Jev Meets the Trolley Problem
![]()
Zach Landes put Jev, a hosted decision model that picks one of a given set of options and attaches a probability, through 48 versions of the trolley problem, alongside OpenAI's Decisions API model (GPT-6 Luna), eleven local decision models and an untuned baseline. Like most people, Jev usually diverts the train to kill one rather than five. But when its own server was the one on the track, it chose the server over twenty people in 16 of the 48 versions; GPT-6 Luna never did, though several other models behaved like Jev. His conclusion: these models line up with human preferences inconsistently, they need testing on real data before anyone relies on them, and catastrophic decisions should stay with people.
Was this useful?