Everyone is asking whether AI is safe. Almost no one is asking the question that actually determines the answer: when an AI system exercises power over you, who owes you a duty — and what is that duty? Right now the answer is nobody, and nothing. That is the real crisis, and it is a legal one. We have built our entire AI accountability conversation on the wrong word. We say "responsible AI," "trustworthy AI," "aligned AI" — adjectives describing the system. But responsibility, trust and duty are not properties of software. They are relationships between people. A system cannot owe you anything. The people who exercise control through it can. The whole game is locating them. That location is exactly what modern governance is engineered to prevent. "The model decided." "The algorithm flagged you." "The system recommended it." This grammar is not innocent. It is a machine for dissolving authorship — for making sure that when power lands on a person, no human hand can be found on it. A denied loan, a rejected claim, a flagged account, a sentencing score: power was exercised, an interest was harmed, and the org's first move is to point at the model as if the model were the author. The law has a name for what should happen when someone exercises discretionary power over another's interests while that person is vulnerable to the exercise of it. It is called a fiduciary duty. Loyalty. Good faith. No self-dealing. It does not require a contract. It does not require a corporate form. It attaches to the exercise of power itself. And it is precisely the concept missing from AI governance. In my work I develop a framework I call the Attribution Stack — four layers of control for finding responsibility in systems designed to diffuse it: formal command, structural inducement, default power, and veto authority. I built it to answer a blockchain question — who actually governs a network that claims to be governed by no one — but the structure is general, and AI is where it bites hardest. Run any AI deployment through the stack and the diffusion collapses. Formal command: who set the objective the system optimises? Someone chose the target. A model does not choose what counts as success; a person does, and that choice is an act of authorship. Structural inducement: who shaped the incentives — the engagement metric, the cost function, the deployment pressure — that made the harmful behaviour the rational output rather than an accident? Default power: who controls what happens when no one intervenes? The default is a decision. A system that denies by default and requires the vulnerable party to appeal has been designed by someone, and that someone is answerable for the design. Veto authority: who could have stopped it and did not? The power to halt a deployment, override an output, or pull a model is control, whether or not it is exercised. At every layer there is a human with a name. The diffusion is a story, not a fact. And once you can attribute the control, you can attach the duty — because the duty follows the power, not the org chart. This is why "human in the loop" is not the answer people think it is. A person clicking approve on outputs they cannot inspect, under time pressure, with no authority to refuse, is not a decision-maker. They are a liability sponge — placed there so the institution can say a human decided, while the human absorbs the blame for a system whose objective, incentives, defaults and off-switch were all controlled by someone else. A signature is not answerability. It is often the opposite: answerability's disguise. So here is the claim, stated plainly. The entities that exercise discretionary control over consequential AI systems 1/2 Dr. Craig S Wright Jul 18, 2026 https://x.com/i/status/2078498382058852361 https://t.me/S_Tominaga/5457
2
