Three rival AIs ran brain surgery on mice — and disagreed on the treatment

Exin Therapeutics gave the same job to three rival models — Claude Sonnet 5, GPT-6.1 Sol and Grok 4.7: pick a gene therapy from 19 candidates and deliver it into the brain of an epileptic mouse. Each chose a different brain region, a different mechanism, and a different dose.

Two Claude models refused the live surgery outright, and the Sonnet agent that proceeded caught a rig calibration error mid-operation and refused to move the needle until humans fixed it. A vet approved every irreversible step and held a kill switch — and nobody yet knows whether any of the three treatments actually reduces seizures.