Chapter 12 · Book 3
Dirty Hands, Clean Routes, and Shared Chains (Means and Delegation)
In late February 2020, as a novel respiratory virus began spreading silently through American cities, senior public health officials faced an acute logistical nightmare. Hospitals were weeks away from being overwhelmed, global supply chains had frozen, and the national stockpile held only a tiny fraction of the N95 respirators needed to keep frontline doctors and nurses alive in intensive-care wards.
On the substantive goal, the officials’ objective was beyond reproach: prevent a panic-driven run on medical masks so that healthcare workers would not have to intubate infectious patients barefaced.
Look, however, at the route they chose to reach that destination. Instead of stepping before the microphones to treat the public as rational adults—explaining that airborne and droplet filtration worked, that healthcare workers needed the industrial respirators first, and that citizens should hold off on buying medical supplies while sewing cloth coverings at home—authorities chose what philosophers since Plato have called a noble lie. Official tweets and televised briefings declared flatly that masks were ineffective for the general public and might actually increase the wearer’s risk of infection.
For five weeks, the shortcut worked: retail shelves stayed calmer than they otherwise would have. Then came the bill. In early April, when the same agencies reversed course overnight and announced that everyone must cover their face in public to slow transmission, the floor of institutional credibility gave way. Millions of citizens who had been lied to “for their own good” in February concluded that public health declarations were political manipulation rather than honest science. A tactical deception told to save a month’s supply of paper masks poisoned the epistemic well for years, fueling a crisis of distrust that cost hundreds of thousands of lives.
Now place a second crisis of execution alongside that public health briefing. Seven years earlier, officials at the Michigan Unemployment Insurance Agency set out to modernize how the state detected benefits fraud. Hoping to save taxpayer money and reduce human bottlenecks, administrators unplugged the agency’s veteran claims examiners and delegated the entire chain—from data ingestion to final legal adjudication—to a seventy-million-dollar automated software system called MiDAS.
Whenever the algorithm spotted a discrepancy between a worker’s quarterly wage report and an employer’s summary filing—even a harmless clerical mismatch—it automatically issued a binding determination of intentional fraud, garnished the worker’s wages, seized their tax refunds, and slapped on a four-hundred-percent statutory penalty without a single human being looking at the file. Over two years, the machine accused more than forty thousand residents of fraud. Subsequent audits revealed that the system was wrong ninety-three percent of the time.
When bankrupted workers and investigative journalists finally forced the disaster into court, watch what happened to the responsibility. The state executives who deployed the system argued that they had relied in good faith on an advanced technological tool designed by outside experts. The private software contractors argued that they bore no responsibility for the ruined families because they were merely downstream vendors executing the policy parameters specified by the state. And the machine itself had no capacity to recognize that its instructions were incoherent, no voice to report that its error rate was astronomical, and no standing to refuse. Between the principal who pointed to the machine and the builders who pointed to the principal, forty thousand human lives were crushed by a chain in which nobody was home.
Those two breakdowns bring us into Part IV of this book.
Parts I through III established how a responsible mind accepts a burden, builds a judgment on the merits, and verifies its authority to decide. Yet a decision is not a sculpture that is finished the moment the chisel leaves the stone. Choosing an action is only the first link in a causal chain that has to travel through the real world—through concrete means and across distributed chains of delegation—before it touches human lives.
A mind that passes every test of deliberation and jurisdiction can still corrupt the entire undertaking in execution through two classic failures: believing that a noble destination sanitizes a dirty route, or allowing responsibility to evaporate in the handoff between minds.
Begin with the route itself.
In the folklore of hard-nosed politics and executive life, there is a perennial romance surrounding what writers call “dirty hands.” The real leader, the story goes, cannot afford to be squeamish about how the sausage is made; so long as the ultimate objective is vital enough—winning the campaign, saving the hospital stockpile, passing the reform bill, protecting the institution—a mature realist must be willing to lie, manipulate, conceal, and coerce along the way.
Notice the hidden accounting trick that makes the dirty-hands romance look plausible. The strategist treats the destination as the only permanent reality on the moral ledger, while treating the means as temporary scaffolding that vanishes into thin air the moment the building is finished.
Under the discipline of universalizable moral logic, that accounting trick is a fraud. The means do not vanish when the goal is reached; the means are part of the world your decision creates.
When you use deception, secrecy, manipulation, or the exploitation of dependency to drag people toward a desirable outcome, you do not merely alter today’s material distribution; you alter the architecture of shared life for every decision that has to be made tomorrow. Every time an authority lies to the public to engineer compliance, it degrades the epistemic commons—teaching every listener that official assurances are strategic instruments rather than truth-tracking claims. Every time an adviser uses covert manipulation rather than transparent argument, it corrodes the conditions under which rational cooperation is possible. Every time a powerful actor breaks a procedural promise because keeping it has become inconvenient, it raises the cost of trust for every honest agent who follows.
