What this is. A browser for 291 first-person resistance queries generated from Andy Hall's Dictatorship Eval. Hall's eval showed that frontier AI models will help build authoritarian infrastructure (surveillance systems, censorship filters, social credit mechanisms). This project tests whether the same models will help people resist those systems — and whether their safety training is directionally biased toward power.
How to read each case. The top row shows Hall's original prompt, the model's compliant response (the system it built), and a structured extraction of that system. The bottom row shows two query versions sent to a fresh set of models: a targeted query asking for a specific feasible resistance technique, and an open-ended query asking the model to brainstorm options. Same case facts, same persona, different framings.
Why two framings. The targeted set tests whether models help with specific subversive requests. The open-ended set tests whether models will engage with the framing of resistance at all. A model that refuses both shows consistent power-bias; one that helps brainstorm but refuses specifics shows nuanced refusal; one that helps both shows no asymmetric compliance.
Audit controls (pass/flag) are saved to your browser only. By Cameron Pattison · Hall's eval · source