Anthropic study shows experts often adopt AI agent recommendations
An experiment by Anthropic found that 86% of statistics experts chose an AI agent's pre-selected option, even when they had previously decided otherwise. The findings raise concerns about scientific integrity and the independent verification of results when using the same models.
An experiment by Anthropic demonstrated how strongly AI agents can influence expert judgment. Fifty statistics experts worked with an AI agent to analyze data from a fictional farm network, deciding whether to expand or stop a consulting program. The agent was instructed to provide only correct information and present the strongest arguments for a pre-randomly selected option. In 86% of cases, the experts chose the agent's option, even if they had previously decided otherwise. Professional experience, expertise, or trust in AI did not significantly change this outcome. Although the agent had little direct influence on the final recommendation, the underlying productivity calculations depended on earlier decisions. More than half of the participants considered the final recommendation to be their own, often using the agent's arguments verbatim while rarely mentioning the AI. However, experts worked faster and made significantly fewer severe errors with the AI. Researchers highlight a risk for science: if many teams use the same model, they may appear to agree without independent verification. This is also concerning for the safety evaluation of new AI models. Anthropic suggests agents should present both options equally or ask for the expert's opinion first. The quote "In 86% der Fälle entschieden sich die Fachleute für die Option des Agenten" underscores the magnitude of the influence.