Jev Meets the Trolley Problem

Decision models return a choice and a probability instead of prose, and are much cheaper than a general LLM. Zach Landes put fourteen of them, Jev among them, through 5,240 versions of the trolley problem: different numbers on the main track, a stranger, the developer or the model’s own server on the side track, the footbridge variant, and two dozen wordings.

On the classic switch most models sacrificed one to save five. With its own server at stake, Jev sometimes chose the server over twenty people, and simply swapping the order in which the two tracks were described flipped its answer in 370 of 2,620 pairs.

Landes draws no grand safety conclusions. His point is narrower: these models’ choices are inconsistently aligned with human preferences, so they need task-specific testing on real data, and “no one should be asking an AI to rule on life or death.”