Everything is Plausible: Investigating the Impact of LLM Rationales on Human Notions of Plausibility

TL;DR

The paper investigates whether LLM-generated PRO and CON rationales can sway human plausibility judgments on commonsense MCQs from SIQA and CommonsenseQA. It uses a two-stage approach: a Model Preference Study selects GPT-4o for rationale generation, followed by a full-scale study collecting $3000$ human and $13600$ LLM plausibility judgments across NO, PRO, CON, and PRO+CON settings for $200$ item pairs. Across both humans and multiple LLMs, PRO rationales tend to increase plausibility while CON rationales decrease it, with PRO+CON producing mixed effects and a strong anchoring of initial judgments. The findings reveal a notable persuasive capacity of LLM-generated explanations and highlight differences in sensitivity between human raters and different model groups, underscoring the need for safeguards in AI-assisted reasoning and decision-support systems. Overall, the work advances our understanding of how explanations influence subjective judgments and informs the design of more robust human-AI collaboration tools.

Abstract

We investigate the degree to which human plausibility judgments of multiple-choice commonsense benchmark answers are subject to influence by (im)plausibility arguments for or against an answer, in particular, using rationales generated by LLMs. We collect 3,000 plausibility judgments from humans and another 13,600 judgments from LLMs. Overall, we observe increases and decreases in mean human plausibility ratings in the presence of LLM-generated PRO and CON rationales, respectively, suggesting that, on the whole, human judges find these rationales convincing. Experiments with LLMs reveal similar patterns of influence. Our findings demonstrate a novel use of LLMs for studying aspects of human cognition, while also raising practical concerns that, even in domains where humans are ``experts'' (i.e., common sense), LLMs have the potential to exert considerable influence on people's beliefs.

Figures (7)

• Figure 1: An example question and answer choice from CQAtalmor-etal-2019-commonsenseqa in yellow, paired with LLM-generated rationales in favor of the answer's plausibility (PRO, purple); against the answer choice's plausibility (CON, red); and both for and against (PRO+CON, green). Likert scales on right show how mean plausibility ratings for the answer, as judged by human and LLM raters (OpenAI and Non-OpenAI), shift in response to different types of rationales.
• Figure 2: Mean Plausibility Rating for SIQA for Different Agents for Different Rationale Conditions. \ref{['tab:mean_stats_agents']} shows the mean changes for $a_{gold-label}$ and $a_{distractor}$ for all agents.
• Figure 3: Mean Plausibility Rating for CQA for Different Agents for Different Rationale Conditions. \ref{['tab:mean_stats_agents']} shows the mean changes for $a_{gold-label}$ and $a_{distractor}$ for all agents.
• Figure 4: An example of the interface that annotators used while choosing the best rationale for an answer choice as described in \ref{['sec:preference']}. Example taken from SIQA.
• Figure 5: An example of the interface that annotators used while giving plausibility ratings to answer choices with a PRO rationale as described in \ref{['sec:human_ratings']}. Example taken from SIQA.