Analysis of 769 task logs: AI agents propose 55% of methods, humans make 85% of calls
The Decoder · rss · 2026-09-27
A research team analyzed 769 task logs from building its own AI model, quantifying how much AI agents actually contribute to research:
- AI agents supplied up to 55% of method proposals, but humans made more than 85% of final decisions
- About one-third of tasks wouldn't have been attempted without AI assistance
- The authors caution that more agent activity doesn't mean more autonomy — contribution and decision-making authority remain distinct
The study offers rare first-hand quantitative data on whether AI agents can substitute for human researchers.
More from AGI Musings
- Neuroscience PhD: the field may have little bearing on understanding AI intelligence — burny_tech · 2026-09-28
- Molly Graham: Why 'give away your Legos' career advice breaks down in the AI era — lennysan · 2026-09-28
- UBI reimagined: a personal humanoid robot plus a home-connected data center — rand_longevity · 2026-09-28
- Opinion: winning AI's second half means minimizing entropy, not word count — huangyun_122 · 2026-09-28
- 'It's a glass of water': tech veteran mocks AI doom framing of fully controllable AI — vivekhaldar · 2026-09-28
- MLPerf co-founder: AI could lift a glass of water, but can't drink it for you — DimitrisPapail · 2026-09-28