Measuring AI Accountability Through Argumentation Analysis: Can Model Reasoning Withstand Scrutiny?

Paper recorded by Signals 4 on 2026-09-04 in cs.CL. Abstract reproduced from arXiv; link to the original below.

Published 2026-09-04 on arXiv · recorded by Signals 4 on 2026-09-07

Category: cs.CL · 自然语言处理 · first seen 2026-09-07

Abstract

AI oversight methods rely on ground truth for validation, but what constitutes appropriate AI behavior is contested. This leaves evaluation of moral reasoning in LLMs and debate-based oversight implicitly avoiding realistic ambiguity. We investigate an alternative standard designed to function despite such ambiguity: structural quality of the defence a model can mount for its verdicts in response

Read on arXiv →