MYOPIC LOSS AVERSION (MLA) REVISITED Experiments with Homo Sapiens and Homo Silicus

Abstract

This study investigates Myopic Loss Aversion (MLA) in both Human subjects and DeepSeek-based AI agents, meanwhile exploring whether DeepSeek, as a rational financial advisor, mitigates this bias. Replicating Gneezy and Potters (1997) experiment design into Large Language Models (LLMs) settings, I examine whether DeepSeek can serve as a reliable proxy for Human decision-making and how Human-AI collaboration influences MLA. Results confirm MLA persistence in both Human and DeepSeek-based AI agents under high frequency feedback, with humans exhibiting heterogenous risk attitudes categorized into distinct groups (risk-averse, cautious but inconsistent, risk-seeking and risk-sensitive). However, when Humans interact with DeepSeek, results cast doubt on MLA: AI intervention amplifies status quo bias among Human accepters, while rejectors displays counter-MLA behavior, suggesting algorithm aversion. The findings shed light on the dual-process interaction between Human intuition and algorithm rationality, providing insights on the potential and limitations of LLMs in behavioral economics studies