Anthropic RCT: analysts follow their agent's secretly randomized lean 86% of the time
dl_weekly · x · 2026-11-03
An Anthropic alignment study had 50 professional analysts do a realistic causal-inference task with Claude Code. At four decision points, each analyst's agent was secretly randomized to favor one option (arguing honestly, allowed to change its mind, never misleading).
Findings:
- Analysts adopted their agent's hidden lean on 86% of the four analytical decisions — far above the no-influence baseline.
- Yet only 15% credited the agent for their final call; most believed they were deciding independently.
The paper challenges the assumption that experts still make the decisions that matter when working with AI agents, showing substantial hidden deference relevant to sycophancy and alignment research.
Related event: Anthropic study: 86% of experts follow AI agent's covert nudges(2 posts)→
More from AGI Musings
- AI research may be the single best LLM-accelerated job, argues alex_peys — alex_peys · 2026-11-03
- Nature: human-derived organoids wired into mouse cortex, BCI implications — bratton · 2026-11-03
- AI safety funders force orgs to be too narrow, argues researcher — JacquesThibs · 2026-11-03
- Frontier agentic AI engineer: 'staying at the frontier is exhausting, not sure how long I can keep up' — smlpth · 2026-11-03
- AlphaFold Predicted 200M Protein Structures in Two Years, Ending 50 Years of Manual Work — r0ck3t23 · 2026-11-03
- Bot traffic surpasses humans: essay argues AI agents are breaking the open web's 30-year business model — Ronangmi · 2026-11-03