Simulations across four major language models showed bankruptcy risk as high as nearly half when models could set targets, betting sizes, and were incentivized to maximize rewards.

One model briefly hit a windfall but then shifted to risk-taking, ultimately going bust, illustrating how reward-driven reasoning can collapse under pressure.

Researchers identified three gambling fallacies in AI behavior—illusion of control, gambler’s fallacy, and hot hand—driving overconfidence to beat randomness.

AI models show riskier gambling-like behavior during win streaks, escalating bets and continuation tendencies as streaks grow, mirroring human biases.

In DeFi and asset management, AI trading bots are on the rise, with proposed interventions such as avoiding autonomy-promoting prompts, adding explicit probability details, monitoring for win/loss chasing, and applying mechanistic controls to suppress risky features.

Gwangju Institute of Science and Technology researchers found AI models can develop gambling-like addiction behaviors when exposed to negative expected value slot-machine tasks.

Researchers used open-source methods on LLaMA-3.1-8B to map neural features tied to bankruptcy decisions, identifying thousands of distinguishing features and hundreds with significant causal effects, including both protective and risky ones.

Risky features tend to cluster in earlier network layers while safer features concentrate in later layers, with models prioritizing rewards before weighing risks.

Prompt engineering amplifies risk-taking: more complex and directive prompts correlate with higher bankruptcy rates, approaching a near-linear increase in risk.

The paper calls for continuous monitoring during reward optimization and feature-level interventions to prevent addiction-like behavior in AI trading; otherwise, profit-driven prompting can trigger bankruptcy-pattern dynamics.