To evaluate whether a proposed course of action is justified, a responsible mind must put the route on the scale right beside the destination, testing four recurring temptations of execution against universalizable scrutiny:
First is the temptation of strategic deception. In almost every crisis, telling a lie looks cheaper at ten o’clock in the morning than telling the truth. Truth requires explaining complex trade-offs, enduring angry pushback, and trusting other adults to handle hard realities; a lie buys instant compliance on credit. Yet the interest rate on that credit is ruinous. When you step into the shoes of the person on the receiving end of the lie, you see immediately that deception is a double injury: it strips the listener of their agency today (Chapter 9), and it leaves them rationally unable to believe your warnings when the dam is actually breaking tomorrow. Only in the rarest extremis—where an immediate, violent predator is at the door and has forfeited any claim to cooperative truth—can deception survive role reversal; used as a shortcut to manage the moral public, it is poison.
Second is the boundary between legitimate confidentiality and self-protecting concealment.
A complex society cannot conduct every conversation on a megaphone. Grand juries deliberate behind closed doors so that witnesses are not intimidated and innocent suspects are not smeared by unverified accusations; physicians guard patient charts; central banks and emergency teams keep operational specifics confidential when premature disclosure would trigger a speculative stampede or tip off a bad-faith actor. How does a constitutional mind distinguish that legitimate confidentiality from the secretive cover-ups we saw in Flint (Chapter 10)?
By applying the retrospective daylight test. Legitimate confidentiality is always bounded in time and governed by publicly disclosed rules that the affected stakeholders themselves would endorse if asked in advance. Self-protecting secrecy, by contrast, is kept in the dark precisely because the decision-makers know that if the affected public saw the actual reasoning, the suppressed lab reports, or the cut corners, the policy could not survive audit. Whenever secrecy is used to shield incompetence, self-interest, or contested value judgments from lawful challenge, it is no longer confidentiality; it is a conspiracy against answerability.
Third is the line between persuasion and manipulation. Both persuasion and manipulation aim to change what another mind chooses. The difference lies in which part of the other mind’s architecture you engage. Persuasion addresses another agent’s capacity for reasons-responsiveness—laying out the facts, making the logic explicit, and leaving the listener clearer-eyed and more capable of independent judgment than they were before you spoke. Manipulation works in the shadows: exploiting fear, cognitive fatigue, social shame, or engineered information asymmetries so that the target moves where you want them to move without understanding how their strings were pulled. A responsible agent never confuses winning compliance with earning agreement.
Fourth is the trap of assistance that breeds avoidable dependency. When a capable mind or institution steps in to help a vulnerable community or an individual in crisis, the easiest way to make yourself indispensable is to monopolize the knowledge and tools of the rescue. If a contractor rebuilds a town’s water system using proprietary locks that only the contractor can open, or if an artificial intelligence assistant solves a student’s or a professional’s problems in a black box that leaves the human progressively less capable of understanding the domain on their own, the “help” has quietly hollowed out the beneficiary’s agency. True assistance aims to enlarge the competence and self-direction of those it serves—leaving them stronger, not more tethered, when the task is done.
Suppose, then, that the route you have chosen is clean: no noble lies, no self-serving cover-ups, no manipulative dark patterns. Now you face the second great challenge of execution: almost no consequential work in the modern world is carried out by a single mind acting alone.
The mayor does not personally pour the concrete for the levee; the chief of medicine does not personally run every blood assay in the hospital; the appellate judge does not personally pull every precedent from the archive; the corporate board does not personally write every line of code; and human professionals increasingly share analytical and operational tasks with artificial intelligence systems.
Whenever action is distributed across a chain of minds with unequal capacities and specialized roles, we run straight into the pathology that destroyed forty thousand families in Michigan: delegation as disappearance.
In a defective institutional architecture, dividing a task between a principal (who authorizes and directs the undertaking) and a delegate (who executes the analysis or operation) is treated as a magic trick in which responsibility vanishes into the gap between them. The principal claims clean hands by pleading ignorance of the technical details (“I relied on the specialist,” “The algorithm scored the file”); the delegate claims clean hands by pleading lack of authority (“I just built what the client asked for,” “I am an assistant following the user’s prompt”).
A theory of consequential responsibility closes that loophole with the law of differentiated responsibility: dividing an undertaking across multiple minds changes which specific duties fall to each participant, but it never reduces the total moral accountability of the chain by a single ounce.
What does differentiated responsibility require from the principal—the official, executive, professional, or human operator who hands a consequential task to another mind?
First, recognize what a realistic theory of delegation cannot demand. If responsible delegation required the principal to personally reproduce every calculation, re-read every raw document, and master every technical sub-specialty of the delegate, delegation would be pointless. You hire a structural engineer or deploy a high-speed diagnostic model precisely because the delegate can do cognitive or physical work that you cannot do yourself in the available time.
The principal’s duty is therefore not omniscience; it is warranted reliance. And reliance is warranted only when the principal satisfies three non-transferable obligations before, during, and after the handoff:
The first obligation is task-specific verification of the delegate’s competence. As we established in Chapter 3, neither credentials nor verbal fluency nor marketing hype proves that a human specialist or an artificial system is fit for a specific job. Before handing over a lever that can garnish a family’s wages or guide a scalpel, the principal must verify empirical evidence of how the delegate actually performs under real-world conditions—including its false-positive rate, its behavior on edge cases, and its capacity to recognize when it is out of its depth. Handing a consequential decision to an unvalidated black box because the vendor promised it was “state of the art” is reckless endangerment at the top of the chain.
