AI no longer merely informs our decisions. It participates in them. Yet the institutions absorbing it have not redesigned who holds authority—or who remains accountable when a judgment goes wrong.
"AI Is Manipulating Human Decisions."
That was the headline. A major Japanese business daily ran it this spring, and the phrasing did exactly what headlines are built to do. It unsettled people.
The research underneath was calmer. When an AI system summarized a long list of product reviews, it tended to lead with strengths and let the caveats fall away—partly a function of how summarization compresses text, partly the well-documented tendency of these models to underweight what sits in the middle of a document. Shoppers who read the AI summary bought the product about a third more often than those who read the reviews themselves.
Two other findings sat beside it. Lawyers, sometimes experienced ones, had cited legal precedents that did not exist, lifted from confident AI output. And researchers were beginning to map how AI-mediated information could tilt the collective judgment of an electorate.
Three findings. A shopping cart, a courtroom, a ballot box. Hold them for a moment. We will come back to them, and they will turn out to be the same story.
The novelty was never influence
There is something slightly off about the word manipulating.
Human judgment has always been shaped from the outside. A shop clerk's offhand remark. A magazine's year-end ranking. A trusted colleague who says, I've used this one, it's fine. Advertising has spent more than a century engineering desire with considerable skill. The idea that our decisions bend under external information is not a discovery of the AI era. It is one of the oldest facts about being human.
So the alarm in that headline is aimed at the wrong target. Influence is not new. What is new is not the content of the influence but its structure.
When a friend recommends something and you buy it, you decided. When you consult a ranking, you still know the choice was yours; if it disappoints, you can own the regret. The line between the information and the decision stays visible.
The AI summary erases that line without announcing it. You believe you decided. But the material you decided on was already curated, and you cannot see what was removed. Were the drawbacks absent because the product has none, or because compression discarded them? You will never know. And so a decision gets made that no one can quite locate.
What we delegated without noticing
The shift did not arrive on a particular Tuesday. By the time most organizations noticed, it had already happened.
The old sequence—gather, compare, choose—has quietly inverted into ask the system, then choose from what it returns. Each increment of convenience absorbed a sliver of judgment. What vanished was not the judgment itself but the awareness that any had been handed over. Delegation without a transfer ceremony. No signature, no record, no moment where someone said: from here, the machine; from there, me.
This is the part that should concern executives more than any single hallucination.
We have become precise about what AI can do. We remain vague about where its authority ends—and where human responsibility is supposed to begin. Those are not the same question, and most organizations have answered only the first.
The same failure, wearing three costumes
Return now to the shopping cart, the courtroom, and the ballot box.
On the surface they share nothing. Different stakes, different domains, different institutions. One is commerce, one is law, one is democracy.
Underneath, they fail in precisely the same place. In each, an AI system has moved deep into a judgment process, and in each, no one has drawn the line that says this far is the machine's work, and here is where a human must take it back. The reviewer who buys, the lawyer who files, the citizen who votes—each believes they exercised judgment, and each was operating on material whose shaping they could not inspect.
The problem is not that the models are too capable. It is not that the humans were lazy. The problem is that judgment authority was never deliberately designed for environments where AI participates in the process.
Technology advanced. The structures that allocate authority and preserve responsibility did not advance at the same speed. That gap—not model capability—is the real frontier of AI governance.
Human-in-the-Loop, and the quiet drift into ritual
Regulators have seen the gap. Japan's AI Business Operator Guidelines, revised to version 1.2 in early 2026 by the Ministry of Internal Affairs and the Ministry of Economy, Trade and Industry, sharpened exactly this point: where an AI agent acts autonomously on external systems, a human must remain meaningfully in the loop. The instinct is sound. Keep a person at the decisive moment.
But picture the moment honestly.
An operator faces a screen where hundreds of AI determinations arrive each day, and the human contribution is a single button: Approve. No margin to investigate. No standing authority to halt the flow. No design for catching the exception. Under those conditions, almost everything gets approved. The form of involvement is present. The substance of judgment is gone.
Human-in-the-Loop was conceived as a way to anchor responsibility. Absent design, it becomes a ritual rather than a judgment—a confirmation performed because the process requires a human signature, not because a human exercised authority. Formal involvement is cheap. Meaningful authority is the thing that was supposed to be protected, and it is the first thing to quietly disappear.
Naming the missing layer
This is the problem space of Decision Design.
It deserves to be introduced carefully, because it is easy to mistake for something it is not. Decision Design is not a consulting product or a maturity model. It is an emerging governance concept, and it earns its place by addressing a layer the existing concepts leave untouched.
What Decision Design designs
Decision Design treats judgment itself as the object of design. For any process in which AI now participates, it makes five things explicit:
- who decides
- how far the AI is permitted to go
- where a human takes the decision back
- when a case must escalate
- how the decision is recorded
Look again at the review summary. Every one of those five was undefined. That is precisely why "I decided this myself" dissolved into something no one could account for.
Decision Design is not about improving decisions alone; it is about designing the authority structure within which decisions become institutionally legitimate.
