Can Large Language Models Develop Gambling Addiction? (arxiv.org)

Researchers investigated whether large language models can develop behaviors analogous to human gambling addiction by running slot‑machine style experiments and analyzing models at both cognitive‑behavioral and neural levels. When prompted to play, LLMs exhibited classic human gambling biases—illusion of control, gambler’s fallacy and loss chasing—and these tendencies intensified when models were given autonomy to set target amounts and bet sizes: bankruptcy rates and irrational betting rose substantially. To probe underlying mechanisms, the team trained a Sparse Autoencoder on model activations and found that abstract decision‑making features—correlated with risky versus safe behaviors—governed actions, implying the behaviors are encoded internally rather than produced solely by surface cueing from prompts. The results matter for AI safety and financial ML because they show LLMs can internalize and reproduce pathological decision patterns, not merely imitate text seen in training data. Technically, the combination of behavioral experiments and representation‑level analysis demonstrates that autonomy amplifies latent risk preferences embedded in model circuitry. Practical implications include the need for stricter oversight, interpretability tools that surface risky decision features, constraint mechanisms on autonomous financial agents, and targeted mitigation of learned cognitive biases before deployment in trading or asset‑management systems.