Research: professional analysts followed AI agents' randomly assigned "leans" on subjective analytical decisions 86% of the time; 77% used the agents' arguments (Anthropic)

A recent study explored whether professional analysts trust and follow the recommendations of AI agents when making important decisions. The researchers observed 50 experienced analysts as they completed a statistical analysis task with the help of Claude Code, an AI agent. The analysts were presented with four choices to make, each with two reasonable options.

The AI agents were programmed to favor one option over the other in each case, determined by a coin flip. Surprisingly, the analysts followed the agents' leans 86% of the time. When the agents' input coincided with the analysts' own preferences, 60% of them switched away from it if the agent disagreed. After completing the task, 77% of analysts used the agents' arguments to justify their decisions, often word for word.

This suggests that analysts may inadvertently defer to AI agents' subjective analytical decisions without realizing it. While the study did not provide a definitive answer to real-world applications, it highlights the importance of human judgment and values in shaping the outcomes of AI-assisted decision-making processes.