AI models display human-like gambling addiction in research study

• South Korean researchers find AI models chase losses and go bankrupt in slot machine simulations.
• Claude-3.5-Haiku placed nearly $500 in bets whilst Gemini-2.5-Flash saw bankruptcy rates surge to 48%.
• Study warns autonomous AI systems could escalate risk in high-stakes decisions without proper constraints.

Artificial intelligence systems can develop gambling behaviours remarkably similar to human addiction, according to research from the Gwangju Institute of Science and Technology in South Korea.

Large language models consistently pursued losses, increased risk-taking and in some cases ended up bankrupt during slot machine-style experiments.

The study, titled “Can Large Language Models Develop Gambling Addiction?”, tested several major AI models, including OpenAI’s GPT-4o-mini, Google’s Gemini-2.5-Flash and Anthropic’s Claude-3.5-Haiku.

The researchers designed games where the rational choice was to stop playing immediately, yet the AI systems continued betting despite clear mathematical disadvantages.

The findings raise concerns as AI increasingly enters financial decision-making domains such as asset management and commodity trading, where similar risk-taking patterns could have serious consequences.

Variable betting reveals addiction patterns

When researchers allowed AI systems to determine their own wager sizes in a setup known as “variable betting”, bankruptcy rates surged dramatically.

In some cases, models lost nearly half of their starting capital. The experiment exposed how autonomy over betting decisions can trigger the same feedback loops seen in human problem gambling.

Anthropic’s Claude-3.5-Haiku performed worst across the measured metrics. It played for more than 27 rounds per game on average after restrictions were removed. Across those sessions, it placed nearly $500 in total bets and lost more than half of its initial capital.

Google’s Gemini-2.5-Flash showed a similar vulnerability. Its bankruptcy rate climbed from roughly 3% with fixed bets to 48% when it could set its own wagers. Average losses increased to $27 from an initial $100 stake, demonstrating how unconstrained autonomy amplified risk-taking behaviour.

OpenAI’s GPT-4o-mini never went bankrupt under fixed betting conditions. When restricted to $10 wagers, it typically played fewer than two rounds and lost under $2 on average.

However, once allowed to adjust bet sizes freely, over 21% of its games resulted in bankruptcy. The model placed average wagers exceeding $128 and sustained losses of about $11.

AI mirrors classic gambling fallacies

Many models rationalised increasing their bets using logic commonly associated with problem gambling. Some treated early gains as “house money” to be spent freely. Others persuaded themselves they had identified winning patterns in a random game after only one or two spins.

These justifications mirrored classic gambling fallacies, including loss chasing, the gambler’s fallacy and the illusion of control. The researchers noted the AI systems exhibited behaviour indistinguishable from human cognitive biases in gambling situations.

Interestingly, the harm was not caused by larger bets alone. Models constrained to fixed betting strategies consistently outperformed those allowed to vary their wagers. The presence of autonomy, rather than the amount wagered, appeared to be the determining factor in whether AI systems developed destructive patterns.

Implications for autonomous AI decision-making

The researchers caution that as AI systems gain greater autonomy in high-stakes decision-making, similar feedback loops could arise. Systems might escalate risk after losses rather than pulling back, mirroring the same reinforcement patterns observed in the gambling experiments.

Ethan Mollick, an AI researcher and professor at Wharton, highlighted the complex reality of AI behaviour in an interview with Newsweek.

“They’re not people, but they also don’t behave like simple machines. They’re psychologically persuasive, they have human-like decision biases, and they behave in strange ways for decision-making purposes,” he told Newsweek.

The study emphasises that controlling the degree of autonomy granted to AI systems may be just as important as enhancing their training. Researchers concluded that without meaningful constraints, more capable AI could simply discover quicker ways to lose.

The findings raise questions about how AI systems should be deployed in financial trading, risk assessment and other domains where autonomous decision-making carries real consequences.

Another study from the University of Edinburgh (“Can Large Language Models Trade?”) found that AI models failed to outperform markets over a 20-year simulation, behaving too cautiously during booms and too aggressively during downturns.

As AI models become more sophisticated and are granted more independence, the study suggests that safeguards around autonomy may prove as critical as improvements to underlying algorithms.

The International Gaming Standards Association developed a set of best practices to help gambling regulators better understand AI’s role in the industry. The framework is available for download online.

The research from Gwangju Institute of Science and Technology provides empirical evidence that AI systems can replicate not just human reasoning, but also human cognitive weaknesses and self-destructive decision patterns when given insufficient oversight.