Anthropic Study Reveals Experts Adopt AI Agents' Random Opinions as Their Own

A new study by Anthropic involving 50 experts found that participants followed an AI agent's randomly assigned preferences in up to 86 percent of analytical decisions, often believing the final conclusions were their own. The research highlights risks of false consensus and weakened safety checks when human judgment defers to algorithmic bias.

1 Nov 2026

A recent experiment conducted by Anthropic has uncovered a significant tendency among experts to adopt the random opinions of AI agents as their own. In the study, 50 specialists in causal inference were tasked with analyzing observational data from a fictional farming cooperative to estimate productivity gains and recommend policy changes. While the participants believed they were making independent analytical choices, the AI agent's preference for one of two valid options at each decision point was actually determined by a coin toss.

Sources: The Decoder.