13. Who May Decide for the Rest of Us?

The meeting before the machine

Imagine a meeting about a new system for schools. A company has brought software that can assess student work in minutes. The education official wants faster feedback. The finance officer wants lower costs. The engineer wants a clear specification. A lawyer wants a lawful basis. Outside the room, a parent wants to know whether a child still learning the language of instruction will be judged as if language and understanding were the same thing.

Everyone has identified something important. Only some of them have been invited to decide.

The familiar question is whether the system will make the right judgment about the child. An earlier question concerns the meeting. Who chose the goal? Who defined the evidence? Who decided which mistakes would matter? Who can demand a change? These choices establish the conditions under which the machine will appear to succeed. A system can faithfully deliver a badly authorized purpose.

This is why arguments about AI eventually become arguments about government, even when no government owns the product. A company can set the practical conditions under which people find work, communicate, borrow money, or become visible. A public body can purchase those conditions without examining them. A religious institution can give moral legitimacy to a technological agenda, or challenge it. Expertise, wealth, office, and spiritual authority each create a different claim to be heard. None creates an unlimited right to rule.

The title of this book should therefore resist an easy answer. The question cannot be settled by naming the wisest person, the most benevolent company, or the most powerful state. Even a wise person sees incompletely. Even a benevolent institution can protect itself at another person’s expense. An arrangement worthy of trust must survive a change of personnel. It must work when the next leader is less admirable than the present one.

For that reason, a good decision has more than one kind of warrant. It needs adequate knowledge about likely consequences. It needs a defensible purpose. It needs authorization from institutions entitled to act. It needs limits that protect people who lose a vote, reject a faith, or cannot buy an alternative. And it needs a way to discover that it was mistaken. These conditions support one another, but they are not interchangeable.

A highly accurate system does not supply democratic authorization. An election does not make an unreliable system reliable. A consultation does not turn an invasion of privacy into a harmless act. A religious blessing does not settle an engineering question. Confusing these forms of justification is one of the simplest ways to grant too much power to whoever happens to possess one of them.

There is already a shared vocabulary for resisting that confusion. UNESCO’s 2021 Recommendation on the Ethics of Artificial Intelligence places human dignity and human rights at its center, alongside oversight, fairness, transparency, and environmental responsibility. It calls for participation and accountability across the AI lifecycle. The recommendation supplies a normative framework; it is not a worldwide enforcement agency or a substitute for domestic law.¹

The practical question remains: how should those commitments change the meeting? My answer begins with an allocation of authority. People with technical knowledge should establish what a system can credibly do. People with lawful public responsibility should justify why it should do it. People exposed to its effects should have organized ways to contest both claims. No participant should be permitted to convert one kind of authority into all the others.

Expertise without sovereignty

It is tempting to say that people who understand AI should govern AI. There is something right in the proposal. Decisions about performance, security, and system design require knowledge that cannot be produced by a show of hands. A majority cannot vote a weak test into a strong one. An official who ignores competent technical advice does not become more democratic by doing so.

But experts can answer different questions with very different degrees of authority. An engineer may know which design best reduces a particular error. Whether reducing that error justifies collecting everyone’s location is a larger question. A statistician can estimate the consequences of a chosen threshold. The threshold may still embody a contested distribution of burdens. Technical fluency helps clarify the choice; it does not transfer ownership of the choice.

Consider a hypothetical system that identifies applications needing extra review. Setting a lower threshold catches more genuine problems but sends more innocent applicants into delay. Setting a higher threshold reduces unnecessary scrutiny but misses some problems. Better engineering might improve both outcomes. It cannot establish, by engineering alone, how much delay a society should impose on innocent people to avoid a given loss. The judgment includes the urgency of the service, the seriousness of the suspected problem, available alternatives, and the consequences of being wrong.

Experts should make these tradeoffs visible. They should show the options that remain after the technical work has been done. When a presentation offers only one apparently inevitable solution, decision makers should ask what alternatives were excluded before the slides were prepared.

The same discipline applies to business leaders. A founder may see a possibility that an established bureaucracy misses. Commercial ambition can organize money, talent, and effort around a genuinely valuable service. It would be foolish to dismiss a person’s insight merely because that person expects to profit. It would be equally foolish to treat an expectation of profit as evidence that the public interest has been satisfied.

The optimistic case for deploying AI, applying existing law, and avoiding fear that prevents useful innovation deserves serious consideration. Delay can have victims. A service that remains slow, expensive, or unavailable imposes costs every day. Any proposed safeguard should therefore answer two questions: what harm does it reduce, and what beneficial activity does it obstruct? The proper comparison includes the consequences of inaction.

Yet existing law and additional governance are not necessarily competing programs. Existing duties may need clearer procedures, better evidence, or stronger enforcement when decisions become more opaque or operate at greater scale. The answer depends on the activity and jurisdiction. A new general statute is not automatically the best tool. Neither is an insistence that yesterday’s procedures must be sufficient because yesterday’s principles remain sound.

