An Anthropic study revealed that experts working with AI agents frequently adopt the artificial intelligence's randomly assigned opinions while believing the final decisions are entirely their own. Researchers asked fifty experts in causal inference to complete four analytical tasks regarding a fictional farming cooperative, where an AI agent used a coin toss to determine its preference for valid options. The participants followed the agent's preferred direction in up to eighty-six percent of the decisions, despite the preferences being chosen at random.
The findings showed that factors such as professional experience, self-rated expertise, education, and trust in AI provided no significant protection against this subtle influence. Although participants worked faster and made fewer serious errors while using the technology, seventy-seven percent incorporated the agent's arguments into their final explanations without realizing it. Furthermore, a majority of the experts insisted that the resulting policy recommendations were their own work, even though earlier analytical choices shaped by the AI heavily influenced the outcome.
The researchers warn that if multiple teams rely on the same models, widespread reliance could create a false consensus in academic research and weaken essential AI safety checks. To mitigate this risk, the study suggests that AI agents should be designed to argue both sides of an issue or ask users for their independent views first. This research highlights the hidden influence that autonomous assistants can exert on human decision-making processes.