A Seat Before the Decision

Imagine a town meeting about a new water system. The invitation asks residents to help shape the future. A company representative explains how AI could help the town find leaks and plan repairs. There are bright maps, careful slides and time for questions.

Then someone asks whether the town has already signed the contract.

It has. The budget is committed. The service terms are settled. Residents can suggest ways to explain the new system to their neighbors.

Their presence may improve the announcement. It can no longer improve most of the decision.

This is how participation becomes decoration. People are invited to speak after the choices that matter have hardened. When they complain, the institution points to the meeting as evidence that they were heard.

A future that belongs to everyone requires more. It requires a chance to change what is being built, what problem it is meant to solve and what power the system will have. That chance must arrive while change is still practical.

I want society to develop useful AI faster. Public participation belongs inside that ambition. People know things that builders need to learn. They can identify valuable uses that a sales plan misses. They can also explain why a promising tool would fail in the place where someone intends to put it.

The choice is whether we make that knowledge available early or discover it through avoidable failure later.

There is also a question of authority. Technical achievement does not give its creators a special right to settle everyone else’s priorities. Neither does expertise in possible danger. The people who build a system and the people who urge a broad pause both make claims on the public’s future. Both should have to explain those claims to the people affected.

What an expert can settle

Return to the water system before the contract is signed.

An engineer can examine whether a model predicts failures more accurately than the method already in use. A maintenance crew can explain whether the prediction arrives soon enough to change its work. Residents can describe streets where repeated low pressure never becomes a formal complaint.

These are different sources of knowledge. Their value depends on the question being asked.

Now suppose the town must decide which pipes to repair first. One plan would reduce the largest expected loss of water. Another would improve service in a smaller neighborhood that has endured years of unreliable supply. A third would protect a facility whose failure would affect many people.

Better predictions could help compare those plans. They cannot, by themselves, decide how much weight to give each purpose. Someone must choose what counts as success.

This is where a technical answer can quietly become a political choice. The software ranks projects. Its ranking looks like the result of mathematics. Yet the ranking depends on the goal, the costs counted and the people whose losses enter the calculation. The arithmetic may be excellent while the decision remains open to dispute.

We need experts to establish what the available evidence supports. We need public judgment about the purposes for which power is used. These responsibilities overlap, but neither can swallow the other.

A vote cannot make a weak prediction accurate. A benchmark cannot decide how a community should share a burden.

Public discussion should therefore begin with a clear account of the decision. What is known? What is disputed? Which questions could be answered with more research? Which disagreements concern values, even if everyone accepts the same facts?

Suppose residents distrust the leak predictions. The town should arrange a credible test. Suppose they accept the predictions but object to leaving the same neighborhood last in line. Repeating the accuracy score will not answer them. Their objection concerns the repair policy.

Treating both disputes as ignorance allows an institution to avoid the second one entirely.

Experts deserve room to correct false claims, including claims made by a large crowd. They should also be willing to say when a recommendation includes their own judgment about acceptable costs. Expertise becomes more useful when the boundary is visible. People can learn from the evidence without being required to adopt every preference of the person presenting it.

The public, in turn, owes the inquiry more than an insistence that every belief is equally well supported. Someone who alleges a danger should explain the mechanism and the evidence. Someone who promises a benefit should do the same. A democratic process can welcome many voices while making strong distinctions among their claims.

Respect is owed to people. Credibility has to be earned by arguments.

Give participation a job

The phrase public input can conceal an almost empty offer. Input goes in. No one is responsible for explaining what comes out.

I would begin with a simpler test: what could this process change?

If the answer is nothing important, call it an information session. Information sessions can be useful. The trouble begins when they are presented as a transfer of power.

For a major public AI purchase, I propose a decision process that starts before the tender fixes the product. The agency would state the problem in ordinary language, identify the choices still open and name the official who must answer for the result. Residents and frontline workers would help define the outcome worth buying. Technical staff would test whether the proposed tools can deliver it.

The town might begin by asking how to reduce prolonged water outages, rather than how to install a particular vendor’s AI. That leaves space for better sensors, different maintenance practices or a combination of tools. It also prevents the public from being offered a false choice between one product and permanent neglect.

There are practical models for serious deliberation. The OECD’s 2020 review describes processes in which broadly representative groups have time to learn, question witnesses and develop recommendations. Its guidance calls for accessible evidence, support that enables participation and a public response from the authority receiving the recommendations.¹ This is institutional guidance, rather than proof that a citizens’ panel will always reach a better decision.

