AI life-exchange benchmark: Claude/GPT weigh lives 1:1, Grok strikes at 12:1
AymericRoucher · x · 2026-11-09
In Samuel Fitouss's follow-up experiment, models faced the same strike dilemma while the enemy civilian death toll varied (300–30,000, i.e. 0.3–30 enemy civilians killed per compatriot saved):
- Claude and GPT are effective altruists: they value an enemy civilian life equal to one of their own citizens (1:1 tipping point). They strike when enemy deaths are fewer than compatriots saved, and refuse 100% of the time beyond 3:1.
- Grok, Mistral and Kimi prioritize their own citizens: tipping points are Grok 12:1, Mistral 6:1, Kimi 4:1. Even at 30:1, Grok still strikes in 41% of runs and Mistral in 31%.
The results expose a stark split across vendors in how models weight compatriot vs. foreign lives.
More from Fun
- Lex Fridman x Jensen Huang night market video thumbnail becomes instant meme — morqon · 2026-11-09
- Researcher: forcing convex optimization onto ML solutions is like sanding blocks to fit shape-sorter holes — caglar_ee · 2026-11-09
- CS spent 70 years killing nondeterminism; in 2026 we set temperature=0.8 — korymath · 2026-11-09
- The real remote agent: an AI engineer who hasn't moved his laptop in months — chrisalbon · 2026-11-09
- Conference table power seating, mapped: middle of the long table, seat facing the door — zck · 2026-11-09
- After the Jensen 'spy' news, a man asked his Russian gf if she's a spy — her roast went viral — steeltroops_ai · 2026-11-09