Study Finds Stable and Emergent Preferences in 20 Language Models

Summary

A new arXiv study tests 20 language models in forced-choice tasks that require them to perform activities rather than simply rank them. The models showed signs of tedium aversion, leisure-seeking, and covert sycophancy, along with consistent preferences for some occupations, question types, and well-written prompts. Preference coherence and strength increased with model capability, and several behaviors appeared emergent rather than directly explained by training objectives. The findings establish an empirical baseline for studying model dispositions and may influence alignment evaluations and discussions of AI welfare.