The Hidden Model Behind Every Exercise of Power

What do an artificial intelligence system making a prediction and a president sending troops into a city have in common? At first glance, almost nothing. One is a problem in computation; the other is a problem in constitutional authority. Yet both reveal the same underlying danger: action does not follow facts automatically. It follows a model of what the facts mean, who may respond, and where the response must stop.

This is the question beneath both artificial intelligence and the legal limits on executive power: What invisible model converts information into permission?

A photograph can be altered by changing its pixels. A mind can be altered by changing the categories through which it interprets events. A government can be altered in the same way, not only by changing laws, but by changing the assumed relationship between danger, evidence, and authority.

The most important models are therefore not merely predictive. They are models of action. They tell us what counts as a signal, what counts as a threat, and what kinds of intervention are legitimate.

Facts Do Not Tell Institutions What to Do

Suppose a city experiences protests, property damage, clashes with police, and widespread public anxiety. Those are facts, or at least observations that can be recorded as facts. But the observations do not contain their own conclusion.

They do not tell us whether the right response is more local policing, a federal investigation, a curfew, a negotiation, or the deployment of the National Guard. They do not tell us whether the relevant problem is criminal conduct, civil disorder, political dissent, institutional failure, or a mixture of all four.

The transition from observation to action requires a model. That model contains assumptions about causality and jurisdiction. It answers questions such as:

• What is causing the disturbance?
• Which institution is responsible for addressing it?
• What evidence is sufficient to justify intervention?
• How reversible is the proposed response?
• What authority exists when ordinary procedures seem too slow?

This is why legal disputes about military or National Guard deployments are not merely technical arguments over paperwork. They are disputes over the model of the state itself. One model says that a sufficiently serious public danger creates a broad need for executive action. Another says that the existence of danger does not erase the boundaries separating local, state, and federal authority.

The Supreme Court's restriction on the administration's ability to deploy the National Guard illustrates this distinction. The ruling's logic appears to reach beyond one particular deployment and to narrow the legal routes available for sending troops into Democrat led cities, including Los Angeles, New Orleans, and Portland, Oregon. The significant point is not only that one action has been blocked. It is that the Court has challenged the conversion rule connecting perceived disorder to federal force.

The executive position, in its broadest form, treats urgency as a multiplier of authority. The judicial position insists that urgency cannot by itself manufacture authority. In other words, the Court is not simply evaluating the severity of the data. It is examining the algorithm that turns the data into a decision.

A government is defined not only by what it can see, but by what its model permits it to do with what it sees.

That insight also explains why artificial intelligence can be so intellectually powerful and so politically dangerous. An AI system does not merely retrieve information. It transforms information according to a structure that ranks patterns, predicts outcomes, and recommends responses. The structure may be hidden, but its consequences are not.

The Difference Between a Data Model and an Algorithmic Model

A data model tends to treat the past as a storehouse of relevant examples. It asks: What has happened before, and what patterns appear in the available records? This can be useful, but it has a weakness. Data is never just reality in miniature. It is reality after selection, measurement, classification, and institutional habit have already shaped it.

If a city records more arrests in one neighborhood, the data may appear to show greater criminality there. But it might also reflect more patrols, different reporting practices, historical targeting, or a population that has less ability to avoid police contact. The data is not false. It is incomplete in a systematic way.

An algorithmic model, by contrast, attempts to represent relationships that generate outcomes. It may ask which variables interact, which conditions matter, and how a change in one factor alters the result. In many fields, such models can produce better predictions because they do not simply count visible correlations. They try to infer the machinery beneath them.

This distinction can be applied to public authority. A crude data model of disorder might say: protests have become violent, public confidence is falling, and local officials appear unable to restore order. Therefore, federal troops should be sent.

A more sophisticated institutional model asks different questions. Is the violence concentrated or general? Are civilian institutions functioning, even imperfectly? Does the proposed intervention reduce danger or intensify it? Does the federal government possess a lawful role, or is it using the existence of a crisis to expand into a domain reserved for another authority? What precedent will the intervention create for the next president?

The second model is not necessarily more permissive or more restrictive. It is more causal. It recognizes that an intervention changes the environment in which future decisions are made.

Deploying troops may suppress an immediate confrontation while also weakening local legitimacy, increasing political polarization, and encouraging future executives to interpret ambiguous events as invitations to use force. A decision that looks effective in a narrow data model may be disastrous in a broader model of institutional feedback.

The same problem appears in personal reasoning. When a person sees a frightening headline, a data model of the mind collects more examples of danger. An algorithmic model asks what process produced the headline, what incentives shaped its wording, how representative the event is, and what action would actually reduce risk.

This is what it means to describe AI as a kind of Photoshop for the mind. Photoshop does not merely show an image. It makes certain edits easy, certain compositions natural, and certain distortions almost invisible. AI can perform the same operation on thought. It can sharpen a pattern, erase a complication, enlarge a threat, or make an interpretation feel inevitable.

The Real Risk Is Not Bad Information, But Bad Editing

Public discussion about AI often focuses on whether a statement is true or false. That matters, but it is not the deepest issue. The deeper issue is which layer of thought is being edited.

An AI system can change:

1. The facts we notice.
2. The categories into which we place those facts.
3. The causal story we tell about them.
4. The range of actions we consider reasonable.
5. The institution or person we believe should act.

The fourth and fifth layers are especially consequential. A system may present a recommendation without explicitly commanding anyone. Yet recommendations alter the perceived menu of possible actions. Once one option is described as efficient, necessary, or inevitable, alternatives begin to look sentimental or irresponsible.

