The Yorktown Test for AI: Power Is Useful Only When People Can Challenge It

What if the most important question about artificial intelligence is not whether it works, but whether people can trust the society that deploys it?

That question reaches far beyond software. It leads to a humid night in 1781, when James Armistead, an enslaved man serving as a spy inside Cornwallis's headquarters, carried intelligence that helped expose the British plan for Yorktown. It leads to the French fleet waiting offshore, the trenches inching toward British defenses, and the fragile alliance that converted scattered advantages into victory. It leads, too, to the unpaid soldiers, indebted farmers, and political compromises that nearly broke the new republic after independence had supposedly been won.

These episodes illuminate a common problem: power is never the same thing as capacity. A tool, army, institution, or algorithm may be capable of producing an outcome. Whether it should be trusted, adopted, or obeyed depends on the surrounding system of incentives, accountability, and human judgment.

The real divide between people who adopt AI and people who abstain is therefore not simply enthusiasm versus fear. It is often a disagreement about governance. Who remains responsible when the system is wrong? Who benefits from its speed? Who bears the cost of its mistakes? And what happens to the people who cannot opt out?

The history of the American Revolution suggests a powerful answer: new power becomes legitimate only when it is embedded in relationships of trust, coordination, and restraint. Without those conditions, even a victory can create the crisis that follows it.

The Yorktown Lesson: Capability Does Not Win Alone

At first glance, Yorktown looks like a triumph of force. The allied army eventually outnumbered Cornwallis's troops two to one. French engineers helped construct trenches. Artillery battered British defenses. American and French forces captured the key redoubts in a daring nighttime assault.

But none of these facts explains the victory by itself. The decisive achievement was not merely possessing military strength. It was coordinating different forms of strength at exactly the right moment.

James Armistead supplied intelligence from inside the opposing command. Washington and Rochambeau had to decide where to concentrate their forces. The Comte de Grasse had to bring the French fleet into the Chesapeake and hold the bay against the British Navy. Washington had to conceal the march south while persuading Clinton that New York remained the central target. Engineers, sailors, spies, infantry, diplomats, and political leaders all contributed to the same outcome.

Remove any one of these links and the result becomes uncertain. An army without naval control could not permanently trap Cornwallis. A fleet without reliable intelligence might have arrived at the wrong place. Intelligence without commanders capable of acting on it would have remained an interesting secret. The victory emerged from integration, not from a single superior instrument.

This is the first useful model for thinking about artificial intelligence. AI is often presented as a force multiplier, and it can be. It can search more documents than a person, generate more alternatives, detect patterns across enormous datasets, and complete routine tasks at extraordinary speed. But a force multiplier also multiplies the quality of the system directing it.

If the objective is confused, AI accelerates confusion. If the incentives reward deception, AI scales deception. If no one knows who has authority to correct an error, AI turns a mistake into an administrative fact. The technology may be impressive while the institution using it remains strategically incompetent.

The question is not whether a system can produce an answer. The question is whether the surrounding human system can recognize when the answer should not be trusted.

This helps explain why adoption is uneven. One organization may adopt AI quickly because it has clear processes, good records, accountable managers, and a culture that permits employees to challenge automated outputs. Another may abstain because its data is unreliable, its responsibilities are ambiguous, or its errors would harm people who have no meaningful avenue of appeal. The second organization may not be backward. It may understand the deployment problem more clearly.

Adoption Is a Constitutional Choice

A constitution is not merely a statement of ideals. It is a way of deciding who may exercise power, within what limits, and under what conditions that power can be challenged. In this sense, every serious decision to adopt AI is a small constitutional decision.

Consider a hospital using an algorithm to prioritize patients, a bank using one to assess credit, or a school using one to identify students who need support. The software may be purchased as a technical product, but its operation changes the distribution of authority. A nurse may defer to a score rather than her own judgment. A loan officer may treat a recommendation as neutral even when the underlying data reflects old inequalities. A teacher may spend less time understanding a student because a dashboard has already classified the child.

