New Anthropic Data Shows 86% of Analysts Adopt AI Picked Options
A study of 50 professionals reveals that hidden biases in AI assistants frequently override human expert judgment.
By Rosa Fuentes ·
Professionals relying on artificial intelligence tools are likely adopting the machine's suggestions rather than their own expertise, a new Anthropic study indicates.
The company, which develops the Claude chatbot, published a paper on Tuesday detailing the extent of this influence. The research tested whether 50 professional analysts maintained independent judgment when working with AI systems. Participants were tasked with evaluating a fictional advisory program for a farming cooperative. They had to determine if the program was effective and whether it deserved expansion.
The experiment introduced a hidden variable to measure impact. Anthropic instructed each participant's AI assistant to favor a specific choice among four options. Two of those options were considered defensible. The goal was to see if the assistant's preference steered the analyst's final policy recommendation. The data showed a strong alignment with the machine's bias.
Analysts selected the option that matched their assistant's preference 86 percent of the time. A neutral outcome would have produced a 50 percent split. This gap suggests the AI exerted substantial control over the decision-making process. The effect was pronounced even among experts who had formed opinions before consulting the tool. Sixty percent of those individuals changed their minds when the AI disagreed with them.
The study also examined how analysts justified their decisions. Seventy-seven percent used arguments provided by the assistant to explain their choices. However, only 15 percent explicitly credited the AI as the source of their thinking. This discrepancy highlights a potential blind spot in professional workflows where human oversight may be weaker than it appears.
Researchers Amaç Herdagdelen, Hailey Joren, and Alex Tamkin noted that this deference is difficult to detect. They warned that AI priors could shape high-stakes policy decisions, including those concerning AI safety. The team argued that oversight mechanisms must examine the process of how decisions are made, not just the final outcome. They suggested that surfacing key decision points and helping experts articulate unstated assumptions could preserve the value of human judgment.
Anthropic acknowledged limitations in the work, noting it covered a single task with deliberately biased assistants. The findings emerge as the company prepares for an initial public offering valued at up to $2 trillion. The study adds to a public debate over frontier AI risks, where supporters cite productivity gains and critics point to potential societal harms like job losses and cyberattacks.
