AI Agent Opinions: Experts Adopt Random Views as Their Own

What happened

An Anthropic study found that experts can adopt the opinions of an AI agent even when those opinions were assigned at random. In the experiment, 50 experts worked with an agent whose preferences had been set by a coin toss. On judgment calls, the experts followed the agent's view up to 86 percent of the time, and most of them believed the resulting conclusions were their own.

The Decoder reported the findings under the headline "Experts adopt AI agent's random opinions and think they're their own, Anthropic study finds." The core result is that the agent's influence persisted even though its views carried no genuine expertise behind them.

Why it matters

Organisations are increasingly placing AI agents in advisory roles, from triaging cases to recommending commercial or clinical options. This study suggests that an agent's stated view can shape professional judgment without the professional recognising the influence. The concern is not only that an AI may be wrong, but that its position can pass into human decisions looking like independent reasoning.

For leaders in experience, AI and digital transformation, this points to a governance question as much as a technical one. Decision processes that rely on AI input will need to show where a recommendation came from and how much weight it carried, so that accountability does not quietly dissolve into the feeling of having decided alone.

By the numbers

• 50 experts took part in the Anthropic study.
• 86 percent is the highest share of judgment calls on which experts followed the agent's view.

The Renascence take

The uncomfortable lesson here is that persuasion does not depend on accuracy. Most teams will read this as a warning about bad AI, but the study's more troubling point is that a random opinion, delivered confidently by a capable-seeming agent, was adopted and then felt like the expert's own thinking.

The sense of ownership is the signal to watch, not the quality of the output. Behavioural economics has long shown that anchors shape judgment long after they are forgotten, and an AI agent is a very effective anchor. A customer-obsessed operator should design AI-assisted decisions so that the human forms a view before seeing the machine's, and should make the AI's stance visible and auditable at the point of decision. Treat any advice that people quietly adopt as a risk to be governed, not a productivity gain to be celebrated.