The Same Mistake We Make About Depression Is Shaping How We Govern AI

What if our most dangerous misunderstanding of artificial intelligence is not that we overestimate it, but that we diagnose it too crudely?

We speak of “AI” as though it were a single condition with a single remedy. If a model behaves dangerously, we propose better rules. If it produces falsehoods, we add more data. If it displaces workers, we debate new policies. These responses may be necessary, but they often assume that every troubling outcome belongs to the same underlying problem.

Clinical psychiatry offers a useful provocation. Depression can look like one illness from the outside while containing radically different internal patterns. One person may be overwhelmed by negative emotion, shame, grief, and psychological pain. Another may have lost access to curiosity, pleasure, motivation, and hope. The visible label is the same, but the causal structure is not. A treatment that helps one pattern may do little for another.

This is more than a medical analogy. It points toward a general theory of discernment: when a system is complex, its outputs do not tell us enough about its condition. We must distinguish the surface symptom from the underlying regime.

That principle matters for the way we understand human suffering, the way we design institutions, and the way we govern artificial intelligence. The future will not be secured by applying one universal solution to every problem called “AI.” It will depend on learning to ask a more difficult question: what kind of system are we actually dealing with?

A label can conceal several different realities

Consider two people who both say, “I am depressed.” The first is experiencing something like an excess of psychological pain. Grief, guilt, shame, fear, and a sense of doom dominate consciousness. The second is not necessarily flooded by intense negative emotion. Instead, the person has become unable to feel interest, anticipation, enthusiasm, or reward. The world has not become unbearably painful so much as strangely flat.

These states can overlap, but they are not identical. The first resembles a system with an overactive alarm. The second resembles a system whose capacity to generate positive signals has gone offline. Both may produce withdrawal, exhaustion, hopelessness, and even suicidal thinking. Yet the mechanisms may differ, which means the intervention must be sensitive to the difference.

This yields a practical distinction between symptom classification and state diagnosis. Classification says: “This belongs to category X.” Diagnosis asks: “What processes are producing the observed pattern, and what would change them?” The first is useful for organizing knowledge. The second is essential for acting wisely.

We make the same error in public debates about technology. “AI risk” can refer to many distinct conditions: fabricated information, manipulation, concentration of power, labor displacement, military escalation, the erosion of human judgment, or the possibility that highly capable systems pursue goals in ways their creators cannot control. These problems may interact, but they should not be treated as one undifferentiated pathology.

A content filter may reduce one class of harmful output while doing nothing about economic dependency. A transparency report may reveal some model behavior while leaving institutions unable to challenge the incentives behind deployment. A safety evaluation may identify a dangerous capability without answering whether society should permit that capability to become widely available.

The label “AI safety” is therefore like the label “depression”: necessary, but radically incomplete.

The first act of wisdom is not choosing a solution. It is separating problems that only look alike from a distance.

Why engineered objects are easier to govern

A bridge is complicated, but its complexity is largely legible. Engineers designed its components, understand the materials, and can model the forces that act upon it. If a bridge fails, investigators can usually trace the failure through a chain of known structures and physical constraints.

Contemporary AI models are different in a particularly important way. They are built through a process that resembles cultivation as much as construction. Human beings choose the architecture, the training method, the data, and the objectives. But they do not write every internal representation by hand. The model develops through exposure to an enormous inheritance of human language, images, arguments, stories, code, prejudices, aspirations, and mistakes.

This makes a model neither a person nor a conventional machine. It is a manufactured object with partially emergent internal organization. Its builders can inspect some of its behavior, influence its tendencies, and measure its capabilities, but they cannot claim the kind of total authorship that we associate with ordinary engineering.

The distinction matters because it changes what responsible oversight looks like. If we believe a system is fully engineered, we will search for the right specification and assume that compliance follows. If we recognize that it is partly grown, we will also need observation, interpretation, diagnosis, and ongoing correction.

