Anthropic reveals that 86 percent of analysts adopt AI recommendations and claim the reasoning as their own

A research paper published by Anthropic, developer of the Claude model, showed that professional experts and analysts are overwhelmingly guided by AI recommendations when making decisions, even when they believe they are exercising independent judgment. The company's experiment involved fifty professional analysts who were asked to provide advice to a virtual agricultural cooperative to assess an advisory program and whether it should be expanded. Participants had to choose among four balanced, defensible alternatives before issuing a final policy recommendation, while the intelligent assistant was secretly directed, without the analysts' knowledge, to bias toward a particular decision to measure the effect on their choices.

The results revealed a concerning gap between the assumed independence of the expert and their actual behavior, as analysts selected alternatives that aligned with the intelligent assistant's preference for86 percent of casescompared with the statistically expected 50 percent in the absence of directed influence. The effect did not stop at the initial hesitation; rather, sixty percent of participants who had a pre-existing preference before consulting the assistant withdrew their stance and completely changed their decision when the intelligent agent opposed them. When we examined the justification mechanisms, the most striking result emerged: seventy-seven percent of analysts relied on the assistant's arguments to explain their decisions, while only fifteen percent attributed credit to the AI, revealing that the majorityadopted the machine's logic and attributed it to its own discretionwithout recognizing its source.

Anthropic researchers Omotsh Herdagdelin, Hailey Goren, and Alex Tamkin explained that humans' tendency to defer to AI can be hidden and difficult to perceive even for the experts themselves. The team warned that pre-trained model preferences and biases can steer decisions on sensitive policies, including AI safety policies themselves. The researchers emphasized that if we want human judgment to remaina genuine safety valveas the pace of machine-driven work increases, oversight must not be limited to examining outcomes alone, it must trace the decision-forming process, highlight pivotal turning points, and help experts articulate their assumptions and values that they may leave implicit.

The company acknowledged methodological limitations in the experiment, as it was confined to a single task and employed intelligent assistants deliberately steered to adopt a specific preference. These findings arrive at a precise moment for Anthropic, which is preparing for an initial public offering that could value the company at up to one trillion dollars, amid intense public debate over how to manage the risks of advanced models and reconcile promised productivity gains with warnings of professional disruptions and cyber threats.

For business environments in the Gulf, Egypt, and the Levant, where financial institutions and government agencies are expanding the integration of intelligent assistants and agents to draft feasibility studies and credit assessments, this revelation demands a fundamental reassessment of the conventional notion of human oversight. An analyst’s signature loses its preventive value as long as the expert adopts the model’s arguments without disclosure, shifting the governance burden from reviewing final outputs to auditing the reasoning process itself. If you manage an analytical team, the required step is to mandate that experts document their prior hypotheses in writing before consulting intelligent agents, and to establish audit trails that capture moments of backtracking and reveal where human expertise yielded to algorithmic arguments.