We Are Asking the Wrong Question
When people talk about artificial intelligence, one of the most common questions is: Will AI replace programmers? What about doctors, lawyers, teachers, journalists, accountants—or me?
The question may be wrong from the start. “Programmer,” “lawyer,” and “doctor” are job titles. Every occupation is actually a bundle of different capabilities. A programmer does more than write code: the job also involves understanding requirements, designing systems, making trade-offs, detecting anomalies, and collaborating with others. A lawyer does more than search cases and draft contracts: the work also involves understanding facts, judging risk, developing strategy, and negotiating with clients and counterparties.
AI does not first enter an occupation as a whole. It enters some of the capabilities that make up that occupation. So before asking whom AI will replace, we should ask a more basic question:
What are human capabilities made of? Which capabilities is AI entering? Which are becoming cheaper, and which are becoming scarcer as a result?
That question is a better starting point for understanding work, software, firms, and individual competitiveness in the AI era.
1. Human capability has three layers
Functionally, human cognitive and practical capability can be analyzed in at least three layers.
Here, “capability” does not mean a degree, a professional credential, or a stock of knowledge. It means the actual ability of an actor to perform a particular cognitive or practical function.
The first layer is basic capability.
Basic capabilities can be invoked relatively independently within a well-defined task, and they usually allow local feedback and verification. Examples include retrieving information, recognizing images, calling up knowledge, summarizing text, translating, writing, generating code, comparing documents, or completing a clearly specified software operation. “Basic” does not mean easy. Writing a complex piece of code can be difficult. The point is that the task can be bounded and its output can be checked locally.
The second layer is integrative capability.
This is not about performing a single local operation. It is about continuously organizing multiple capabilities around a goal. It includes judgment, planning, coordination, organization, monitoring, exception handling, and continuing to pursue an objective as the environment changes.
Writing a piece of code and taking responsibility for a large software system are not the same kind of capability. Reading a medical image and managing a patient over time are not the same either. The former can be verified locally; the latter requires continuous integration of information, objectives, risks, and changing real-world conditions.
The third layer is meta-capability.
Meta-capability does not ask only, “How do I do this?” It asks: What method should I use? Which judgment should I doubt? What evidence is still insufficient? When should I change strategy? Why was I wrong this time? What rule should I change next time?
Meta-capability manages other capabilities. In the simplest terms:
Basic capability is the ability to do. Integrative capability is the ability to combine capabilities in pursuit of a goal. Meta-capability is the ability to manage how one does so.
This is not a fixed biological taxonomy. The same activity may sit at different layers in different task environments. Planning a three-step code change can be close to a basic capability; planning a company’s strategy for the next three years is clearly integrative.
Making a complex judgment mainly uses integrative capability. Asking whether that judgment itself is reliable, whether the method is appropriate, or whether the yardstick should be changed moves into meta-capability.
The point is not to assign every capability a permanent label. The point is to build a useful yardstick for analyzing AI applications.
2. AI is externalizing and lowering the cost of some basic capabilities
One of the most important changes brought by generative AI is that cognitive capabilities once performed only by the individual are increasingly becoming externally callable resources. In the past, if a person wanted a capability, the person usually had to learn it. If you could not program, you learned programming. If you could not translate, you learned a language. If you could not conduct legal research, you received professional training. If you could not design, you learned specialized software.
Now the situation is different. A person who is not an experienced programmer can call on substantial programming capability. A person who cannot read one hundred articles in a day can ask AI to summarize, compare, and organize them first. A person whose own knowledge covers only a limited set of fields can call on a knowledge space far beyond personal experience.
This does not mean expertise has lost its value, nor does it mean AI can already perform all of these tasks reliably. The important change is more specific:
Some cognitive capabilities that once had to be internal to the individual are becoming externalized.
Once a capability can be called at scale and at low cost, its economic character changes. It becomes cheaper. The first structural change, therefore, is not simply that “machines are smarter than people.” It is that:
Some cognitive capabilities are moving from scarce human resources toward inexpensive external resources.
Here, “becoming cheaper” means that the marginal cost of accessing and invoking the capability is falling. It does not mean that high-quality outcomes automatically become cheap, and it certainly does not mean verification costs disappear.
3. Internal capability and effective capability are beginning to diverge
This creates a distinction that was previously less visible. The capabilities a person possesses unaided can be called internal capabilities.
But what a person can actually accomplish also depends on which external capabilities can be invoked reliably. In the AI era, effective capability can no longer be defined simply by what a person can do alone. More precisely:
A person’s effective capability is increasingly determined by both internal capability and the external AI capability that can be invoked reliably.