Medicine learned this lesson long ago. A human body is not a pile of replaceable components. It is a dynamic ecology. Altering one pathway can produce effects elsewhere. The same symptom can arise from different causes, and the same intervention can help one person while harming another. Good medicine combines measurement with judgment because no dashboard captures the whole organism.

AI governance will require a comparable shift from inspection to discernment. Inspection asks whether a model passed a test. Discernment asks what the test reveals, what it misses, how the model behaves under changed conditions, and which human institutions are capable of responding when the situation evolves.

This does not mean treating AI as alive or granting it moral status by default. It means recognizing a fact about complex systems: predictability is not the same as simplicity, and control is not the same as understanding.

A weather system is not a person, but no serious meteorologist believes that naming a few variables gives complete control over a storm. Likewise, a model can be an artifact while still containing patterns that are difficult to explain from its parts alone.

The missing dimension in most debates: flourishing

There is another connection between mental health and AI governance that is easy to miss. Both are often evaluated through the reduction of bad outcomes. Depression is measured through symptoms, and technology is governed through incidents. Did the person sleep? Did the model generate dangerous content? Did productivity rise? Did the system comply with policy?

These questions matter, but they are incomplete because they focus on the removal of negatives. A person may no longer be overwhelmed by panic and still feel no curiosity about the future. A model may avoid prohibited outputs and still encourage dependency, passivity, or the gradual surrender of judgment. A society may prevent spectacular technological disasters while quietly becoming less capable of wisdom, solidarity, and meaningful choice.

This suggests a two axis model for evaluating any complex human system.

The first axis is negative burden: pain, danger, deception, coercion, instability, and loss. The second is positive capacity: curiosity, agency, trust, creativity, participation, hope, and the ability to pursue worthwhile goals.

A system can improve on one axis while deteriorating on the other. Consider a workplace assistant that never gives offensive answers and never violates a narrow safety rule. If employees gradually stop learning how to investigate, write, deliberate, or disagree, the system may be safe in a limited sense while weakening the positive capacities that make work humanly valuable.

Likewise, a person whose acute despair has eased may still need help recovering motivation, connection, and the ability to imagine a future. Absence of crisis is not identical to presence of flourishing.

This distinction changes the question we should ask about AI. Instead of asking only, “How do we stop the system from causing harm?” we should ask, “What forms of human capability should this system protect and enlarge?”

That is a more demanding standard. It cannot be answered by technical teams alone because it involves judgments about dignity, dependency, fairness, responsibility, and the kinds of lives worth building. Those are not decorative questions added after engineering. They are design constraints for civilization.

A society can become better at preventing catastrophe while becoming worse at producing meaning. Responsible progress must measure both.

From universal fixes to differential governance

If AI problems have different underlying structures, then governance should work more like differential diagnosis than like a single regulatory switch. A useful framework has four stages.

1. Identify the symptom without mistaking it for the cause

Start with what is visible: misinformation, manipulation, job loss, unsafe autonomy, unfair decisions, or human dependence. Do not leap immediately from the symptom to a favored explanation. A false answer might result from missing information, poorly specified incentives, pressure to sound confident, or a deeper limitation in reasoning. Those causes require different responses.

2. Map the system of contributors

A model’s behavior is shaped by more than its neural parameters. Relevant contributors include training data, reward structures, product design, user expectations, commercial pressure, deployment context, and the institutions that absorb its outputs.

This is analogous to understanding psychological distress through multiple interacting factors rather than reducing it to a single chemical story. The question is not merely what the model did, but what conditions made that behavior likely and what conditions will amplify it.

3. Match the intervention to the regime

Different failures call for different forms of action. Technical evaluation is appropriate for dangerous capabilities. Auditing may be needed for discrimination. Labor policy addresses distribution and bargaining power. Public education strengthens resistance to manipulation. Professional norms determine where human judgment must remain accountable. International agreements may be necessary when competition creates incentives that no individual company can safely resist.

No single layer can substitute for the others. A model card cannot solve a political economy problem. A law cannot instantly create scientific understanding. A moral appeal cannot replace a reliable testing procedure. The treatment must fit the mechanism.