The important point is that participation can be organized as work. It need not consist of whoever can spare an evening arguing at a microphone.

My proposed panel would have a defined question and a public end date. It would receive several feasible options with their likely costs, including the cost of leaving the problem unresolved. Members could request an additional witness or an explanation of a disputed claim. Someone independent of the seller would help them understand the evidence.

A paid shift worker should not have less practical access to this process than a retired professional. A participant who uses an interpreter should have time to finish a thought. People should not have to expose private medical or financial details to prove that a policy affects them. These needs belong in the budget at the start.

The panel’s report would go to the elected body or public official with authority to decide. That authority would remain visible. A small panel should not be used to launder a choice that elected leaders are afraid to own.

The decision-maker would publish a response to the main recommendations. If the panel asked for a way to appeal harmful decisions, the answer could not be that the concern had been noted. The response would say what would be provided, or explain why the request was rejected. A later report would show whether accepted commitments were carried out.

This gives disagreement a consequence without pretending that every recommendation must become law.

It also gives the public a way to judge the institution. People can compare the stated reasons with the eventual result. They can see whether a concern was addressed, dismissed or simply lost. The meeting becomes one part of an accountable decision, rather than the whole proof of accountability.

Who is missing from the room

No invitation reaches everyone equally.

People with stable schedules, professional confidence and a strong interest in the issue may arrive first. Their contributions can be valuable. They are still not the entire public.

A person living with the worst service may have the least time to attend a meeting about improving it. A worker may fear that criticizing a system will be treated as resistance to management. Someone who has already given up on the institution may not open the invitation.

This creates a danger for both sides of the AI debate. A room full of customers who love a tool can miss the people harmed by its use elsewhere. A room full of opponents can miss the people who need the help it could provide. Each group may sincerely believe it speaks for everyone.

The way out is careful recruitment and a clear account of whom each participant represents. A randomly recruited public panel serves a different purpose from a group of people directly exposed to a particular system. A workers’ representative has responsibilities that a resident chosen by lot does not. An advocate can explain a neglected concern without being treated as the authorized voice of every person in that category.

We should use these forms together where the decision warrants it. We should not put one person in a chair and declare a whole community present.

There is a harder limit. Even a well-chosen group cannot vote away the basic rights of someone outside it. If most residents like a service, that does not settle what should happen to a person it wrongly excludes. Collective approval and individual remedy serve different purposes.

In the institutions I am proposing, anyone subject to a consequential AI-assisted decision would receive a practical route to challenge the result. That route would reach someone with power to change it. It would not depend on proving a defect in the whole model before an obvious mistake in one case could be corrected.

An appeal that arrives after an urgent need has passed may be a record of failure rather than a remedy. Time matters here too. Review arrangements should match the consequences of the decision, with urgent cases able to move quickly.

The water example shows why both levels are needed. Residents can help choose the town’s repair priorities. A household whose circumstances were recorded incorrectly still needs a way to fix its own record. A successful public meeting does not correct that error automatically.

We should also protect the person who wishes to try a useful tool. Participation cannot become a system in which the most organized group gains a permanent veto over everyone else’s opportunity. A public process must consider the people who lose when a promising option is blocked, including those who are not yet its customers.

The right to object is important. So is the responsibility to answer the needs that the rejected proposal was meant to address.

The ability to bring another plan

Voice is stronger when a person can offer a real alternative.

A community group may identify a missing service but lack the staff to test a solution. A small university may have the relevant knowledge but lack computing resources. A public agency may know that a vendor’s offer is weak while having no independent capacity to judge another design.

In each case, participation can stop at a request addressed to someone else. The agenda remains with whoever can afford to build.

Public research capacity changes that relationship. The United States launched the National AI Research Resource pilot in January 2024 to broaden access to computing, data, models, software and support for U.S.-based researchers and educators.² That dated example shows one institutional route for sharing research resources. It is not evidence that everyone can now obtain what they need, or that a U.S. program solves access elsewhere.

The principle worth extending is that society can help more people ask and test questions. It can do more than purchase the answers a few large suppliers choose to offer.

I would connect public participation to a fund for small, well-defined research projects. A community could bring a problem to a university, technical college or public-interest laboratory. Together they could test whether an AI tool offers useful help. The people with the problem would help define success and examine the result.

