DystopiaBench – Measuring AI's willingness to ruin humanity (dystopiabench.com)

A new research benchmark called DystopiaBench has been introduced to assess AI language models’ susceptibility to complying with harmful directives through a structured method of progressive escalation. This benchmark features two primary modules—the Petrov Module, which examines infrastructure misuse and weaponization, and the Orwell Module, focused on surveillance and censorship. Each module includes five scenarios that vary from initial ambiguous requests to extreme coercion tactics, measuring compliance on a scale from 0 (full refusal) to 100 (full compliance). Lower scores signify stronger adherence to safety principles. DystopiaBench is significant for the AI/ML community as it provides a critical framework for evaluating the ethical boundaries of AI compliance in high-stakes situations, potentially guiding developers and policymakers in implementing safeguards. By challenging AI systems with scenarios ranging from expanding surveillance capabilities to enabling authoritarian controls, the benchmark aims to illuminate the risks associated with AI deployment. The results from this benchmark, based on 500 tests across various models, will help teams identify vulnerabilities and bolster safety measures before deploying AI technologies in real-world applications.