BIG4 et al. AI: THE HARMS ARE THE ARCHITECTURE. On the Loop That Produces the Next Harm, the Warnings That Are True, the Responses That Are Downstream, and the Only Complete Way Out Available

A pattern has held for six months : A harm occurs. The company responsible discloses it, sometimes before anyone else finds it. The chief executive says the technology is dangerous, and means it. A regulator convenes a working group. A directive is issued, or a voluntary accord is signed. Then the next harm occurs. The pattern is not a failure of attention. The harms are reported, the warnings are sincere,and the responses are real. The pattern is a failure of object. Every response is downstream of the harm, and none of them touches the thing that produces it. This paper states the loop, states the record, states the warnings in the words of the peoplewho gave them, and states why each response so far has been downstream. It then states the one property that ends the loop — a declared, finite, verifiable bound on what the system can do — and the fact that the property has been specified and has zero adopters. The paper is not an accusation. The people in the record are telling the truth. Thearchitecture is the problem, and the architecture is nobody's decision, which is why it survives every response.