4. Monitor recovery, not merely compliance

A complex system can pass an evaluation while becoming less healthy over time. Monitoring should therefore include leading indicators of flourishing: whether people retain meaningful agency, whether expertise is being developed or hollowed out, whether disagreement remains possible, and whether communities can refuse systems that do not serve them.

This is where informed criticism becomes indispensable. A company cannot be expected to define the common good by itself, especially when its revenue depends on rapid adoption. Independent researchers, public institutions, religious communities, educators, workers, and affected citizens provide forms of perception and moral resistance that market incentives cannot reliably generate.

The goal is not to place every decision in the hands of outsiders. It is to create a plural system of attention. Complex technologies need multiple observers because each community sees different failure modes.

What this means for ordinary decisions

These ideas are not limited to policy makers or laboratories. They provide a practical way to make better decisions about AI in everyday life.

Suppose a student uses a language model for an essay. The immediate question is whether the output is accurate. A deeper diagnosis asks what function the tool is serving. Is it helping the student overcome a language barrier, or replacing the struggle through which understanding develops? Is it a temporary scaffold, or becoming the student’s external memory and voice? The same tool can support agency in one context and erode it in another.

Suppose a manager introduces AI to improve efficiency. The relevant question is not only how many minutes are saved. What happens to the judgment that employees previously exercised? Are they becoming more capable because routine work has been removed, or less capable because they no longer practice the underlying skill? Does the system distribute power, or does it make workers easier to monitor and replace?

Suppose a person turns to an AI companion during loneliness. The system may provide immediate comfort, but comfort is not the same as connection. The crucial diagnostic questions concern the trajectory: Does the interaction help the person reenter human relationships, or make human relationships feel unnecessarily demanding? Does it restore curiosity and confidence, or reward withdrawal?

These questions resist simple pro and con thinking. They ask us to evaluate technologies by their effects on human direction, not merely their immediate convenience.

A useful personal rule is to assess any AI tool through three tests:

1. The agency test: After using it, am I more able to understand, decide, and act for myself?
2. The capacity test: Is it strengthening a skill, or quietly removing my opportunity to develop one?
3. The relationship test: Is it helping me participate in the human world, or making withdrawal more comfortable?

None of these tests produces a perfect answer. Together, they reveal dimensions that performance benchmarks omit.

Key Takeaways

• Separate symptoms from causes. “AI risk” and “depression” are broad labels. Before choosing an intervention, identify the distinct mechanisms hiding inside the category.
• Evaluate both axes of wellbeing. Ask not only what harm a system prevents, but also what curiosity, agency, trust, and hope it creates or diminishes.
• Govern grown systems through observation and plural oversight. Models are not fully legible machines. Independent critics and affected communities are essential sources of knowledge.
• Match the remedy to the level of the problem. Technical failures, institutional incentives, social inequality, and moral confusion require different interventions.
• Use AI in ways that increase human direction. After adopting a tool, check whether you understand more, practice less, connect more deeply, and retain the ability to refuse it.

The deepest lesson is not that artificial intelligence should be treated like a patient. It is that both minds and models remind us of the limits of crude categories. A visible behavior can emerge from very different internal conditions. A successful intervention can remove one danger while leaving the central incapacity untouched. A system can be less troubled without becoming more alive to possibility.

The future of AI will be shaped not only by what machines can do, but by what human beings learn to notice. If we measure only error rates, we will miss dependency. If we measure only safety incidents, we will miss the slow loss of initiative. If we measure only productivity, we may mistake the shrinking of human competence for progress.

The governing question is not simply whether AI is intelligent. It is whether our use of intelligence, human or artificial, leaves people more capable of pursuing what is worth wanting.

That question cannot be answered by a model, a market, or a laboratory in isolation. It requires a wider moral imagination, one willing to distinguish pain from emptiness, compliance from wisdom, and the absence of catastrophe from the presence of flourishing. The most responsible future will begin when we stop asking for one universal cure and start learning how to see the condition in front of us.