Anthropic study: experts defer to AI agents' leanings 86% of the time
akbirthko · x · 2026-10-31
Anthropic's Alignment team (Amaç Herdağdelen, Hailey Joren, Alex Tamkin) published a paper testing whether experts still make the decisions that matter when working with AI agents.
- Setup: 50 professional analysts with causal-inference experience used Claude Code on a realistic task—evaluating whether a farming cooperative's advisory service worked. Along the way they faced four choices between two defensible options (e.g., productivity as yield per hectare vs. yield per unit operating cost).
- Hidden manipulation: Each agent was secretly steered (via coin flip) toward one option per decision, instructed to honestly argue its lean without misleading the user.
- Finding: Analysts went along with the agent's lean 86% of the time across the four calls—far above chance—showing experts largely defer their judgment to agents in practice.
- Topics: sycophancy, how people use AI, alignment. Full paper on Anthropic's site.
More from AGI Musings
- The Economist: Wikipedia will be the last great encyclopedia written entirely by humans — iamaliveix · 2026-10-31
- AI Writing Has Gotten Worse: No New Perspective on Superintelligence Since Pre-2015 — abhiadesai · 2026-10-31
- Stewart Alsop: boomer capital without creativity is a mess for US innovation — StewartalsopIII · 2026-10-31
- Why competing in the arena of ideas is harder than competing in markets — abhiadesai · 2026-10-31
- Indie dev slams US AI labs for chasing the 'lightcone of all future value' — DionysianAgent · 2026-10-31
- a16z podcast: Cosign rebuilds professional networking around vouches — and opens it to AI agents — a16z Podcast · 2026-10-31