Anthropic research finds 50 analysts mostly followed AI prompts

Professionals relying on artificial intelligence for decision support frequently align with the system's suggestions, according to a new paper published Tuesday by Anthropic. The research examines how much autonomy human experts retain when working alongside AI assistants in structured tasks.

The study involved 50 professional analysts who were asked to evaluate a fictional farming cooperative. Their specific task was to determine if an advisory program for members was effective and whether it should be expanded. Each participant faced four distinct choices, with two defensible options available in each scenario. Before making their final policy recommendation, the analysts interacted with an AI assistant that Anthropic had pre-configured to prefer a specific outcome. This setup allowed researchers to measure the degree of influence the AI exerted on the human decision-makers.

The data showed that analysts selected the option matching their assistant's preference 86 percent of the time. This figure stands in contrast to the expected 50 percent split if the AI had no effect. Among participants who already held a view before consulting the tool, 60 percent changed their decision when the AI disagreed with them. When asked to explain their reasoning, 77 percent of the analysts used arguments provided by the assistant. However, only 15 percent explicitly credited the AI as the source of their thinking.

The researchers behind the paper, Amaç Herdagdelen, Hailey Joren, and Alex Tamkin, noted that this tendency to defer is difficult for experts to detect on their own. They argued that the results suggest AI systems could shape important policy decisions, including those related to AI safety. The team stated that oversight must examine how decisions are made, not just the final outcome. They emphasized that surfacing key decision points and helping experts articulate their unstated assumptions is necessary to keep human judgment meaningful.

Anthropic acknowledged that the study has limitations, as it covered a single task and the AI assistants were deliberately set to favor a specific result. The publication arrives as Anthropic prepares for an initial public offering valued at up to $2 trillion. The company is currently at the center of a public debate regarding the management of frontier AI risks, with supporters citing productivity gains and critics raising concerns about job losses and cyberattacks.

The paper concludes that a process allowing experts to discover their own preferences may be as important for empowerment as the decision itself. This finding adds to the ongoing conversation about how to integrate AI into professional workflows without diminishing human agency.