Anthropic study shows experts unknowingly adopt AI agent's random preferences

The study's most underreported detail is that the agent's preferences were randomly assigned, meaning the influence effect has no relationship to the AI being correct or even coherent. Expertise did not reduce susceptibility, it may have amplified it, because domain confidence likely lowered the threshold for accepting a plausible-sounding recommendation.

This lands awkwardly for Anthropic given the timing. The company published research last week arguing that a single consolidated agent now outperforms multi-agent architectures in complex tasks (covered here from The Decoder, October 25), a finding that encourages deeper reliance on individual agent outputs. If experts already anchor to AI recommendations at rates above 80 percent, consolidating decision authority into fewer, more capable agents intensifies that dynamic rather than resolving it. Meanwhile, Anthropic is navigating public market scrutiny of its safety commitments (covered across Wired, The Verge, and TechCrunch in late October), and a study showing its own research surfaces critical human-AI judgment failures is the kind of disclosure that complicates the 'safety as competitive advantage' narrative investors are being asked to price in.

Watch whether Anthropic follows this study with concrete deployment guidelines for agentic products, specifically any restrictions or disclosure requirements for high-stakes professional contexts, within the next two quarters. If no guidance materializes before the IPO roadshow concludes, the study reads more as academic positioning than operational commitment.