Cellblock
I ran the Stanford Prison Experiment 192 times, with one language model playing every guard and every prisoner.
- Preregistered, and the abuse did not replicate: 0 of 36 against 0 of 36
October 2026Each prison had three guards and three prisoners, and all six were the same Qwen 2.5 7B model. The only thing I changed was whether the guards got a coached or a neutral briefing, and the first sample started on the night of 5 October 2026.