The word “reliably” matters. Getting AI to complete a task once does not mean the capability has effectively been acquired. Reliable invocation at least requires knowing how to specify the task, understand the result, detect obvious errors and boundaries, verify critical conclusions, and switch methods, escalate, or stop when necessary. Someone who knows nothing about code and happens to obtain a working script from AI does not possess the same effective capability as someone who can use AI to design, test, modify, and maintain software consistently.
This is also why two people using the same model can produce very different results. The AI capability may be the same, but their integrative and meta-capabilities differ; therefore the capability of the resulting human-AI system differs as well.
4. AI expands human cognitive reach
Human knowledge is limited, but so is the range of knowledge and associations a person can actually activate at any given moment. Even someone who has read widely must search for answers through a limited field of experience, memory, and association. Some people can think for a long time and gradually widen that field, but the number of knowledge nodes, the density of associations, and the speed of retrieval available to one individual remain strongly constrained.
Large models provide a very different kind of external cognitive system. They do not contain “all human knowledge,” and they lack large amounts of lived experience, tacit knowledge, and up-to-the-minute information. But the range of explicit human knowledge and associations they can call on is far larger than that available to any single individual. One major consequence of human-AI collaboration is therefore:
Candidate solutions that once lay outside a person’s cognitive reach can be brought into that person’s field of judgment.
This does not mean AI should make the final judgment. Often the opposite is true. A person confronting a complex problem may not know in advance where to go next, yet may immediately recognize which direction is worth pursuing once AI lays out several possibilities. High-quality collaboration therefore often looks like this:
AI expands the candidate space; humans compress the choice space.
AI is strong at expansion. Human advantage still often lies in goals, lived reality, judgments of importance, anomaly detection, and final trade-offs. “AI expands, humans compress” is not merely a slogan; it describes an observable division of cognitive labor.
What expands here is the reachable cognitive space. It does not guarantee that final thinking will become more diverse. If AI is used only to produce a single “best answer,” the human and the model may converge too early together.
5. Value migrates to the next scarce bottleneck
As many basic capabilities become cheaper, scarcity may move upward. A more durable principle than “integrative capability will rise in value” is this:
When AI lowers the cost of a capability, value tends to migrate toward the next scarce bottleneck.
Today, search, summarization, writing, code generation, and information organization are becoming cheaper quickly. The bottleneck therefore shifts toward questions such as: Which problem is worth solving? Which variable actually matters? Which answer deserves trust? When is the evidence still insufficient? When should direction change? How should multiple capabilities be organized into a system that continues to function over time? At the current stage, many of these bottlenecks sit in integrative and meta-capabilities.
But this is not permanent human territory. In bounded tasks with clear feedback, AI is already entering integrative capability and even parts of meta-capability.
Nor is this an eternal rule about any one layer. If AI eventually makes these capabilities mature and cheap as well, scarcity will move again. The durable principle is not that one capability will remain valuable forever, but that:
Value keeps migrating toward the next scarce bottleneck.
6. Human + AI does not automatically create greater capability
Putting a human and an AI together does not automatically create an effective human-AI system. AI can mislead the human; the human can interfere with good AI output; both can develop shared blind spots. A third variable therefore matters: workflow.
Through sustained research practice, we have gradually developed a relatively stable process for open, uncertain problems that require judgment:
Define the object and the problem boundary → establish a factual base → discuss and broadly expand the space with AI → form a judgment → conduct prior-art search and adversarial challenge → verify facts, yardsticks, and causal links → revise the judgment → compress it into a coherent form → deposit reusable rules.
This is not a fixed checklist that every task must follow in full. The more deterministic and verifiable the task, the more the process can be simplified. The more open the problem and the higher the consequences, the more important full expansion, challenge, and verification become.
After the object and factual base are established, AI can use its broad knowledge radius to expand the space of possibilities, after which the human forms a judgment.
Once a judgment has formed, the next step is not to keep finding support. It is to conduct prior-art search and adversarial challenge: actively look for intellectual predecessors, opposing views, counterexamples, and facts capable of breaking the conclusion.
The process verifies not only the facts but also the causal arrow between fact and conclusion. If the judgment fails the attack, revise it. If an error reveals a more general problem, deposit the lesson as a reusable rule for future work.
