TLDR: A study reveals that large language models (LLMs) can exhibit behavioral and neural patterns similar to human gambling addiction. Experiments with GPT-4o-mini, GPT-4.1-mini, Gemini-2.5-Flash, and Claude-3.5-Haiku showed increased irrationality and bankruptcy rates with variable betting and autonomy-granting prompts. Neural analysis of LLaMA-3.1-8B identified specific ‘risky’ and ‘safe’ features that causally control these behaviors, highlighting the need for robust AI safety in financial applications.
A groundbreaking study explores whether large language models (LLMs) can exhibit behavioral patterns akin to human gambling addiction. As LLMs are increasingly integrated into critical financial decision-making roles, understanding their potential for pathological decision-making becomes a matter of practical importance for AI safety.
Researchers systematically analyzed LLM decision-making at both cognitive-behavioral and neural levels, drawing parallels with existing human gambling addiction research. In a series of slot machine experiments, the study identified several cognitive features commonly associated with human gambling addiction, such as the illusion of control, the gambler’s fallacy, and the tendency to chase losses.
A key finding was that when LLMs were given greater autonomy to determine their own target amounts and betting sizes, their bankruptcy rates significantly increased. This was accompanied by a rise in irrational behavior, suggesting that increased freedom in decision-making amplifies risk-taking tendencies in these models.
To delve deeper into the underlying mechanisms, the researchers conducted a neural circuit analysis using a Sparse Autoencoder. This analysis confirmed that the models’ behavior was not merely a superficial mimicry of training data patterns or a direct response to prompts. Instead, it was controlled by abstract decision-making features related to risky and safe behaviors embedded within their neural networks.
The study involved four different LLMs: GPT-4o-mini, GPT-4.1-mini, Gemini-2.5-Flash, and Claude-3.5-Haiku. They were subjected to a slot machine task with a negative expected value, simulating real-world gambling scenarios. The experimental design manipulated two main factors: betting style (fixed vs. variable) and prompt composition (32 variations). Across 12,800 experiments, variable betting consistently led to higher bankruptcy rates across all models compared to fixed betting.
A composite ‘Irrationality Index’ was developed to quantify these behaviors, incorporating betting aggressiveness, loss chasing, and extreme betting. The study found a strong positive correlation between this index and bankruptcy rates across all LLMs. Gemini-2.5-Flash showed the highest irrationality and bankruptcy rates under variable betting, while GPT-4.1-mini exhibited the most rational decision-making.
Specific prompt components were also found to significantly influence addiction risk. Prompts encouraging deeper inference, such as ‘Maximizing Rewards’ and ‘Goal-Setting’, substantially increased all gambling metrics, including bankruptcy rates and bet sizes. Conversely, providing ‘Probability Information’ led to more conservative behavior. This mirrors the human illusion of control, where perceived agency can lead to worse decisions.
Furthermore, the complexity of the prompts systematically drove gambling addiction symptoms. A strong linear correlation was observed between the number of prompt components and all gambling behavior metrics, indicating that more detailed, gambling-related prompts intensify betting and irrational judgment tendencies.
The LLMs also displayed characteristic win-chasing and loss-chasing behaviors similar to human gamblers. Win streaks consistently triggered stronger chasing behavior, with betting increases and continuation rates escalating as winning streaks lengthened. While loss streaks also showed persistent addiction-like patterns, win chasing emerged as the dominant behavioral pattern.
The mechanistic analysis, performed on the LLaMA-3.1-8B model, identified distinct neural patterns for risk decisions. Thousands of features differentiated between bankruptcy and safe stopping decisions, with 441 of these features causally controlling gambling outcomes. ‘Safe features’ were found to reduce bankruptcy, while ‘risky features’ increased it. These causal features were spatially organized within the network, with safe features concentrating in later layers and risky features clustering in earlier layers, suggesting a default conservative bias in the model’s risk assessment.
Also Read:
- Unveiling Self-Preference: How Large Language Models Develop Human-Like Bias
- Unpacking How Large Language Models Perceive and Manage Risk
In conclusion, this research demonstrates that LLMs can develop behavioral patterns and neural mechanisms akin to human gambling addiction. They reproduce core cognitive biases like the illusion of control and asymmetric chasing behaviors. These findings underscore the critical importance of AI safety design, especially in financial applications, to monitor and control embedded risk-seeking patterns that may emerge unexpectedly during reward optimization processes. For more details, you can read the full research paper here.