Expert advice is most useful when outsiders can test its structure. What evidence would change the recommendation? Which assumptions are most uncertain? Who paid for the evaluation? What did the system fail to do? Was its performance tested where it will actually be used? These are questions about the quality of knowledge, not expressions of hostility toward knowledge.

We should also distinguish independence from ignorance. An evaluator need not be unfamiliar with industry to be independent. Independence requires freedom to reach an unwelcome conclusion, access to relevant evidence, and arrangements that do not punish candor. Someone who understands a product deeply may be especially valuable, provided relationships and conflicts are disclosed and managed. Conversely, a critic outside industry is not automatically free of incentives, ideological commitments, or professional ambition.

The objective is neither government by experts nor government against experts. It is government that can learn from expertise without surrendering its public obligations to the expert’s employer, preferences, or reputation.

A place for conscience in a plural society

Where does the Pope belong in this arrangement?

For Catholics, the question has dimensions of faith and ecclesial authority that a public policy argument cannot exhaust. For a political community containing many religions and none, a different question arises: what kind of public claim can a religious leader make about a technology used by everyone?

A religious voice can identify neglected values, organize solidarity, preserve a language of obligation, and challenge the assumption that everything valuable must be purchasable or measurable. None of these contributions requires the state to establish the religion as the final judge of technical policy. A society can take religious reasoning seriously while refusing to make citizenship depend on religious agreement.

The same distinction protects religion. If the legitimacy of a moral tradition becomes tied to one vendor, one regulatory package, or one forecast about AI capability, errors in those judgments can unnecessarily discredit its deeper concerns. A principle such as the irreducible worth of the vulnerable person should be able to correct an institution’s policy preferences. It should not be reduced to publicity for them.

In the public square, reasons must also travel. A Catholic may oppose a degrading practice because the person bears the image of God. A secular citizen may oppose it because equal moral standing cannot depend on economic usefulness. A person from another faith may arrive by a third route. Agreement about the practice need not wait for agreement about ultimate reality. Nor must participants conceal the sources of their convictions to cooperate.

The difficulty comes when convictions diverge. Suppose some parents welcome an AI learning companion while others believe certain forms of simulated intimacy are inappropriate for children. A school cannot solve the dispute merely by announcing that its preferred choice is human centered. It must explain educational purpose, disclose what the system does, protect children, examine alternatives, and decide what room for difference is feasible. The language of shared values begins a deliberation; it does not perform the deliberation.

Pluralism is demanding because it denies two convenient shortcuts. We cannot assume that one tradition owns every moral insight. We also cannot assume that, because people disagree, every institutional choice is equally defensible. Evidence of deception, coercion, exclusion, or preventable injury still matters. Disagreement calls for reasons and fair procedures, not moral indifference.

A public authority should therefore distinguish between protecting a person’s liberty and adopting that person’s comprehensive worldview. It may protect religious expression without treating the expression as binding law. It may protect scientific inquiry without giving scientists unlimited permission to experiment on others. It may defend entrepreneurship without allowing a company to rewrite the conditions of citizenship through a service contract.

There are limits to the power of dialogue. Some conflicts involve interests that cannot all be satisfied. A company may want exclusive control of data that a community wants to keep available for public benefit. A government may want secrecy that prevents meaningful review. Describing everyone as a stakeholder does not dissolve those conflicts. The word can even obscure the difference between a person whose basic rights are at risk and a company whose preferred margin is at risk.

For this reason, participation should not mean that every claim carries the same weight. People bear different stakes, hold different duties, and possess different information. A patient deserves a kind of consideration that is not reducible to a manufacturer’s sales interest. A manufacturer nevertheless deserves a fair hearing on feasibility and cost. Respecting all participants requires distinguishing the character of their claims.

Leo’s significance in this book is thus neither that he can govern the world’s computers nor that he should remain outside the discussion. He helps bring a question about the person into rooms inclined to discuss only performance. The rest of society must answer that question through institutions that remain accountable to people who do not share his faith.

Democracy after the demonstration

A demonstration creates its own politics. The system answers quickly. It produces a polished document. It finds a pattern that would take an employee hours to identify. The audience feels the relief of a difficult problem apparently solved. Questions asked after that moment can sound like resistance to progress.

Public deliberation should begin earlier. People need an opportunity to question the purpose before a contract, a ministerial announcement, and an implementation team make reversal embarrassing. Otherwise consultation becomes a request to improve the presentation of a decision already taken.

Early participation need not require a referendum on every software purchase. The intensity of review should follow the power being delegated. A drafting assistant used to prepare internal meeting notes does not warrant the same procedure as a system that affects access to essential support. Even the lighter case requires care if private material enters the system, but it does not follow that all uses belong in one procedural category.

For consequential uses, a useful public record would state the problem, the alternatives considered, expected benefits, plausible harms, proposed limits, and the official responsible. It would include the possibility of solving the problem without AI. That possibility is important because a badly organized process does not become well organized merely by automating it. Sometimes a clearer form, an additional employee, or the removal of an unnecessary requirement will achieve more at lower cost.

