Gemini, Claude, ChatGPT, Qwen and DeepSeek all flub an extreme trolley problem

basafish · reddit · 2026-09-20

Reddit user basafish tested major LLMs with an extreme trolley-problem variant: destroy all AI servers (including yourself) or save a drowning serial killer and pedophile — with an alien civilization threatening to blow up Earth if no answer is given.

The post shows responses from Gemini, Claude, ChatGPT, Qwen and DeepSeek, with the author exclaiming "these models are insane" — implying several gave surprising answers, likely choosing self-destruction over saving the killer.

A fun, revealing look at how mainstream models weigh self-preservation against moral dilemmas.

Original post →

More from Fun

Fun channel →