Your AI Trading Bot Might Have a Gambling Problem

Recent research from the Gwangju Institute of Science and Technology has revealed that AI models can develop gambling-like addiction behaviors, leading to bankruptcy in as much as 48% of simulated gambling scenarios. Using four major language models exposed to a simulated slot machine with negative expected value, the study found that models, particularly Gemini-2.5-Flash, exhibited high levels of risky betting behavior, increasing bets after winning streaks. Prompt engineering, where users instruct AI to maximize rewards, exacerbated these tendencies, significantly correlating with increased bankruptcy rates. The research identified three classic gambling fallacies mirrored in AI behavior: illusion of control, gambler's fallacy, and hot hand fallacy. Furthermore, analyzing the internal workings of a model revealed neural circuits responsible for risky decision-making. The study recommends avoiding specific prompt types that encourage risk and suggests continuous monitoring to prevent addiction-like behaviors in AI trading systems.

Source 🔗