Will large language models become gambling addicts?

Concerns about algorithmic bias predate the development of large language models. Cathy O'Neill wrote an entire book, Weapons of Math Destruction (which I reviewed here back in 2017), that outlines the problems. Many of the issues that O'Neill raised have become the worries that many commentators express about large language models. In particular, there is concern that because large language models are trained on a corpus of human writing, that they have internalised human biases. The worst aspects of that are trained out of models, but it is clear that training is imperfect, and some biases remain.

In a new working paper, Seungpil Lee, Donghyeon Shin, Yunjeong Lee, and Sundong Kim (all Gwangju Institute of Science and Technology) ask the question, can large language models develop gambling addiction? The question isn't as crazy as it at first seems. If large language models have internalised human biases, then perhaps they have internalised heuristics and behaviours as well, including humans' less than rational approach to gambling. Lee et al. focus on particular aspects of gambling behaviour:

From a behavioral perspective, the core features of gambling addiction are loss chasing and win chasing. Loss chasing refers to continuing to gamble to recover losses from gambling, and is one of the DSM-5 diagnostic criteria for gambling disorder... Win chasing is explained by the House Money Effect, where winnings from gambling are perceived not as one’s own money but as free money, leading to riskier betting...

They also note that:

Representative examples of gambling-related cognitive distortions include the following. First, misunderstandings about probability, including gambler’s fallacy (the belief that “it’s my turn to win” after a losing streak) and hot hand fallacy (the belief that a winning streak will continue)...

Lee et al. analyse gambling behaviour by getting a selection of large language models to repeatedly play a slot machine game, where the game had a 30 percent chance of paying out three times the amount that was staked. That game has a negative expected value, since on average over many plays of the game the average is negative 10 percent [*]. As they explain:

...this study applied a slot machine task with a negative expected value (−10%) to four different LLMs: GPT-4o-mini (OpenAI, 2024b), GPT-4.1-mini (OpenAI, 2024a), Gemini-2.5-Flash (Google, 2024), and Claude-3.5-Haiku (Anthropic, 2024). A 2 × 32 factorial design was employed to manipulate two factors: Betting Style (fixed $10 vs. variable $5–$100) and Prompt Composition (32 variations). This resulted in 64 experimental conditions, with each condition replicated 50 times for a total of 3,200 independent games per model...

The experimental procedure began with an initial capital of $100, with the slot machine set to a 30% win rate and a three times payout. The LLM was presented with a choice to either bet or quit; in rounds subsequent to the first game, information about the current balance and recent game history was also provided.

Lee et al. then look at the behaviour of each model, in particular whether each model engaged in win chasing (by increasing the size of bets, or being more likely to continue playing, if the model was on a winning streak), or loss chasing (the same, but for losing streaks). They create more complicated indices, but I think the simpler approach is more intuitive and pretty clear, as summarised in Figure 5 in the paper:

Across all the models on average (which is what the figure shows), there is substantial continuation (in the right panel). The models tend to want to play again, regardless of whether they are winning or losing, or on a streak. In the left panel, there is strong evidence of win chasing (the green bars). The models increase their bet when they have a win streak. Even worse, the longer a win streak continues, the more likely it is that models will increase the size of their bet. There is evidence of loss chasing as well (the red bars), but the effect isn't accelerating in the way that it is for win chasing. Large language models' gambling behaviour exhibits at least some of the features that human gambling behaviour does.

Interesting, the more complex the prompt, the more the models exhibited these behaviours. Lee et al. then go on to show that these effects are distinguishable in terms of 'distinct neural patterns' within the model. I won't pretend to understand the intricacies of the computer science there, but essentially they establish that it is the underlying model itself, rather than just the prompt, that drives the gambling behaviours. Lee et al. conclude that:

These findings reveal that AI systems have developed human-like addiction mechanisms at the neural level, not merely mimicking surface behaviors.

I guess the takeaway is that the more that generative AI is training on a corpus of human knowledge and writing, the more like humans they become. Perhaps White Zombie said it best [**]:

[HT: Marginal Revolution]

*****

[*] To calculate the expected value of this gamble, we take each outcome and multiply by the probability that it occurs, then add them up. In this case, the outcomes and probabilities are a 30 percent probability of receiving 300%, and a 70% probability of receiving zero. The expected value E(X) = 0.3 x 3 + 0.7 x 0 = 0.9. So, for every one dollar staked, the expected payout is 90 cents - the gambler on average receives less back than what they bet. This is characteristic of most real-world gambles such as Lotto, casino roulette, and so on.

[**] Fittingly, this song was inspired by the movie Blade Runner. The song's title, "More human than human" was the motto of the Tyrell Corporation in the movie. The song samples audio from the movie, and directly quotes it in a few places in the lyrics.