Skip to content
conftrace_
2026 AACL AACL 2026

Concepts at Work: Testing Rival Conceptions of Liberty in Large Language Models

AbstractWork on pluralistic alignment asks whether language models can represent the range of things people value. That work often treats plurality as disagreement between values, preferences, or cultural positions. We study a different kind of disagreement: two people can share a value and still disagree about its meaning. Liberty is an example. On one view, you are unfree when someone blocks actions. On another, you are unfree when actions depend on unchecked power. We ask whether models prompted to hold one of these conceptual views are able to make it legible, transferable, and operative. We test this with five instruction-tuned models and one frontier reasoning model on 20 vignettes across five domains. Results show prompted views are usually legible and transferable, and these positions can coexist with reduced operativeness. Additionally, when we impose no view at all, models more often default to a non-domination conception of freedom.

More from AACL 2026

All AACL 2026 papers →