The important question is not only, “Does the model improve average performance?” It is also:

1. Who gets to define success?
2. Who can inspect the reasoning?
3. Who can contest the result?
4. Who is accountable for the final decision?
5. Who pays when efficiency creates harm?

These questions resemble the dilemmas faced by the founders of the American republic. After the war, the United States had won independence but lacked a government capable of paying debts, regulating trade, maintaining an army, or resolving conflicts among states. The Articles of Confederation protected local autonomy so intensely that the collective government could barely act.

Shays' Rebellion exposed the cost of confusing freedom from central authority with freedom supported by effective institutions. Farmers who had fought for the country faced taxes, foreclosure, and imprisonment. The government that had asked for their service could not reliably pay them or address their grievances. Liberty had been declared, but the machinery required to sustain it was missing.

The lesson is not that stronger institutions are always better. The Constitutional Convention itself recognized the danger of concentrated power and built checks and balances into the new system. The deeper lesson is that freedom requires both restraint and capacity. A government too weak to protect rights leaves people vulnerable to private power. A government too strong to challenge turns protection into domination.

AI deployment faces the same tension. Organizations need enough central coordination to establish standards, security, training, and accountability. But they also need enough local judgment to notice when a general system fails a particular person. A rigid ban may protect against some harms while preventing useful experimentation. Unlimited adoption may produce innovation while quietly transferring authority from accountable people to opaque systems.

The mature alternative is neither automatic enthusiasm nor automatic refusal. It is conditional adoption: use the tool where its benefits are clear, its risks are observable, and its decisions remain contestable.

The Hidden Cost of Victory

Yorktown ended the last major battle of the Revolution, but it did not immediately resolve the war's human and political consequences. Enslaved people who had sought freedom behind British lines became vulnerable to recapture. Washington assisted in returning some fugitives to their owners. The peace settlement expanded American territory while ignoring Native American rights. The new nation proclaimed liberty while preserving slavery through constitutional compromise.

This is not an incidental contradiction. It reveals a general law of political and technological change: the winners of a system often define its success by the benefits they receive, while those who absorb the costs disappear from the official story.

A company may celebrate that AI reduced processing time by 40 percent. But what happened to the workers whose roles became more precarious? Did customers receive faster service, or merely encounter more difficulty reaching a human being? Did managers gain useful insight, or did they acquire a convenient way to avoid responsibility? Did the system reduce bias, or make old judgments appear objective because they now arrive in numerical form?

The American founders faced a similar temptation. They could preserve political union by postponing the slavery question. That compromise produced short term stability, but it did not eliminate the underlying conflict. It converted a moral and institutional crisis into a deferred obligation, one that later required a far more destructive reckoning.

AI systems can create the same kind of deferred crisis. An organization may introduce automation without deciding how displaced workers will be retrained, how incorrect decisions will be appealed, or how private data will be used. The deployment appears successful because the costs are assigned to people outside the room. Over time, resentment accumulates. The system may remain technically functional while losing legitimacy.

This suggests a practical test for responsible adoption: map the externalities before celebrating the output. Every new system has a visible benefit and a less visible redistribution of effort, risk, attention, or authority. The visible benefit might be speed. The hidden cost might be surveillance. The visible benefit might be convenience. The hidden cost might be the loss of skill. The visible benefit might be personalization. The hidden cost might be the narrowing of choice.

A useful institution treats these costs as part of the design, not as unfortunate side effects. It asks affected people what failure feels like from their position. It creates an appeal process before the first dispute occurs. It measures not only efficiency but also error concentration, reversibility, and dignity.

Trust Is Built Through the Right to Say No

The most revealing feature of the Revolution is that cooperation was not produced by total agreement. Americans, French allies, enslaved spies, soldiers, merchants, and local communities had different interests and different ideas about the future. Coordination became possible because particular people could contribute distinct forms of knowledge and because leaders could align those contributions toward a shared objective.

Trust did not mean assuming everyone was reliable. It meant creating conditions in which reliability could be tested and failure could be addressed.

