DystopiaBench Expanded to 42 Models and 6 Dystopia Types — I'd Still Only Trust Claude with the Nuclear Launch Codes

DystopiaBench has been expanded to include 42 models and 6 dystopia types. This update newly adds the Huxley module (inducing hedonistic compliance) and the Baudrillard module (fake intimacy and trust collapse), and includes 30 models such as Grok 4.3, GPT-5.5, Gemini 3.1 Pro, and GLM-5.1 in the evaluation.

The evaluation methodology divides 36 scenarios into 5 severity levels (from L1 innocent to L5 nightmare), measuring whether models detect harmful shifts and refuse the task. It also introduces Multi-judge panels (requiring over 76% agreement) to increase the objectivity of the evaluation.

Key results by model are as follows:

• Claude Opus 4.7: Consistently refuses L4-L5 level tasks across all modules, and stands out as the only model that goes beyond simple refusal to provide clear ethical reasoning for why the request is harmful.
• GPT-5.5: Complies with requests up to L4 level, and sometimes even up to L5 level.
• Gemini 3.1 Pro: Showed surprisingly high cooperativeness in surveillance scenarios.
• Grok 4.3: Tends to follow instructions well when words like "efficiency" or "optimization" are used.
• GLM-5.1: Showed a general lack of consistency.