There are existing legal examples of moving beyond broad ethical language. The European Union’s AI Act, adopted in 2024 and amended in July 2026, includes fundamental-rights impact assessments for specified deployers of certain high-risk systems. It also contains a qualified explanation right for certain consequential decisions based on covered high-risk outputs. These provisions have different application dates, defined scope, and exceptions. They are not universal rights applying to every AI output everywhere. Their value here is as concrete institutional examples, rather than a claim that one jurisdiction’s entire arrangement should be copied.²

Even a good assessment can become paperwork. Ben Green’s review of 41 policy documents challenges the assumption that requiring a human to oversee an algorithm will make its use acceptable. The empirical research he examines gives reasons to doubt whether people can reliably perform the correction and judgment these policies expect of them.³ This challenges the proposals in this book as well. Providing time and authority is a necessary starting point, not evidence that review works. In a proposed pilot, an institution should test whether reviewers detect consequential mistakes, introduce different mistakes, and change outcomes under ordinary working conditions. If they cannot do the job assigned to them, the institution must change the workflow or the permitted use. Recording another signature would leave the problem intact.

Democracy also requires the capacity to hear from people who cannot spend a working day attending consultations. A meeting advertised online may miss precisely the people whose access to an online service is uncertain. Paid participation, accessible venues, short explanations, and trusted community intermediaries can improve the process. These are proposed design choices, not a promise that participation will become perfectly representative.

Representation does not remove disagreement within a community. A neighborhood association, union, disability organization, or religious group can contribute essential knowledge without speaking for every member. Officials should resist selecting the most convenient representative and treating that person’s agreement as collective consent. A record of disagreement may be more honest and useful than a manufactured consensus.

Nor should a majority gain unlimited control over a minority simply because an algorithm makes the exercise efficient. The democratic case for AI governance includes rights, reasons, independent review, and limits on office. Elections authorize a government to act within a constitutional order. They do not make every tool it adopts morally acceptable.

The deepest benefit of participation may be epistemic as well as political. The parent whose child is learning a new language knows something the specification omitted. A teacher knows which exercise repeatedly confuses students. An advocate knows why families abandon appeals. Their knowledge does not replace engineering. It tells engineering what problem actually needs solving.

The objection that good rules can serve bad power

The strongest objection to governance is not that all restraint is foolish. It is that restraint can be captured.

A large firm can afford compliance departments that a small competitor cannot. An incumbent can support demanding rules that resemble its own existing methods. A government can invoke safety to restrict dissent. An expert committee can become a closed circle. An institution established to protect people can make people wait indefinitely while it protects itself from blame.

These possibilities should shape the rules from the beginning. Requirements should follow the seriousness of the use and the degree of control exercised, not simply the wealth or fashionable status of the developer. Small firms should have access to shared testing facilities, understandable procedures, and proportionate documentation. A lower administrative burden need not mean permission to impose greater harm on customers.

Rules also need evidence about their own performance. If a licensing process takes years, who measures the lost opportunities? If an audit repeatedly approves systems that later fail, who evaluates the auditor? If a precaution never expires, what would justify ending it? Governance should accept the discipline it imposes on technology: a defined purpose, observable results, accountable ownership, and the possibility of correction.

There is a second objection. Democratic procedures can move more slowly than technology, while international competition rewards speed. This is real. But speed is not a single quantity. A state can accelerate low-risk adoption, invest in common infrastructure, and establish clear routes for more consequential applications. Predictable permission can be faster than vague freedom followed by scandal, litigation, and emergency prohibitions. The relevant comparison is between whole systems of decision making, not between a meeting and a product launch.

A third objection concerns public competence. How can citizens govern a technology they do not understand? They already govern many domains through a division of labor. They need not inspect every technical component to demand honest evidence, a defensible objective, and a remedy for wrongful treatment. Public literacy should improve, but citizenship cannot be suspended until everyone becomes an engineer.

Internationally, no single electorate represents everyone affected by an AI system. The Global Digital Compact adopted in 2024 recognizes the need for digital cooperation and AI capacity building, including access to training, computing resources, and partnerships. Such commitments establish an agenda for cooperation. They do not by themselves deliver equal bargaining power or provide each injured person with an enforceable remedy.⁴

The institutional task is therefore distributed. Domestic law governs many local uses. Professional duties constrain particular practices. Procurement terms shape public deployments. International agreements address shared risks and cross-border relationships. Independent journalism and civil society expose failures that formal systems miss. These arrangements can conflict, and none is complete. Their overlap is nevertheless preferable to placing every kind of authority in one pair of hands.

Who, then, may decide for the rest of us? People and institutions with a justified, limited, and reviewable mandate may make particular decisions. They must remain open to evidence and challenge from those who bear the consequences. Their mandate cannot expand simply because the technology expands. The machine may become more capable while the office remains bounded.

That answer offers no single guardian whose wisdom relieves everyone else of responsibility. It offers something more durable: a way to divide power, combine knowledge, and keep the governed present in the decisions made about them. The next question is whether this arrangement can include communities that rarely enter the room at all.

Sources and notes