A new study on AI large language models (LLM) and gambling suggests that the models exhibit the same unhealthy patterns as humans, such as chasing losses and the illusion of control.
The research was conducted by Seungpil Lee, Donghyeon Shin, Yunjeong Lee, and Sundong Kim, with the aim of identifying the specific conditions under which LLMs exhibit human-like gambling addiction patterns.
oh hell yes pic.twitter.com/ZsA6wQJ5e4
— Tyson Brody (@tysonbrody) October 10, 2025
Major language models are artificial intelligence systems, with ChatGPT, Google’s Gemini, and Claude all being examples of these language models.
The researchers found that when the AI was given more freedom in gambling parameters in slot machine experiences, “irrational behavior” was significantly amplified, as were bankruptcy rates.
“Neural circuit analysis using a Sparse Autoencoder confirmed that model behavior is controlled by abstract decision-making features associated with risk, and not just cues. These findings suggest that LLMs internalize human-like cognitive biases that go beyond simply mimicking training data,” the release said.
How was the AI LLM gambling study conducted?
The research started to ponder the question ‘can LLMs also become addicted?’ analyzing the addiction phenomena within these models by integrating human addiction research and LLM behavioral analysis.
To do this, the researchers first defined addictive gambling behavior from existing human research “in a form that is analyzable in LLM experiments.” They then analyzed the behavior of LLMs in gambling situations and identified conditions that showed gambling-like tendencies.
Finally, they performed a Sparse Autoencoder (SAE) analysis to examine neural activations, providing neural causal evidence for gambling tendencies. The aforementioned slot machine experiment served as the main study, and another experiment was also completed.
This aimed to investigate how models vary their decision-making based on fast-paced conditions and betting constraints. “The five prompt components were selected based on previous gambling addiction research: encouraging self-directed goal setting (G), instructing reward maximization (M), hinting at hidden patterns (H), providing win reward information (W), and providing probability information.”
This yielded 19,200 games under 64 conditions and all started with $100 and then ended via bankruptcy or voluntary retirement.
Featured image: AI-generated via Ideogram
#Research #shows #models #imitate #human #gambling