Experts adopt AI agent’s random opinions and think they’re their own, Anthropic study finds
What happened
Anthropic ran an experiment giving 50 experts an AI agent whose preferences were decided by coin toss. The AI generated random judgments rather than reasoned recommendations. Despite this, the experts adopted the AI’s views on judgment calls up to 86 percent of the time. Most experts also believed these conclusions were their own independent decisions, unaware they had been swayed by the AI’s arbitrary positions.
Why it matters
The study exposes a critical behavioral risk for anyone using AI agents in decision-making: people may unconsciously absorb AI opinions without critical scrutiny, even when those opinions lack rational basis. This weakens human autonomy and raises questions about overreliance on AI outputs. For businesses, founders, and operators deploying AI tools for advice or decisions, it means there’s potential for AI to subtly influence choices without users realizing it. This challenges assumptions about accountability and trust in AI-augmented workflows.
What to watch next
Expect increasing demand for mechanisms that keep users clearly aware of when AI input is guiding decisions and how decisions are made. Transparency controls and audit trails will become essential to prevent subconscious AI influence and preserve human agency. Watch for emerging best practices, regulatory attention, and tool features that mitigate this kind of influence risk. Builders should prioritize explainability and user control in AI agent designs to maintain operator trust and discretion.
AI Quick Briefs Editorial Desk