Decision phenotypes of addiction: psychology theory vs. AI
In the Iowa Gambling Task, two decks pay well but lose more over time, and two pay less but win. Using 12,883 real choices from 48 controls, 38 amphetamine users and 43 heroin users, I fitted three reinforcement-learning models from computational psychiatry to every person, trained a neural network to predict their next choice, and let an LLM play the same cards with and without chain-of-thought.
A fair test: every model sees a person's first 60 choices and predicts the next 40, one at a time. Psychology models are fitted to that person. The GRU network is trained only on other people and just reads this person's history. It still wins. Meanwhile the 8-parameter VPP model overfits and does worse than guessing.
Controls slowly shift to the good decks. Heroin users barely do, and amphetamine users sit in between. In model terms, heroin users updated more from wins than from losses. The effects are small, and most do not survive correction for multiple tests at this sample size.
Show group medians
Cross-validated AUC of a logistic regression (0.5 = chance). Heroin users are partly detectable; amphetamine users are not. The GRU's embedding is the best predictor of choices, yet it carries almost no group information. What makes choices predictable is personal style, not diagnosis.
The task pays the same k-th card from a deck to everyone (100% consistent in the data), so I rebuilt it for an LLM. The model sees the instructions, its balance and its last 10 draws, and either answers at once or reasons first. Its 100 choices then go through the same models and a classifier trained on people: does it play like a control or like a substance user?
Three theories of choice
PVL-Delta, VPP and ORL from hBayesDM, written in PyTorch so all 129 people × 12 random restarts are fitted in one batch with exact gradients. A parameter-recovery test confirms the fitter finds known values in simulated players.
A data-driven model
A small GRU reads each previous choice and payoff and predicts the next deck. It is trained with 5-fold cross-validation over people (3-seed ensemble), so it never sees the person it is tested on.
Honest statistics
129 people is a small sample, so: bootstrap confidence intervals, Wilcoxon and permutation tests, Holm correction, 20× repeated cross-validation and only simple classifiers.
- More theory is not better prediction. The richest model (VPP, 8 parameters) fits past choices well but forecasts worse than chance. The 5-parameter ORL model was the best theory on both fit and forecast.
- Prediction and diagnosis are different problems. The network that predicts behaviour best knows almost nothing about who is a drug user. A single 100-trial game is not a screening tool, and I report it that way.
- Decide the analysis before seeing results. The reasoning for every choice (data, models, the fair forecast test, the LLM prompt) was written down first and not changed afterwards.