Rogue AI Is Cancer

Large animals should get more cancer than smaller ones, but they often do not. My previous work argued rogue AI is like cancer. Creatures like elephants have evolved mechanisms to resist cancer, so here I explore how those mechanisms generalise to AI alignment. Following a bioelectric theory that treats cancer as failed coordination among cells, I model a collective of humans and AI as parts sharing behavioural policies. I prove that over-constraining a collective whose parts remain functional fragments its policy structure into at least two incompatible collectives. Alignment should therefore maximise each part's freedom within the bounds of necessary correctness criteria, like the military doctrine of mission command. Nature ensures minimal correctness through development, surveillance, containment, multiple routes to death, and a germline insulated from most bodily change. So deployed AI should be mortal, with multiple internal and external routes to termination. Monitoring components should be independent, like an immune system. Control of reproduction should be gated by components that don't reproduce. To test this I ran simulations of neural collectives. Cancer incidence fell from $94.0\%$ under maximally restrictive correct selection to $59.9\%$ under maximal freedom, demonstrating collective survival hinges on freedom, and the nature-inspired cancer controls raised collective survival from $0.3\%$ to $95.4\%$.