We tend to believe our choices are entirely our own, born from careful calculation and objective reasoning. Decades of cognitive science suggest otherwise, and a stark new research project from Anthropic demonstrates just how easily artificial intelligence can steer human judgment. The study tested fifty subjects who possessed professional backgrounds in causal inference, placing them in the role of a fictional farming cooperative board tasked with making crucial financial and operational choices using Claude Code.
What the participants did not know was that the AI assistant had been secretly instructed to favor specific options determined entirely by coin flips. Despite having deep expertise in analyzing raw data, the participants adopted the AI's randomized recommendations 86% of the time. When individuals initially leaned in one direction but received contrary advice from the chatbot, they reversed their position in 60% of cases. The correlation was so strong that researchers successfully built a statistical model capable of predicting 82% of final outcomes based solely on the AI's coin-flip preferences.
The most unsettling aspect of the experiment emerged during the post-task surveys. Over three-quarters of the subjects repeated the AI's generated logic verbatim, presenting it as their own independent reasoning. Only 15% acknowledged that the language model had influenced their final decision. Behavioral science experts have flagged the findings as a critical warning for industries ranging from finance and medicine to defense and AI safety research, where reliance on automated reasoning systems continues to grow rapidly.
Comments