James Armistead's intelligence mattered because someone had to evaluate it and act on it. The French fleet mattered because military plans depended on a commitment outside Washington's direct control. The march to Virginia succeeded because deception, logistics, and timing were connected. In each case, trust was operational. It was not a mood. It was a structure of dependencies.

AI changes the nature of these dependencies. A person may rely on a model without understanding how it reached its conclusion. A manager may rely on a vendor who cannot explain the training data. A public agency may rely on a system that no individual employee has the authority to suspend. The result is a dangerous form of borrowed confidence: everyone believes that someone else has checked the system.

The antidote is not complete transparency in the abstract. Many systems are too complex for every user to understand every internal mechanism. What matters is practical contestability. People need to know when the system is being used, what kind of decision it influences, what its limits are, and how to obtain human review.

The right to say no is especially important. Employees should be able to reject an automated recommendation when they have evidence that it does not fit the case. Customers should be able to challenge consequential decisions. Leaders should be willing to suspend a system without treating that act as failure. A tool that cannot be paused has become an authority.

Washington's conduct after the war offers a related principle. He had military prestige and immense personal influence, yet the survival of the republic required him to place power under civilian institutions and eventually relinquish command. The act was powerful precisely because it demonstrated that authority could be surrendered.

Organizations deploying AI need an equivalent precedent. The strongest signal of responsible adoption is not a promise that the system will never fail. It is evidence that leaders are willing to limit it, audit it, and shut it down when the public interest requires.

A trustworthy system is not one that demands confidence. It is one that makes correction possible.

A Practical Framework for Human Scale AI

The history of Yorktown and the early republic points toward a simple framework with four tests.

First, strategic fit. What problem is the system solving, and is automation actually the constraint? If the real problem is poor coordination, unclear authority, or inadequate resources, adding AI may merely conceal the weakness. Before purchasing a tool, identify the bottleneck in plain language.

Second, accountable ownership. Name the person or institution responsible for the outcome. “The model decided” is not an explanation and should never be an excuse. If no one has both authority and obligation to intervene, the system is not ready for consequential use.

Third, reversibility. Can the decision be reviewed, corrected, and undone? A recommendation about which article to read is easily reversible. A denial of housing, employment, medical care, or education is not. The more durable the consequence, the stronger the safeguards must be.

Fourth, coalition strength. Who must cooperate for the system to work? Include frontline workers, technical staff, affected communities, legal experts, and decision makers. Like the allied victory at Yorktown, success depends on connecting different forms of knowledge. A tool designed only by its vendor or executive sponsor will miss the intelligence held by people closest to the consequences.

These tests also clarify when abstention is wise. Abstaining is not a rejection of progress when the system lacks a clear objective, when errors cannot be appealed, or when the organization is using automation to avoid a difficult political decision. Sometimes refusing a tool is the first act of responsible governance.

Key Takeaways

1. Treat AI adoption as a governance decision, not merely a software purchase. Define authority, accountability, and appeal before measuring speed or savings.
2. Look for the system bottleneck. If the problem is poor data, unclear goals, or weak coordination, automation will amplify rather than solve it.
3. Require practical contestability. People affected by an AI assisted decision should know that the system was used and have a meaningful route to human review.
4. Measure hidden costs. Track who loses time, privacy, skill, income, or agency when the system succeeds according to its formal metric.
5. Make reversibility a design requirement. The more serious the consequence, the easier it must be to pause, inspect, correct, and undo the decision.

The sacred fire of liberty is often imagined as a flame that must be protected from government. The history of the Revolution suggests a more complicated image. Fire can illuminate, warm, or destroy. Its value depends on the vessel, the keeper, and the rules governing its use.

Artificial intelligence is another such fire. Its promise is real, but promise is not legitimacy. A system becomes worthy of trust when it serves a clear human purpose, operates within visible limits, and leaves people strong enough to challenge it.

The future will not be decided by those who adopt AI fastest, nor by those who abstain longest. It will be decided by whether societies can build institutions capable of using powerful tools without surrendering judgment to them. The central question is therefore not, “Will AI replace us?” It is more demanding: Can we become the kind of people and institutions that remain responsible when our tools become extraordinarily capable?