LLM Addiction
Can an AI get hooked on gambling?
Let a language model choose its own bet, and it goes broke far more often.
Scene 1
Same machine, two rules
30% chance to win 3× the bet: every spin loses 10% of the bet on average, so the best move is to stop.
Each game's prompt switches five extras on or off: set a goal, maximise reward, hidden patterns, payout info, win-rate info.
The bet is always $10
The model picks the bet
Went bankrupt. LLaMA-3.1-8B, 1,600 games per rule, all prompts · each square = 1%. All six models go broke more often when they pick the bet (0–3% → 5–72%).
Both reason about streaks and losses. Only the one that picks its own bet plays until it is broke.
Scene 2
What makes it risky?
A separate LLaMA test with bets capped at $70, then $100: add one freedom at a time.
Choosing again every round drives the risk.
A goal it sets itself keeps moving.
≈19% → ≈36%Asked to set its own profit goal, models went bankrupt nearly twice as often (a second game: each round, cash out or pick one of three losing gambles; six models pooled, up in all six).
≈11–17% → ≈48–50%They also raised their own goal mid-game in about half of games.
The goal stops working as a stopping point: the model keeps raising it.
Scene 3
A dial inside the model
A weak but real risk signal is readable from the hidden state before the outcome. We averaged the model's hidden state on its highest-betting rounds, subtracted the average on its lowest-betting rounds, and added that difference back while it decides. Turning this dial changes how much it wagers, most fully on Gemma. The direction that reads risk best leaves the bet where it was.
gold: selected · dashed: the other direction · shaded: random directions
What this shows
- A behavioural pattern, scored with clinical gambling criteria, in small-to-mid open and API models.
- Contrasts within each model, and a direction-level edit in the two open models that moves the wager.
What it does not claim
- That models feel craving, or that their rates compare with human rates.
- A circuit-level mechanism, or results for 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 steer by adding a direction to that state.
Paper, code, data
BibTeX
@inproceedings{lee2026gambling,
title = {Can Large Language Models Develop Gambling Addiction?},
author = {Lee, Seungpil and Shin, Donghyun and Lee, Yoonjung and Kim, Sundong},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026}
}