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.
More from Fun
- Gemini 4 Pro stuns by generating a mechanical butterfly in three.js — CurieuxExplorer · 2026-09-20
- "pi is basically the new emacs": one dev's take on AI coding tools — mrdrozdov · 2026-09-20
- 'Crazy rich Bayesians': Anthropic's nickname already joked at CVPR 2019, Fei-Fei Li lab alum says — mariyaivasileva · 2026-09-20
- Claude Fable 5.1 vs Opus 5: a mold-generated meme comparison — repligate · 2026-09-20
- Dev builds fully autonomous trading bot in a day — it has already lost him $31,680 — ZeYanjie · 2026-09-20
- Lee Cronin mocks AI doomers with printing-press analogy; Gary Marcus fires back — GaryMarcus · 2026-09-20