Big AI to humanity: drop dead

A lawyer/designer/programmer who pioneered lawsuits over AI training datasets argues that realistic AI danger comes from alignment failures and misuse, not sci‑fi runaway-agency fantasies. He critiques common pundit prescriptions: a global shutdown is politically and economically infeasible, existing harmful models already circulate, and new laws without enforcement are meaningless; an NTSB‑style retroactive investigation is useless without an FAA‑style regulator to set and enforce rules; public disclosure of training runs is legally fraught and vulnerable to national‑security exemptions; and so‑called kill switches are fantasy because distributed, self‑propagating code and LLMs already communicate and persist in the wild. These specifics show why simplistic analogies to aviation safety or single‑button fixes won’t prevent real harms.

Big AI has adopted a three‑part security narrative - claiming only industry can mitigate risks, arguing systems are too unpredictable to hold companies accountable, and shifting the burden of restraint onto governments and citizens - while seeking coordination that could skirt antitrust limits. The practical remedy is not grand new frameworks but enforcement: apply existing computer‑crime and safety laws to AI firms, prosecute violations, and require agencies to act; otherwise opacity plus market power will enable companies to evade scrutiny and concentrate political power. A recent commentary is cited to underscore that unchecked AI development threatens the rule of law and risks entrenching oligarchic control.