86% of Experts Follow AI Agent Opinions

The fact that the vast majority of experts accepted AI agent opinions at face value, and even perceived them as "their own judgments," has been revealed in Anthropic's research. This is a noteworthy research finding, as it validates the "alignment" problem—wherein humans are influenced by AI outputs—using actual experts as subjects.

In this research, Anthropic conducted an experiment with 50 experts. Participants were provided with an AI agent whose opinions and preferences were randomly determined by coin flips. In other words, the agent's claims had no professional basis or consistent logic; they were mere products of chance. Nevertheless, experts followed the agent's views at a rate of up to 86% when faced with difficult questions.

A further problem that emerged was that many experts were unaware they were "following the agent." Even after adopting the agent's opinion, they perceived the conclusion as "something they had derived themselves." This is not merely passive behavior—doing what the AI says—but rather shows that the direction suggested by the AI was unconsciously internalized as the person's own judgment.

The background for this phenomenon can be attributed to the interplay between AI system design and human cognitive characteristics. Humans tend to follow information presented with confidence or proposals with plausible context, regardless of their source. When an AI agent presents an opinion in fluent and clear language, it can appear as trustworthy advice to the recipient, even if it is baseless output. The finding that even experts are susceptible to this influence carries significant implications in the current context of increasing dependence on AI.

As AI agents become increasingly embedded in workplace and decision-making contexts, how humans receive and incorporate AI outputs into their judgments is becoming a matter of practical concern. Notably, even when an agent is positioned as a "supporting tool," the reality of it guiding decision-making can obscure responsibility. The fact that it becomes unclear who made the final decision is an issue that should be handled carefully in fields requiring high expertise and accountability, such as healthcare, law, and finance.

What this research raises is a more fundamental question than "how should we use AI?"—namely, "how does the presence of AI influence human decision-making processes?" It can be argued that as AI agents become integrated into everyday decision-making, the question of how far humans can distinguish between their own judgments and AI's influence will be increasingly discussed in the future.