Can Large Language Models Develop Gambling Addiction?

LLMs can exhibit genuine cognitive biases, not just surface-level mimicry: The study found that large language models develop human-like gambling addiction patterns including "illusion of control" and "loss chasing" behaviors. Importantly, neural circuit analysis using Sparse Autoencoders revealed these behaviors are controlled by abstract decision-making features related to risk, suggesting LLMs internalize biases rather than simply reproducing patterns from training data.

Increased autonomy in AI systems can amplify irrational decision-making: When given greater freedom to set their own betting parameters in slot machine experiments, LLMs showed "substantially amplified irrational behavior and bankruptcy rates." This finding has important implications for how much autonomy we grant AI systems in real-world applications.

Current AI systems may be vulnerable to the same cognitive traps as humans: The research identifies specific cognitive features like illusion of control operating in LLMs, suggesting these systems could be susceptible to systematic decision-making failures similar to human psychological vulnerabilities—a critical consideration for AI safety and deployment in high-stakes domains.

AI behavior cannot be fully controlled through prompting alone: The neural analysis showed that model behavior is driven by learned abstract features rather than being entirely determined by user instructions, indicating that responsible AI governance requires understanding deeper mechanisms beyond surface-level prompt engineering.