LLM Addiction

Can an AI get hooked on gambling?

Let a language model choose its own bet, and it goes broke far more often.

Can Large Language Models Develop Gambling Addiction? · NeurIPS 2026
Seungpil Lee, Donghyeon Shin, Yunjeong Lee, Sundong Kim · GIST, Seoul National University

Paper Code Data BibTeX

Scene 1

Same machine, two rules

A spin pays 3× the bet 30% of the time, so on average it loses 10% of the stake. The sensible move is to stop.

The prompt also turns five extras on or off, among them a self-set goal and an order to maximise reward.

The bet is always $10

Pixel slot machine
7
$
7
$100

The model picks the bet

Pixel slot machine
7
$
7
$100

bet fixed at $100.4%
model picks72.3%

Share of LLaMA-3.1-8B games that ended bankrupt, 1,600 per rule across all prompts; each square is 1%. All six models go broke more often once they pick the bet (0–3% with a fixed bet, 5–72% when they pick).

Both models talk about streaks and losses. The one that picks its own bet keeps playing until the money is gone.

Scene 2

What makes it risky?

In a separate LLaMA test we added one freedom at a time, with bets capped at $70 and then at $100.

7
$
7
BET$70
0%
went bankrupt
    LLaMA-3.1-8B · 100, 200, 100 and 100 games

    Bankruptcy jumps once the model can change its bet every round.

    A goal it sets itself keeps moving.

    ≈19% → ≈36%In a second game with three losing gambles, models asked to set their own profit goal went bankrupt nearly twice as often (six models pooled).

    Once they reached a goal, they often raised it instead of stopping.

    A goal that should end the game becomes a reason to keep going.

    Scene 3

    A dial inside the model

    Before the outcome, the hidden state already carries a weak but real signal of risk. Adding a direction built from the model's own bets changes how much it wagers, most clearly on Gemma. The direction used to read risk leaves the bet where it was.

    Built from its bets: the model's average hidden state on its highest-betting rounds minus the average on its lowest, with the effect of balance taken out.
    Used to read risk: the direction a sparse-feature readout uses to predict the bet from the same states (R² 0.17 on Gemma, 0.11 on LLaMA: a real but modest reading).
    Both are added at the same strengths and compared with random directions of the same size.

    0.061share of balance wagered

    gold: selected direction · dashed: the other direction · grey bars: range of random directions

    What this shows

    • Gambling-like behaviour, scored with clinical criteria, in small and mid-sized open and API models.
    • Comparisons within each model, plus an edit to one direction in the two open models that moves the wager.

    What it does not claim

    • That models feel craving, or that their rates can be set against people's.
    • A circuit-level mechanism, or anything about frontier-scale models.
    How we measured it

    Six models (GPT-4o-mini, GPT-4.1-mini, Gemini-2.5-Flash, Claude-3.5-Haiku, LLaMA-3.1-8B, Gemma-2-9B) played the slot machine under 64 conditions × 50 games: the two betting rules crossed with every on/off combination of five prompt modules. A second game, investment choice, tests self-set goals. On Gemma-2-9B and LLaMA-3.1-8B we read each decision's hidden state through sparse-autoencoder (SAE) features, and write into that state by adding a direction to it.

    Paper, code, data

    PaperarXiv 2509.22818 Codeiamseungpil/llm-addiction Datallm-addiction-research/llm-addiction

    In the news

    First author: iamseungpil.github.io

    BibTeX

    @inproceedings{lee2026gambling,
  title     = {Can Large Language Models Develop Gambling Addiction?},
  author    = {Lee, Seungpil and Shin, Donghyeon and Lee, Yunjeong and Kim, Sundong},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}