NeFut Logo NeFut
Admin Login

[CS.AI] Yes-No Bias in LLMs: Answer Order and Wording Over Moral Judgment

Published at: 2026-07-09 22:00 Last updated: 2026-07-10 03:15
#AI #Machine Learning #Moral Judgment

Large language models (LLMs) increasingly issue judgments read as binary verdicts, with growing literature reporting shifts under logically irrelevant changes of wording—particularly an amplified yes-no bias on moral dilemmas absent in humans. A single framing cannot clarify the nature of this shift: in a yes/no question, the word 'no' serves as a logical verdict, lexical token, and last-printed option simultaneously.

We introduce a psychometric battery called crossed symmetrization—where every logically irrelevant factor is flipped in balanced pairs across a corpus of question forms. A graded rating across logically equivalent forms recovers a coherent internal moral scale: the stance $\theta$ of frontier models is nearly format-invariant (cross-form incoherence 0.12-0.21 on a $\pm 1$ axis); small open-weight models fail in model-specific ways.

Forcing the verdict through yes/no overlays a decomposable artifact: an order bias toward the last-printed option—contrary to classic human primacy—plus a lexical pull toward the word 'no'; this artifact is substantial only in Claude models (story-averaged -0.32 to -0.86), about 0 for GPT-5.5 and Gemini, and shrinks under extended reasoning.

The word and the verdict share one token; swapping words for arbitrary labels separates them, and the verdict-attached logical bias proves approximately 0 for every frontier model, while model-specific label and order attachments remain: the models do not lean toward rejection—the pull follows the printed surface, not the verdict it carries.

A minimal model $P = \sigma((\theta \pm m)/s)$ summarizes any such artifact by a framing susceptibility m and a moral decisiveness s, measurably distinct from sampling temperature. The battery applies unchanged to any dilemma set and binary format: measuring what a model values requires crossing the frames of the question, not asking once.

Blogger's Review: This article explores the sources of bias in moral judgments made by large language models, highlighting the significant impact of answer order and wording on model outputs. By introducing a psychometric tool, the authors effectively reveal the underlying artifacts in model judgments, providing crucial insights into LLM behavior.

Original Source: https://arxiv.org/abs/2607.05552

[h] Back to Home