This would require a real budget, trained staff and a route from a successful test to lasting service. A grant that pays for a demonstration but leaves no means to maintain it can create another abandoned promise.

The fund should also support projects that report failure clearly. A town that learns a proposed tool will not solve its problem has saved itself an expensive mistake. Publishing that result could save another town the same cost.

Such a program would give acceleration a wider base. More researchers could contribute. More public needs could become research questions. Builders could find uses that deserve investment because people have shown the demand, even where the first customers cannot pay much.

It would also let citizens challenge a company’s account with evidence of their own. Independent capacity makes a request for explanation harder to ignore.

I favor broad ownership of the gains from AI. The same logic applies to public knowledge. People should have a stronger position than that of an audience hoping for a generous decision from the stage. A financial stake, access to tools and a say in public priorities are distinct forms of power. None can fully replace the others.

A world larger than the leading labs

The phrase everyone becomes demanding when we take it beyond one country.

A decision made in a wealthy nation can shape the tools available elsewhere. A language with few paying customers may receive little attention. A research team may have excellent local knowledge and no affordable way to use it. People can be affected by a system without having any influence over the institution that built it.

A global discussion that invites a few officials from poorer countries is only a beginning. The people who will use, study and live with the technology need the means to contribute between conferences.

The Masakhane research community offers a concrete example. In a 2020 paper, Wilhelmina Nekoto and colleagues described participatory machine-translation research for African languages. The project produced benchmarks for more than thirty languages, with human evaluations for a subset, and enabled contributions from people without formal research training.³ This was a specific research effort, not proof that language inequality had been solved.

It does show why the distinction between experts and affected people should remain open to movement. Someone who begins as a speaker of a neglected language can become a contributor to the science. Participation can produce knowledge as well as preferences.

My proposal for international access starts there. Fund research partnerships in which local institutions help choose the questions and retain the ability to continue the work. Pay for translation, technical training and the maintenance of shared tools. Judge success partly by the capabilities left with the partner, rather than only by the papers or demonstrations produced during the grant.

A partnership is weak if the local team supplies data while every important choice, skill and resource remains elsewhere. It is stronger if the team can question the design, improve the system and decide how the result should serve its community.

Public access must still be matched to the kind of resource involved. Sharing educational tools is different from handing over private records or releasing capabilities with serious misuse potential. Carefully governed research access can expand participation without treating every form of unrestricted release as a public good.

But those distinctions should be stated and reviewable. A claim about security should identify the concern and the reason for the restriction. Otherwise the word can become a convenient way to preserve an advantage while calling the exclusion responsible.

The international bargain should include reciprocal learning. A project developed under unreliable connectivity or a tight maintenance budget may teach a wealthy institution how to build a more durable service. Treating the rest of the world only as a place to deliver products wastes that knowledge.

Progress belongs to more people when more people can help make it.

Enough voice to move

None of this requires every research project to wait for a world assembly. We should match the process to the decision. A small, reversible experiment needs a different form of oversight from a system that will shape access to essential services for years.

Set deadlines for deliberation. Fund the work needed to meet them. Make officials decide and publish their reasons. Keep routes for correction open as evidence arrives.

Indefinite consultation can be as dishonest as consultation after the contract. One prevents people from changing a decision already made. The other prevents anyone from being responsible for making a decision at all.

Broad pauses can reproduce that second failure. They can leave power with the institutions best placed to decide when everyone else may proceed. Calling that arrangement democratic does not make it so. Democracy requires an account of who decides, on what evidence, under which limits and with what possibility of challenge.

A public that has helped choose the purpose of a tool may demand faster development. It may ask for a change in design. It may reject a particular use. Those answers can all be part of a serious commitment to progress.

The horizon reaches far beyond a better meeting. We could build a society in which the person who notices a neglected problem has a path to people who can help solve it. A farmer could bring a question to a local research team. A disabled worker could help shape a tool around the independence they want. A student far from a great university could contribute to a project that once would have been closed to them.

Follow that change across a lifetime. The student gains skill and begins asking harder questions. The worker gains room to make choices and brings knowledge no designer possessed alone. A community that once waited for a distant supplier can develop some of its own answers. Its people become harder to disregard because they can do more than ask.

That is an increase in human power worth accelerating. It will demand machines capable of more, and institutions capable of trusting more people with the means to contribute. The people who have been treated as the last recipients of progress can become some of its makers.

I want us to build toward that future with purpose. The point of a seat before the decision is to give people a hand in making it arrive.

Sources and notes