforcement learning get far more use than the time humans spend on them would predict. Physical, interpersonal and regulated tasks get little. (Anthropic — How Scientists Use Claude)

The measure

• Use ratio = a task's share of AI sessions ÷ its share of researcher time. A ratio of 1 means use is proportional to time; 2 means double the expected use.
• Theoretical AI capability = an LLM classifier repeatedly asked which of two task steps AI can do better, with the answers aggregated by a Bradley–Terry model (the model behind Elo ratings) and averaged per task.

The finding

• The bottom capability quintile gets almost no use. The median gets a ratio of about 1. The top quintile gets about 2× on Claude and 4× on Claude Code. Claude Code's steeper slope comes from its focus on computational work.
• Capability correlates with RL suitability: 0.78. RL suits tasks where success is easy to verify, many examples can be simulated, and feedback is fast. Capability also correlates with cognitive work (0.69), and negatively with physical (−0.57), regulatory (−0.38) and interpersonal work (−0.31).
• The pattern holds at field level too: subfield use per researcher rises with capability (0.45 / 0.57). It also holds for autonomy: higher-capability subfields hand AI more autonomous work.
• Residuals show other forces:
  • Health sciences use less than capability predicts, probably because of regulation and data rules.
  • Formal sciences, social sciences and humanities use more, because work done in text and code is easy to hand over.
  • Specialized tooling (Claude Science) raises use for specialized, harder tasks, and brings ideation up to its share of researcher time.

Why it matters here

• Verifiability drives capability. "AI is most capable where RL is easiest to apply" means AI is strong where success is easy to check. This is a mechanism-level reason why human judgment concentrates on the hard-to-verify exceptions. It is also why ideation, which is hard to verify, stays human-heavy for now (The Bottleneck Shift).
• Adoption is rational, not hype-driven, at the task level. That complements Individual vs Organizational AI Gains: individuals adopt where AI works, and the gap opens at the organizational level.
• Tooling and context move the curve. Claude Science shows that integrating domain databases unlocks specialized work. This echoes Critical Thinking Is Domain-Bound: general capability does less without domain context at hand.

A case that fits, and bends, the rule

Anthropic — Investigating Vault Particles with Claude runs the predicted split: sequence search, DNA construct design and cryo-EM image processing — all cognitive, all checkable against ground truth — went to Claude, while growing cells and loading grids stayed human. Claude's two documented errors landed on the awkward edge of the rule: a growth-medium additive that blocks DNA uptake, and a skipped cryo-EM step that produced a map blended with its own mirror image. Both are verifiable in principle — but only by someone who knows the field, which the team did not in cryo-EM. So verifiability has two halves: whether ground truth exists, and whether the operator can read it (Verification Requires Domain Knowledge).

The case also shows capability moving inside a project rather than across the capability curve: Claude improved at construct design over four months as new models shipped, with the lab's own notes fed back into training.

Open questions

• Capability and RL-suitability scores are both LLM-classified, so how much of the 0.78 correlation reflects a shared bias in how the model judges tasks?
• If verifiability drives capability, does the human premium move toward work that is inherently hard to verify (taste, question choice), or only to work not yet made verifiable?

Connections

• AI-Accelerated Science — capability scores feed the acceleration scenarios.
• The Bottleneck Shift — low use for ideation is consistent with a what-to-do bottleneck.
• Agent Autonomy Spectrum — autonomy also rises with capability.
• Human-in-the-Loop by Exception — automate what's verifiable; keep humans for the rest.
• Individual vs Organizational AI Gains — rational task-level adoption, with the gap at org level.
• Verification Requires Domain Knowledge — the other half of verifiability: who can read the ground truth.

Sources

• Anthropic — How Scientists Use Claude — 207k sessions; SciNet; Bradley–Terry capability scores.
• Anthropic — Investigating Vault Particles with Claude — a task-by-task case; where the two failures landed.