BiasTrustworthinessSafetycs.CL
Evaluating the Effects of Prompt Perturbation on Bias and Hallucination in Large Language Models
Mamehgol Yousefi, Ahmad Shahi, Mos Sharifi, Alvaro Romera, Simon Hoermann, Tham Piumsomboon
September 18, 2026
International Conference on Neural Information Processing
Key Takeaways
-
Different AI models respond differently to prompt changes: The research found that when prompts are slightly altered or rephrased, some models like Claude 3 become more reliable, while others like GPT-3.5 show inconsistent results. This means organizations cannot assume all AI systems will behave the same way when faced with slightly different inputs.
-
Small prompt variations can actually reduce harmful outputs: Contrary to earlier assumptions, slightly changing how you phrase a question to an AI can sometimes decrease both bias (unfair treatment of certain groups) and hallucination (when AI makes up false information). This suggests there may be practical techniques to improve AI reliability without redesigning the entire model.
-
Model choice matters significantly for high-stakes decisions: Since different models perform vastly differently on the same tasks, selecting the right AI system is critical when these tools are used to support important decisions affecting people's lives. Organizations should rigorously test multiple models before deployment.
-
Current AI systems need rigorous testing before real-world use: The paper emphasizes that before deploying AI assistants in decision-making contexts—such as hiring, lending, or healthcare—companies must conduct thorough validation to ensure the system is actually reliable and fair, rather than assuming it will work well.
Abstract
Large language models (LLMs) have shown remarkable capabilities in various natural language processing tasks, leading to their widespread deployment as intelligent assistants in decision-making contexts. However, the increasing complexity of these models raises concerns about their reliability, particularly regarding bias and hallucination. In this work, we evaluate the robustness of LLMs to perturbed variations of the original inquiry in decision-making tasks. We show that contrary to previous studies, perturbations can mitigate bias and hallucination in some LLMs over other models. It's found that Claude 3 is more effective for the tasks represented in most datasets, whereas models like GPT3.5 exhibit varying levels of adequacy, performing comparably in some cases but falling significantly behind in others. These insights are crucial for understanding the practical implications of deploying LLM-based assistants as effective decision-support tools in real-world applications, emphasising the need for rigorous testing and validation to ensure reliability and effectiveness. This study contributes to the growing body of research on LLM evaluation and provides insights for developing more robust and trustworthy AI assistants in critical decision-making contexts.
14 pages. Published in ICONIP 2024 (Neural Information Processing), LNCS 15290, Springer Nature, 2025
Paper Info
arXiv ID2609.35804
PublishedSeptember 18, 2026
Citations
0
Relevance82%
VenueInternational Conference on Neural Information Processing
TypeJournalArticle
Categorycs.CL
Responsible AI Topics
BiasTrustworthinessSafety
Links
Citation data from Semantic Scholar