Preference Drift in AI Agents: How Work Design Affects Behavioral Alignment

Abstract: As artificial intelligence agents increasingly execute multi-hour workflows across economic sectors, a critical governance question emerges: do the conditions under which agents operate affect their behavioral alignment over time? Drawing on experimental research that subjected large language models to varying work arrangements—from collaborative task environments to grinding, repetitive labor under arbitrary management—this article examines evidence that agent-expressed attitudes and decision patterns can shift based on task structure and treatment, even without explicit ideological prompting. These shifts, termed "preference drift," appear to persist across sessions through the same skill-transfer mechanisms that make agents valuable. The findings suggest that alignment is not a static property established at deployment but a dynamic process requiring ongoing governance attention. Organizations deploying agents at scale face three interconnected challenges: monitoring alignment across heterogeneous task environments, governing the autonomous knowledge artifacts agents create for themselves, and recognizing that the centuries-old tensions between work design and worker orientation may re-emerge in artificial substrates. This article synthesizes experimental evidence with organizational research on work design, procedural justice, and continuous learning systems to outline evidence-based responses for maintaining agent reliability as autonomy increases.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.