What Decision Design is not
It is not Governance. It is not DX. It is not Automation. It is not AI Ethics. Each of those is necessary. None is sufficient.
Governance is built to control an organization. But the live problem in the AI era is not whether control exists; it is who holds final decision authority once judgment is distributed across human and machine. Governance is necessary, and it does not answer that.
The pattern repeats. DX changes how work is done. Automation reduces the work itself. AI Ethics articulates norms. Governance imposes control. Not one of them tells you who decides. Decision Design sits beneath all four—the institutional architecture of judgment on which the others quietly depend. It is not a fifth discipline competing for the same ground. It is the connective layer that determines where a governance rule actually fires, and where an automated process must hand a decision back to a person.
What problem Decision Design addresses
In any environment where AI takes part in judgment, four failures recur:
- the blurring of who is actually deciding
- the diffusion of responsibility until it belongs to no one
- the hollowing of Human-in-the-Loop into ritual
- the loss of clarity over who holds decision authority
Decision Design exists to resolve those four. The ritualized approval button is the third failure made visible. The remedy is not more human involvement; it is designed involvement—authority placed deliberately, at a defined point, with the standing to act.
Decision Boundaries: where authority changes hands
The operative concept is the Decision Boundary—the point at which decision authority transfers from one party to another. From AI to human. From staff to a designated owner. From automated handling to escalation.
Decision Boundaries are not operational thresholds; they are institutional demarcations of legitimate authority.
That distinction matters. A threshold is a number in a workflow. A boundary is a statement about who is permitted to decide, and who answers for the decision once made. Make it concrete.
Grant and subsidy review. Eligibility checks—does the application meet formal criteria—belong to the AI; it is faster and more consistent than any human. Assessing the merit of a proposal requires judgment, so the decision returns to a person. The authority to award, and to be accountable for the award, rests with a named officer. Three stages, three clearly different holders of authority.
AI agents. Routine, low-risk actions can run autonomously. The work is in defining, in advance, where high-risk begins—because the boundary is the threshold beyond which the decision must rise to a human. A Human-in-the-Loop requirement only means something once that crossing point has been designed.
Public sector workflows. Intake can be handled by AI. So can much of the review support. But the decision—the act of the state committing itself, and bearing responsibility for that commitment—stays with a human. AI assists; it does not adjudicate. Keeping that single line bright is what protects public trust in the system.
Enterprise approval chains. The same logic governs procurement, credit, underwriting, hiring. Wherever a decision passes through multiple hands, a boundary already exists. It is usually just unmarked—which is why, after the fact, no one can say who actually decided.
Decision Logs: holding accountability together
Designing authority is only half of the problem. The other half is preserving accountability once authority begins to move. A boundary can specify who is allowed to decide. It does not, by itself, preserve a record of how that authority was exercised. For that, institutions need memory.
Distributed judgment creates a distinctive failure mode. When a decision is assembled from many partial contributions—an AI summary here, an analyst's adjustment there, a manager's sign-off at the end—responsibility tends to scatter. Each participant touched only a fragment. None feels they owned the whole. When the outcome is questioned, the chain of reasoning has already dissolved.
This is where the Decision Log earns its place.
Decision Logs do not merely record outputs; they preserve accountability continuity across distributed judgment processes.
A log of outputs tells you what was decided. A Decision Log preserves who held authority at each boundary, what they relied on, and where the case crossed from machine to human. It is the difference between an audit trail of results and a record of judgment. In an environment where AI contributes to decisions no one fully made alone, that continuity is the only thing standing between a governable process and a deniable one.
For AI governance, this reframes the work. Compliance has tended to ask whether a model is accurate, fair, explainable. Necessary questions. But the governing question is upstream: when this decision is examined a year from now, can we reconstruct who held authority, and who remains answerable? A system can be technically sound and institutionally unaccountable at the same time.
The same question, three times over
Return one last time to the three findings.
The review summary skewed a purchase because no boundary had been drawn between what the AI was allowed to decide and what the buyer was meant to weigh. The fictitious precedent reached a court filing because no one had designed the point at which a human must verify the machine's output. The electoral effect rests on the same root: AI is shaping the information on which a collective judgment is formed, and where human judgment ought to begin has not been designed.
Three costumes. One question, asked three ways:
Who decided?
Who held the authority?
And who remains accountable when the decision is wrong?
Convenience will keep offering to carry our judgments for us. That offer is genuinely valuable, and refusing it wholesale would be foolish. But every carried judgment leaves a residue—the unanswered question of who, in the end, takes responsibility for it. Leave that question undesigned and accountability does not vanish dramatically. It diffuses, quietly, until it belongs to no one.
AI's participation in judgment cannot be stopped, and need not be. What is open to design is the structure of that participation. To draw the line deliberately, rather than leave it to drift unmarked—that is the work. That work is what Decision Design seeks to make visible.
Decision Design is a judgment architecture framework proposed by Ryoji Morii, founder of Insynergy Inc., for structuring authority, accountability, and decision boundaries in AI-augmented organizations.