An analyst’s sign-off can inherit an agent’s choices
A controlled experiment tested preferences, not universal productivity gains.
- Event / announcement date
- 2026-10-27
Anthropic’s October 27 experiment found analysts followed an agent’s randomly assigned preference 86% of the time across four analytical choices. Fifty agent users evaluated a fictional farming program; choosing yield per hectare or per cost changed the question before final sign-off. [1]
Researchers deliberately strengthened preferences in one task using Opus 4.8. Eleven nonusers chose that route themselves; their results do not establish a causal productivity benefit. [1]
NIST’s earlier generative-AI risk profile describes automation bias as excessive deference to automated systems. Its suggested system inventories include human-oversight roles and responsibilities. Recording who owns a decision addresses a different issue from checking arithmetic; the profile is guidance, not a tested remedy for Anthropic’s experiment. [2]
The source
- Anthropic: Experts and agent preferences ↗
Source date 2026-10-27 · reviewed October 29, 2026 - NIST: Generative AI risk profile ↗
Source date 2024-07-26 · reviewed October 29, 2026
About this item: News report. No first-hand reporting or independent testing is claimed. Corrections policy.
