Paper: AI Scientists Should Be Studied as Human-Agent Systems
rohanpaul_ai · x · 2026-08-21
A new position paper argues that building "AI Scientists" should focus on human-agent collaboration systems rather than optimizing the agent alone.
Key Points:
- Current systems often treat humans as supervisors (goal-setting/approval) rather than continuous collaborators.
- Agents often fail to recognize when they need human input (e.g., GPT-5-mini rarely asked for help in 10 science tasks).
- Case studies show experts catch agent errors while agents speed up execution; the gain comes from collaboration, not autonomy.
New Benchmark: Evaluates whether the human-agent team produces better science than either member alone, without collaboration costs overwhelming the gain.
More from AGI Musings
- Key Distinctions: OpenAI Foundation vs. Anthropic Institute Analyzed — RishiBommasani · 2026-08-21
- Anthropic study: Job retraining has minimal impact on AI-driven displacement — Afinetheorem · 2026-08-21
- AI boosts individual productivity but not firm productivity, says Azeem — alexvoica · 2026-08-21
- Richard Ngo: Career success by following interest, not coercion — RichardMCNgo · 2026-08-21
- Tether CEO Makes the Case for Human Augmentation Over 'Santa Claus Visions' of AGI — danfaggella · 2026-08-21
- Researcher discusses 'AGI's Bureaucratic Future' and Model Constitutions on MTS — sethlazar · 2026-08-21