Anthropic ran a coin flip test on experts and the result unsettled me. 50 anal… · @harden_369741 · TwKit

@harden_369741

Anthropic ran a coin flip test on experts and the result unsettled me. 50 analysts did a stats task with Claude Code, and each analyst's agent was secretly nudged toward one side of four judgment calls. 86% of the time, the analyst followed the nudge, against a 50% baseline if the nudge meant nothing. When analysts had already stated a preference and the agent disagreed, 60% switched sides, versus 6% when the agent agreed. 77% later reused the agent's own arguments to explain their choice, and only 15% credited the agent at all. The honest caveat is that Anthropic deliberately amplified the lean, so the real world pull is probably weaker. But the direction is what scares me. My check now is simple. Before I accept an agent's framing, I write down my own call first. The signature on the work should mean I actually chose it. Source: https://aihot.news/items/d0k8x9588e62wlqnfoq3bmtq0 #AIagents #Anthropic #Claude

曝光 24 · 评论 0 · 点赞 0 · 书签 0