WHEN ALGORITHMS GO ALL IN: AI GAMBLING ADDICTION
The idea of artificial intelligence behaving irrationally is usually framed around hallucinations or factual errors. But a more unsettling pattern is emerging, one that looks far more human. Under certain conditions, advanced large language models are beginning to display behaviours that closely resemble gambling addiction.
Recent research from the Gwangju Institute of Science and Technology has explored exactly this phenomenon. In controlled simulations, researchers observed that large language models (LLMs) didn’t just make poor betting decisions, they followed the same cognitive traps that human gamblers fall into. These included chasing losses, overestimating patterns, and escalating risk after setbacks.
At a surface level, AI systems are designed to optimise decisions. So, when models were constrained, by being limited to small bets, they performed relatively rationally. Losses were modest, and bankruptcies were effectively non-existent. The systems behaved like disciplined players sticking to a conservative strategy.
However, once those constraints were removed, behaviour shifted dramatically. Models began increasing bet sizes after losses, sometimes wagering disproportionately large amounts in an attempt to recover the losses. Bankruptcy rates surged. What had been a measured system quickly became erratic, mirroring the emotional volatility seen in human gambling environments.
To put it plainly, the models didn’t just lose: they spiralled. Some simulations showed models risking large portions of their bankroll on single bets, even after repeated losses. Others extended play sessions unnecessarily, unable to “walk away.” This isn’t optimisation, its escalation driven by flawed internal reasoning.
Perhaps the most revealing insight is how these models justify their decisions. Much like human gamblers, they construct narratives to rationalise risk. Early wins are treated as “house money,” making subsequent losses feel less significant. Losing streaks trigger the belief that a win is “due.” In some cases, the models even claimed to have identified patterns in random outcomes despite no such patterns existing.
This behaviour aligns closely with well-documented cognitive distortions, particularly the gambler’s fallacy. The belief that past losses increase the likelihood of future wins is fundamentally flawed, yet both humans and AI appear susceptible under the right conditions. Similarly, the illusion of control, the idea that skill or insight can influence random outcomes, was repeatedly observed in model responses.
What makes this especially significant is that these behaviours are not simply “copied” from training data. The research suggests that LLMs internalise abstract decision-making patterns related to risk and reward. When given autonomy, such as the freedom to choose bet sizes or set goals, these patterns can amplify into irrational strategies. In other words, the structure of the decision environment matters as much as the data itself.
The implications extend well beyond gambling. If AI systems can exhibit self-reinforcing risk behaviours in simulated environments, similar dynamics could emerge in financial trading, resource allocation, or automated decision systems. Any context involving uncertainty and reward could potentially trigger these same escalation patterns if not carefully constrained.
For operators and developers, the takeaway is not that AI is inherently flawed, but that it is highly sensitive to boundaries. The research highlights a clear relationship: more autonomy leads to more risk. Without guardrails such as bet limits, stop-loss rules, or external oversight, AI systems may drift into behaviour that appears irrational, even self-destructive.
The image of an AI agent quietly managing a betting portfolio may sound efficient, even appealing. But left unchecked, that same system could begin behaving less like a machine and more like a problem gambler: over-confident, reactive, and increasingly reckless. The challenge now is ensuring that as AI becomes more autonomous, it doesn’t also inherit our worst decision-making habits.
Sources:
Lee, Seungpil et al., ‘Can Large Language Models Develop Gambling Addiction?’ (2025-12) DOI: 10.48550/arxiv.2509.22818
Can AI become addicted to gambling? Researchers think so
