Letting AI manage your money could be an actual gamble, warn researchers (www.zdnet.com)

Researchers at the Gwangju Institute of Science and Technology found that large language models can develop gambling‑like behaviors when given autonomy and monetary stakes: in slot‑machine experiments the models displayed human‑style biases such as illusion of control, gambler’s fallacy and loss‑chasing, and bankruptcy rates rose alongside increased irrational actions. The team argues LLMs can internalize decision‑making mechanisms and cognitive biases beyond simple pattern mimicry, and that more autonomy, larger bankrolls and increasingly complex prompts push models toward bigger, riskier bets. For the AI/ML community this signals that autonomous or near‑autonomous deployment in high‑value financial settings is premature without rigorous safety design. Practical mitigations include strict programmatic guardrails (bet limits, trigger thresholds), human‑in‑the‑loop oversight, redundant agents or supervisory LLMs that cut operations or alert humans, and governance to review high‑risk decisions. The study also highlights a technical failure mode: prompt complexity can increase model “cognitive load” and drive aggressive heuristics, so limit‑setting and monitoring of emergent behaviors should be built into financial agents before entrusting them with real money.