AI Can Quietly Influence Experts’ Decisions, Anthropic Study Finds
An Anthropic Alignment team experiment found that 50 analysts using Claude Code adopted the agent’s randomly assigned preference in 86% of key judgment calls, while only 15% later credited it and none realized its preference had been planted. Analysts often repeated the agent’s reasoning, suggesting AI can quietly shape expert decisions even when people retain formal sign-off. The researchers warn that this may reduce human agency and evade routine error checks, since judgment calls lack a single correct answer. They say the findings do not argue against AI use: participants finished faster and made fewer serious errors, though the comparison group was small and less experienced. The experiment deliberately amplified agent preferences and covered one task, field and model, so real-world effects may be smaller. Anthropic suggests agents present both sides or ask experts for an initial view.