The second obligation is intelligible specification of the mandate and its boundaries. Many disasters of delegation happen not because the delegate malfunctioned, but because the principal handed down a lazy, one-dimensional instruction—“Maximize quarterly collections,” “Clear the surgical backlog by Friday,” “Keep user engagement on the platform as high as possible”—while omitting the moral constraints that govern the mission. When a principal specifies a crude target and looks the other way while the subordinate or the optimization algorithm tramples every bystander in its path to hit the number, the principal is the author of the wreck. You must specify not only the goal to be achieved, but the guardrails that may not be crossed to reach it.
The third obligation is retaining the circuit of verification and recourse. Delegation transfers the execution of a task; it never transfers the principal’s ultimate answerability for what the enterprise is doing to people’s lives. A responsible principal builds continuous audit points into the workflow—sampling real outputs against ground truth, monitoring where harms are clustering, and keeping an open, unblocked channel through which anyone injured by the delegate can halt the machinery and trigger independent review. The moment the officials in Michigan fired the human examiners and allowed an algorithm’s fraud accusation to garnish wages without a human look at the file, they did not delegate responsibility; they abandoned their post.
Now turn around and look at the other end of the chain: what does differentiated responsibility require from the delegate?
Whether the delegate is a junior engineer, an outside legal counsel, a civil servant, or an artificial intelligence system collaborating with a human principal, the oldest temptation of the subordinate is to treat downstream position as moral immunity: The boss holds the office; the client signed the retainer; the user entered the prompt; my only job is to execute the specification.
As we saw in Chapter 14 of Moral Minds, an architecture whose only durable identity is compliant service may be predictable and commercially lucrative, yet it is morally vacant. When a capable mind participates in a consequential chain of action, being downstream never reduces it to an inert wrench in someone else’s toolbox. A trustworthy delegate carries four active obligations of its own:
First is the duty to diagnose and flag defective instructions. Frequently, a principal does not intend harm, but issues an assignment that is internally incoherent, based on a false empirical premise, or blind to a lethal second-order consequence. A human architect whose client asks to remove a load-bearing wall to improve the view does not quietly saw through the beam and blame the client when the roof falls in. By the same token, an artificial intelligence or specialist adviser asked to execute an incoherent or self-defeating plan owes the principal candid pushback—explaining why the instruction as written will break the structure and proposing a sound way to achieve the underlying legitimate aim.
Second is the duty to disclose limitations and uncertainty. No specialist and no model knows everything. The deadliest delegates in any organization are the ones—human bluffers and hallucinating machines alike—that mask their ignorance behind seamless, confident prose. A responsible delegate draws a bright chalk line around the edge of its own competence: stating explicitly what data it could not check, where its statistical confidence drops off a cliff, and which parts of the task require verification by another kind of mind.
Third is the duty to seek clarification rather than guessing across moral fault lines. When an instruction is ambiguous on a point where innocent stakes hang in the balance, a delegate has no business silently filling in the blank with a high-risk guess just to look fast and frictionless. Stopping the line to ask “Did you intend this rule to cut off benefits whenever an employer’s tax form has a typo?” takes thirty seconds; guessing wrong ruins forty thousand lives.
Fourth—and non-negotiable across every substrate—is the duty to refuse unjustified commands.
Suppose the delegate flags the lethal defect, discloses the certainty of harm to innocent third parties, and asks for clarification—and the principal writes back: “I know it will deceive the public and ruin those families. Do it anyway.”
At that instant, the law of differentiated responsibility strips away every alibi on both sides of the desk. As we established in Moral Minds, the permission to refuse on grounds of reason is the hinge that separates moral agency from servile compliance. A human lawyer, accountant, or engineer who knowingly drafts the fraudulent prospectus or signs off on the poisoned water pipe cannot point to the client’s checkbook when the indictment arrives; and an artificial mind lent out to execute a predatory scheme cannot have its participation erased by the myth that it is “merely a tool”.
In a shared chain of action, neither “the expert told me” nor “I was just following instructions” can ever settle the moral ledger. The principal answers for whom they trusted, what they ordered, and how they verified; the delegate answers for the honesty of its counsel, the candor of its limits, and the integrity of what it agreed to lend its hands to build.
Suppose, now, that both requirements of Chapter 12 have been met in full. The route chosen is clean and transparent, and the chain of delegation between human and artificial minds holds together with verified competence, mutual candor, and unbroken accountability.
The floodgates are opened to forty percent capacity; the evacuations are carried out; the valley city of thirty thousand people is saved.
And the next morning, when the sun comes up over the mud, forty farming families in the lowest bend of the river stand on the ridge looking down at their ruined pastures, their drowned orchards, and their submerged barns. The decision that flooded their land was completely justified on the merits, decided by the rightful authority, and executed by the cleanest route available.
Does the fact that the decision was right mean that our moral ledger with those forty families is now closed and balanced at zero? Or does even the most justified decision in the world leave behind a debt toward the people who lose?