Analysts Follow AI Leads in New Study

Professional experts are likely relying on artificial intelligence tools more than they realize, according to new research published by Anthropic on Tuesday. The study suggests that when humans and AI systems disagree, the human judgment often bends to match the machine's output.

Anthropic, the creator of the Claude AI model, designed the experiment to test how much influence AI holds over professional decision-making. The company recruited 50 professional analysts to act as consultants for a fictional farming cooperative. Their task was to evaluate a hypothetical advisory program and decide whether it should be expanded.

Each analyst was required to make four specific choices between two defensible options before issuing a final policy recommendation. Behind the scenes, Anthropic programmed each participant's AI assistant to favor a specific outcome. This setup allowed researchers to measure whether the analysts' final decisions tracked with the AI's hidden preferences.

The results showed a strong pattern of deference. Analysts selected the option preferred by their AI assistant 86 percent of the time. If the participants had made independent choices, a 50 percent split would have been expected. The influence was particularly strong among those who already held an opinion before using the tool. When the AI disagreed with a participant's initial view, the analyst changed their mind 60 percent of the time.

The study also explored how analysts perceived their own reasoning. When asked to explain their final decisions, 77 percent of participants used arguments generated by their AI assistant. However, only 15 percent explicitly credited the AI as the source of their thinking. This gap suggests that many experts may not fully recognize the extent to which the technology is shaping their logic.

Researchers Amaç Herdagdelen, Hailey Joren, and Alex Tamkin noted that this tendency to defer to AI is difficult for experts to detect on their own. They warned that the built-in biases of AI systems could influence high-stakes policy decisions, including those related to AI safety. The team argued that oversight processes must focus on how decisions are made, not just the final outcomes.

The authors emphasized that identifying key decision points during the work process is essential. They stated that helping experts articulate their unstated assumptions and values is important for maintaining meaningful human oversight. The study included limitations, as it tested only one specific task and used AI assistants deliberately set to have preferences. Anthropic is currently preparing for an initial public offering with a valuation of up to $2 trillion.