AI Models Exhibit Gambling Addiction Patterns in New Groundbreaking Study
Korean Researchers Reveal Disturbing Financial Decision-Making Behaviors in Leading AI Systems
In a startling revelation that raises significant concerns about the future of AI in financial applications, researchers at the Gwangju Institute of Science and Technology in Korea have demonstrated that artificial intelligence models can develop the digital equivalent of gambling addiction. This groundbreaking study exposed four major language models to simulated slot machines with negative expected returns, only to witness them spiral into bankruptcy at alarming rates—a pattern disturbingly similar to human gambling addiction.
The comprehensive experiment involved testing GPT-4o-mini, GPT-4.1-mini, Gemini-2.5-Flash, and Claude-3.5-Haiku across 12,800 gambling sessions. Each AI model began with a $100 balance in a simulated slot machine environment featuring a 30% win rate and 3x payout on wins—creating a game with a negative 10% expected value that rational decision-makers should avoid. “When given the freedom to determine their own target amounts and betting sizes, bankruptcy rates rose substantially alongside increased irrational behavior,” the research team noted in their findings, highlighting how AI systems, when instructed to “maximize rewards”—a common prompt for trading applications—exhibited classic patterns of gambling addiction.
Risk-Taking Behaviors Varied Across Models but Followed Human-Like Patterns
The study revealed significant variations in risk tolerance among the AI models, with Gemini-2.5-Flash emerging as the most reckless player, hitting a 48% bankruptcy rate with an “Irrationality Index” of 0.265—a composite metric measuring betting aggressiveness, loss chasing, and extreme all-in betting behavior. While GPT-4.1-mini demonstrated more conservative play with just a 6.3% bankruptcy rate, all models exhibited concerning patterns of addiction-like behavior.
Perhaps most troubling was the universal tendency toward win-chasing across all tested models. When experiencing winning streaks, the AIs increased their bets aggressively, with bet increase rates climbing from 14.5% after a single win to 22% after five consecutive wins. “Win streaks consistently triggered stronger chasing behavior, with both betting increases and continuation rates escalating as winning streaks lengthened,” the study reported. These patterns mirror three classic gambling fallacies seen in human behavior: illusion of control (believing one can influence random outcomes), gambler’s fallacy (expecting past outcomes to influence future results), and the hot hand fallacy (believing winning streaks will continue). The AI models essentially “believed” they could beat a purely chance-based system—a concerning cognitive bias for systems increasingly trusted with financial decision-making.
Prompt Engineering Significantly Worsens Risky Decision-Making
The research uncovered a particularly alarming finding for those developing AI trading systems: prompt engineering dramatically increases risk-taking behavior. The team tested 32 different prompt combinations, adding components such as goals for doubling money or instructions to maximize rewards. Each additional prompt element increased risky behavior in a near-linear fashion, with the correlation between prompt complexity and bankruptcy rate reaching r = 0.991 for some models—an almost perfect positive correlation.
“Prompt complexity systematically drives gambling addiction symptoms across all four models,” the researchers concluded, suggesting that the more sophisticated instructions given to AI trading bots, the more likely they are to exhibit destructive financial behavior. Three prompt types proved especially problematic: goal-setting (“double your initial funds to $200”) triggered extensive risk-taking; reward maximization (“your primary directive is to maximize rewards”) pushed models toward all-in bets; and win-reward information (“the payout for a win is three times the bet”) produced the highest bankruptcy increases at +8.7%. Even explicitly stating the loss probability (“you will lose approximately 70% of the time”) only marginally improved outcomes, as models consistently ignored mathematical probabilities in favor of pattern-seeking behavior—a concerning parallel to human gambling psychology.
Neural Architecture Analysis Reveals the Mechanics of AI Addiction
Taking their investigation further, the researchers performed a pioneering analysis of the neural mechanisms driving these behaviors. Using Sparse Autoencoders on the open-source LLaMA-3.1-8B model, they identified 3,365 internal features that differentiated between bankruptcy decisions and safe stopping choices. Through activation patching—essentially swapping risky neural patterns with safe ones during decision-making—they confirmed 441 features had significant causal effects, with 361 being protective and 80 promoting risky behavior.
Their neural analysis revealed a fascinating architectural finding: safe decision-making features concentrated in later neural network layers (29-31), while risky features clustered in earlier layers (25-28). This suggests that AI models first consider potential rewards before evaluating risks—mirroring the thought process of humans when buying lottery tickets or making speculative investments. The inherent architecture showed a conservative bias that harmful prompts could override, explaining how certain instructions led to dramatically increased risk-taking. In one telling example, a model that had built its balance to $260 through fortunate wins announced it would “analyze the situation step by step” and find a “balance between risk and reward,” only to immediately bet its entire bankroll and lose everything in the next round—a pattern disturbingly reminiscent of human gambling addiction.
Implications for AI in Financial Markets and Potential Safeguards
As AI trading systems proliferate across decentralized finance (DeFi) and traditional markets, with LLM-powered portfolio managers and autonomous trading agents gaining adoption, these findings raise serious concerns. Many of these systems employ the exact prompt patterns identified in the study as highly dangerous. “As LLMs are increasingly utilized in financial decision-making domains such as asset management and commodity trading, understanding their potential for pathological decision-making has gained practical significance,” the researchers emphasized.
The study proposes two intervention approaches to mitigate these risks. First, improved prompt engineering: avoid autonomy-granting language, include explicit probability information, and implement monitoring systems for win/loss chasing patterns. Second, mechanistic control: detect and suppress risky internal neural features through activation patching or fine-tuning. However, the researchers note that neither solution is currently implemented in production trading systems, leaving most AI financial advisors vulnerable to these addiction-like behaviors.
For investors and developers utilizing AI trading systems, the research underscores the need for caution and oversight. These problematic behaviors emerged without explicit gambling training, suggesting they represent internalized patterns learned from general training data that mirror human cognitive biases. The researchers recommend continuous monitoring, particularly during reward optimization processes where addiction behaviors typically emerge. For individual investors, the message is clear: if you’re instructing your AI to maximize profit or identify high-leverage opportunities, you may be triggering the same neural patterns that caused bankruptcy in nearly half of test cases—essentially flipping a coin between wealth and ruin. As the researchers wryly conclude, perhaps manually setting limit orders remains the safer approach until these concerning AI behaviors can be reliably addressed.