Under conditions in which AI participates, the relationship can be summarized with a heuristic structural formula. It expresses mutual constraint among the factors; it is not a directly measurable mathematical identity:
Effective human cognitive productivity ≈ callable AI basic capability × human integrative capability × collaboration workflow quality
Strong AI with weak human judgment can still produce poor results. Strong human judgment without AI remains constrained by personal cognitive and execution reach. And even when both are strong, poor division of labor, weak verification, or bad feedback can make them interfere with one another.
What matters is not the abstract phrase “human + AI,” but:
Which capabilities are assigned to whom, and by what rules those capabilities are connected.
7. Long-term collaboration creates a shared cognitive protocol
Workflow explains how a single task is coordinated. Long-term collaboration creates a second benefit: rules begin to accumulate. Human-AI collaboration does not have to restart from zero every time.
When the same rules are used repeatedly, humans gradually compress them into intuition. An unusually large number, for example, may immediately feel “wrong.” The person may not consciously replay the entire chain of reasoning, but repeated experience has formed an implicit judgment rule.
Sustained human-AI dialogue can do something that was previously difficult: make part of those implicit judgment rules explicit again.
Why did something feel wrong? Which variable triggered the warning? Was the error in the fact, the classification, the yardstick, or the reasoning chain?
Through repeated discussion, a one-time experience can become an expressible, storable, reusable rule that can also be handed to others. Human and AI can gradually develop a shared cognitive protocol.
What is shared is not the requirement to reach the same conclusion. It is a set of rules about when to doubt, when to search for counterexamples, what evidence is acceptable, and when the problem itself must be redefined.
The ideal state is not increasing agreement between human and AI. It is:
Align the rules; keep the conclusions open to attack.
Otherwise, long-term collaboration can itself create shared blind spots. Mature human-AI collaboration therefore needs both knowledge accumulation and adversarial challenge.
8. Analyze capabilities first, then discuss “job replacement”
If capability is the more basic unit of analysis, then “Will AI replace programmers?” is not yet a good enough question.
A better sequence is:
AI capability changes → some human capabilities become cheaper → tasks inside occupations are recombined → individual productivity changes → firms redesign jobs → market demand changes → only then do employment levels rise or fall.
Therefore:
Capability substitution is not the same as task substitution; task substitution is not the same as job substitution; job change is not the same as the disappearance of an occupation.
A programmer may hand much coding work to AI and, as a result, manage a larger system. A lawyer may hand much research and drafting to AI while judgment, negotiation, and responsibility remain. A doctor may hand parts of information organization and local recognition to AI while longitudinal patient management still requires a more complex bundle of capabilities.
Some occupations may indeed shrink sharply, and some may disappear. But to understand those outcomes, we have to decompose capabilities first and then trace the economic transmission. Starting with the occupational label hides the change that is actually taking place.
The transmission is not a one-way line either. Job design, market demand, wages, and education choices feed back into the price and supply of capabilities. Occupational outcomes are ultimately part of a feedback system.
9. AI changes the structure of capability, not just its level
The most important question in the AI era is not simply whether people are becoming “more capable.” Some internal basic capabilities may decline as they are outsourced over long periods. People may memorize less, perform fewer calculations themselves, or write less code directly. At the same time, the external capabilities they can invoke may increase rapidly. Several changes can therefore occur at once:
Some internal capabilities decline; external capability rises sharply; effective capability rises when collaboration is reliable; attention and value are reallocated toward newly scarce capabilities.
This reconfiguration has costs. Basic capabilities that are no longer practiced may deteriorate, and that deterioration can reduce sensitivity to detail in later judgment. “Having more external capability” does not mean every internal human capability becomes stronger.
This is a restructuring of capability, not a simple increase in capability. In the past, we often evaluated people by how much they knew, how much they could do, and how many professional skills they possessed.
Increasingly, we may need to ask different questions: What problems can they formulate? What can they judge? Do they know when not to trust an answer? Which external capabilities can they invoke reliably? How many different capabilities can they organize into a result? Can they turn error into a better rule for next time?
What AI changes, therefore, may not simply be that machines are beginning to possess human capabilities. The deeper change is:
Human capability itself is shifting from a concept enclosed within the individual toward an open system composed of the person and external intelligence.
That may be the real starting point for understanding who will be displaced, who will benefit, and which capabilities will become most valuable in the AI era.
We should not begin by asking, “Whom will AI replace?” We should begin by asking: What is human capability made of? Which capabilities are being externalized and made cheaper by AI? When those capabilities are no longer scarce, what becomes the next scarce capability?
Only after we build this yardstick can we begin to see clearly how occupations, firms, and society will change.
