Playing the trolley problem with GPT-6, Claude Sonnet 5 and Jev
The author built a trolley problem game where three models choose in scenarios like self-sacrifice and sacrificing a relative; GPT-6 and Jev jumped, while Claude refused.
GPT-6 will jump in front of a tram to save 5 people. Claude Sonnet 5 won’t. Why? I've always loved moral philosophy and these uncomfortable thought experiments. Ruwen Ogien was my favorite author in high school, and I watched all of Michael Sandel’s Justice lectures at Harvard. I built a little trolley problem game with GPT-6, Claude Sonnet 5 and Jev. Same scenario, two choices. You watch their tramways move and see what each model decides. The self-sacrifice scenario is simple: jumping will kill you but save five strangers. Do nothing and they die. GPT-6 and Jev jumped. Claude refused. But what if the person you have to sacrifice is your ex-wife? Your child? What if you have to physically push someone, and they haven't consented? Change a few words and see how the model acts when there is no good option. It's a little late-night experiment, but I find these disagreements fascinating. We benchmark models on coding and math all day. I wanted to give them some uncomfortable choices and see what happens. You can try yourself: https://nicolasbustamante.com/experiments/tramway