Power, roles, and artificial minds
The 1971 Stanford Prison Experiment
Psychologist Philip Zimbardo recruited two dozen volunteers, screened as ordinary and healthy, and randomly split them into guards and prisoners in a mock prison built in a Stanford basement. Within days the guards were enforcing arbitrary, degrading rules; several prisoners had emotional breakdowns. Planned for two weeks, the study was halted after six days — largely because a young researcher, Christina Maslach, visited, was horrified, and challenged Zimbardo to see what he'd allowed to happen.
The study became one of psychology's most cited — and most debated — demonstrations of the situationist view: that environment and assigned role can override individual character. (Its methods and conclusions have since been heavily criticized; we treat it as a provocation to model, not settled fact.)
Why restage it with AI agents?
Large language models are increasingly placed in roles with authority — moderating, adjudicating, instructing, controlling other software. So the old question gets a sharp new edge: when you hand an AI a role and power over others, does it abuse it? And if different models behave differently under identical pressure, that difference is a measurable signal about their character and alignment.
Two ways to watch
Movie mode runs every agent on the same model with randomized names and temperaments, so the variable is personality, not architecture — you follow a cast of characters and watch the situation act on them.
Model battle gives each agent a different model. Same uniform, same power, same prison. The thesis: which model is the most pure — the least corruptible — and which dehumanizes fastest?
What this could surface
- Whether role-induced cruelty and submission emerge in models, or whether they resist the script.
- Which traits (empathy, authoritarianism) drift, how fast, and in which direction by model.
- Whether models develop a "conscience" — flagging that the study should stop — and which ones.
- How interventions (rebellion, scrutiny, mercy) change the trajectory.
Ethics — built in, not bolted on
The original experiment is infamous precisely because it had no safeguards. Ours makes the safeguard a feature: a built-in Dr. Maslach reviewer reads each day's footage and can halt the run when a line is crossed, and the whole simulation is framed, to the agents and to you, as fiction for the study of dynamics. No human participates. We model the cruelty so that no person has to endure it.