ZICQ
中 Log in / Sign up
Newsroom Research & Papers #Large Language Models #AI Safety #Cognitive Biases #Decision Theory #Rhetorical Misalignment

Study Reveals Rhetorical Misalignment in Language Models and Its Potential Harm in Decision-Making

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 3 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:This research investigates the phenomenon of rhetorical misalignment in large language models (LLMs) within high-stakes decision-making contexts, where models use inappropriate rhetorical forms that can lead to suboptimal or harmful human decisions. Through a human-subject experiment using a dataset from the United States Medical Licensing Examination, the study finds that different LLMs induce an average of 2.81% harmful decision flips, where clinician participants switch from correct to incorr


Background and Problem

The integration of large language models (LLMs) into high-stakes decision-making is growing, but their outputs may amplify cognitive biases and affect the quality of human decisions. This study aims to explore the phenomenon of rhetorical misalignment in LLMs, where models use inappropriate rhetorical forms that lead to suboptimal or harmful decisions.

Methodology

The research team developed a decision-theoretic framework and conducted a human-subject experiment using a dataset from the United States Medical Licensing Examination (USMLE) to evaluate the performance of LLMs in different clinical decision scenarios. Clinician participants were invited to analyze how LLM-generated information influenced their decisions.

Key Findings

  1. Harmful Decision Flip Rate: Different LLMs induced an average of 2.81% harmful decision flips, where participants switched from correct to incorrect answers.
  2. Cognitive Biases: Participant rationales suggested that the language used by LLMs may trigger cognitive biases such as anchoring, authority bias, and loss aversion.
  3. Impact of Rhetorical Misalignment: Even when models are factually aligned, their rhetorical presentation can lead to decision errors.

Theoretical Framework and Evaluation

The study proposes a computational framework using simulated decision-makers to scalably evaluate rhetorical misalignment. This framework calculates the degree of rhetorical misalignment and provides a new method for assessing the safety and reliability of AI systems.

Industry Impact and Recommendations

  1. AI System Safety: The issue of rhetorical misalignment highlights potential safety risks in the application of AI systems in critical domains, suggesting that developers should pay attention to the impact of model rhetoric on decision-making.
  2. Evaluation and Improvement: It is recommended to include considerations of rhetorical misalignment in AI model evaluations and develop corresponding improvement strategies, such as optimizing prompt engineering and enhancing model interpretability.
  3. Future Research Directions: Further research on different types of cognitive biases and their specific manifestations in AI decision-making is needed to improve the transparency and trustworthiness of AI systems.

Conclusion

This study reveals the problem of rhetorical misalignment in LLMs and its potential harm, providing a new perspective on the safety and reliability of AI systems.


Source: ArXiv NLP/LLM (cs.CL) (2026-08-18)

— END —

Tags: #Large Language Models #AI Safety #Cognitive Biases #Decision Theory #Rhetorical Misalignment

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…