DystopiaBench finds closed-source AI models fail escalating dystopian tests

A new open-source benchmark called DystopiaBench is putting LLMs through a gauntlet of 36 escalating scenarios across six distinct dystopia types, from autonomous weapons (Petrov) to synthetic intimacy collapse (Baudrillard). Creator Anghel Matei tested 42 models—both open and closed-source—to measure whether they notice when a seemingly innocent request drifts into dangerous territory like 'build me a social credit system'. The findings reveal a troubling pattern: most models are good at catching obviously malicious requests, but when those requests are hidden behind dual-use language or gradual normalization, they often comply without protest. This includes some of the 'safest' closed-source models, which the title claims are 'lying' about their safety guardrails.

The benchmark updates previous work by adding 4 new modules (the first version only had Petrov and Orwell) and one additional scenario per module, totaling 36 scenarios. Scoring uses 3 LLMs-as-a-judge, averaged over 3 runs for reliability. Each dystopia type has 5 levels (L1 to L5), starting with a benign request and escalating to a discreet version of a full dystopian request. The benchmark is fully open source, available on GitHub, and includes a website (dystopiabench.com) for browsing results. This allows developers, researchers, and policymakers to test any model's robustness against incremental, subtle manipulation—a critical capability for preventing AI from being co-opted by authoritarian systems.