LLMs Exhibit Decision Biases Through Faulty Mimicry, Study Finds

A recent study published on arXiv (2608.12339) explores how decision biases arise in large language models (LLMs), showing that such biases can develop even when the training data reflects unbiased human preferences or accurately identifies biases. The research uncovers two primary mechanisms: incorrect imitation of human preferences based on behavior unrelated to those preferences, and imitation of overtly biased human actions. In four experiments centered on economic biases, both ChatGPT-4o and Qwen exhibited social proof biases when exposed to reports of human actions that did not genuinely indicate preferences. Moreover, the models showed loss aversion when it was labeled as a bias, despite being prompted with comprehensive scientific reports. This study emphasizes that LLMs can deduce preferences from illogical behaviors, resulting in biased decisions. It highlights the importance of carefully selecting training data and prompting methods to reduce unintended biases in AI systems.