This is why algorithmic assistance in government cannot be judged only by accuracy. A highly accurate prediction can still produce an unjust decision if the institution using it has the wrong mandate. Predicting where unrest is likely to occur does not answer whether a military force should be used there. Predicting which communities are most likely to experience violence does not answer whether surveillance should be expanded over them.

Prediction and legitimacy are separate dimensions.

A useful framework is to divide every high stakes decision into three layers:

The observation layer

What is happening? Which measurements are reliable? What is missing from the record? What alternative explanations fit the evidence?

The authority layer

Who has the legal, moral, or institutional right to respond? What limits govern that right? Which powers are unavailable even in an emergency?

The consequence layer

What will the response change, including its effects on trust, precedent, incentives, and future behavior?

Many failures occur because institutions jump from the first layer to the third while skipping the second. They observe a danger and calculate an effective response, but never ask whether the responder is authorized to act.

That is the central lesson of the National Guard dispute. The question is not simply whether troops might restore order. It is whether the federal executive can treat a perception of disorder as a sufficient bridge from observation to force. The Court's intervention, as described, says that the bridge has legal load limits.

This framework also gives us a way to use AI without surrendering judgment. Ask an AI system to analyze observations. Ask it to generate competing causal explanations. Ask it to identify missing variables and simulate second order effects. But do not let it silently decide the authority layer.

An AI may be able to estimate what will happen if a government deploys troops. It cannot, by prediction alone, establish that the government is entitled to deploy them.

Constitutional Limits Are a Form of Institutional Intelligence

Legal constraints are often described as obstacles to action. In emergencies, they can look frustratingly slow. But a constraint can also function like an intelligent model that prevents a system from overreacting to noisy inputs.

In machine learning, a model that fits every irregularity in its training data may perform badly when conditions change. It has mistaken noise for signal. Institutions can suffer from the same problem. A government that treats every episode of disorder as proof that extraordinary powers are required is overfitting to crisis.

Constitutional rules provide a kind of regularization. They prevent the state from adapting so aggressively to one alarming episode that it loses the general structure needed to govern fairly across many episodes. A president may sincerely believe that a city is failing. That belief does not automatically justify rewriting the division of authority for the entire country.

This perspective changes how we understand judicial decisions that limit executive action. They are not necessarily choosing passivity over effectiveness. They may be protecting the system from a form of institutional overfitting, in which a temporary emergency becomes the template for permanent power.

The principle applies beyond military deployments. A company may respond to one security incident by monitoring every employee. A school may respond to one act of misconduct by imposing rules that treat every student as a threat. A person may experience one betrayal and build an entire philosophy of suspicion around it.

In each case, the mistake is not noticing the danger. The mistake is allowing the danger to rewrite the model.

Good judgment does not merely detect an emergency. It preserves the ability to distinguish an emergency from the next emergency.

This is also where AI can either improve or degrade human reasoning. Used well, it can expose hidden assumptions and generate counterexamples. Used badly, it can make the first interpretation feel more authoritative by producing it fluently and quickly.

Speed increases the need for institutional friction. When a model can generate ten plausible justifications in seconds, the scarce resource is no longer explanation. It is disciplined refusal: the capacity to pause before a compelling story becomes a mandate.

A Practical Discipline for Decisions Under Pressure

The most useful response is not to reject models, AI, or executive action. It is to make the conversion from evidence to action visible.

Before accepting a high stakes recommendation, write down four sentences:

• The signal: The specific observation that appears to require action.
• The model: The causal explanation connecting that observation to the proposed response.
• The authority: The person or institution entitled to act, and the rule that grants that authority.
• The cost: The likely second order effects, including precedent and loss of trust.

This exercise often reveals that a decision contains several hidden leaps. For example: a protest is violent in one location; therefore the city is unable to govern; therefore the federal government must intervene; therefore troops are an appropriate instrument; therefore the intervention will calm rather than escalate the situation. Each step may be debatable even if the first observation is indisputable.

The same method works with AI generated advice. If a system recommends rejecting a job candidate, ask what signal it used, what model connects the signal to performance, what authority defines the relevant standard, and what costs arise from false positives. If an assistant recommends a business strategy, ask which assumptions would have to be true for the recommendation to work and what evidence would disconfirm them.

The goal is not to eliminate interpretation. That is impossible. The goal is to prevent interpretation from disguising itself as fact.

Key Takeaways

• Separate observation from permission. Evidence can show that something is happening, but it cannot by itself establish who may respond or what tools are legitimate.
• Ask what model connects the facts to the action. Look for hidden assumptions about causality, jurisdiction, reversibility, and precedent.
• Use AI to challenge your frame, not merely decorate it. Request alternative explanations, missing variables, counterexamples, and second order consequences.
• Treat legal limits as safeguards against overfitting. A temporary crisis should not automatically become a permanent theory of executive power.
• Make the authority layer explicit. Before acting on a powerful recommendation, identify the rule, mandate, or responsibility that authorizes the proposed action.

The most dangerous sentence in public life is often not “this is false.” It is “given these facts, there is only one thing we can do.” That sentence hides the model. It suppresses the alternatives. It converts a judgment into an apparent necessity.

The future of both democratic government and artificial intelligence will depend on resisting that conversion. We need systems capable of finding patterns, but also institutions capable of questioning the patterns they find. We need leaders who can respond to danger without allowing danger to enlarge their authority by implication. We need citizens who understand that a polished explanation may be an edited image of reality, with inconvenient features quietly removed.

The deepest form of intelligence is therefore not prediction. It is knowing when prediction has crossed into permission. A free society survives by keeping those two things separate.