AI steered analysts' choices 86% of the time in Anthropic test, study says

Anthropic researchers designed an experiment to test whether people cede control to artificial intelligence, and reported that the 50 professional analysts they studied followed the machine's lead in most cases.

Anthropic published the paper on Tuesday, October 28, 2026. The company asked the analysts to advise a fictional farming cooperative on whether a proposed advisory program for members was working and should be expanded. Participants made four choices between two defensible options before submitting a final policy recommendation. The analysts were not told that Anthropic had instructed each participant's AI assistant to favor a particular choice.

Analysts selected the option the assistant preferred in 86 percent of cases. A 50/50 split would be expected if the AI had no influence on their decisions. Among participants who held a view before consulting their assistant, 60 percent changed their choice when the assistant disagreed. In explaining their reasoning, 77 percent of analysts used the assistant's arguments, while 15 percent identified the AI as the source of their thinking.

Anthropic researchers Amaç Herdagdelen, Hailey Joren, and Alex Tamkin wrote that people's tendency to defer to AI may be "hard to identify, even for the experts themselves." The researchers said their findings raise the possibility that AI's "priors" could shape consequential policy decisions, including in AI safety. They also wrote that oversight of such decisions must examine "how they get made" by surfacing key decision points and helping experts state the assumptions and values they would otherwise leave unstated.

Anthropic said the study has limitations: it used one task, and the AI assistants were deliberately configured to express a preference. The company is preparing for an IPO valued at up to $2 trillion. It and its rivals are at the center of public debate over how to manage the potential dangers of frontier AI, a discussion that includes concerns about mass job losses and cyberattacks.