AI generation is getting cheaper and cheaper, and what has truly become more expensive is verification.

Every significant technological revolution is initially prone to being misinterpreted as a victory of machine capabilities. The steam engine was seen as a triumph of power, electricity as a triumph of energy, and computers as a triumph of computing speed. In the era of artificial intelligence, the popular view is that machines can now write, draw, program, answer questions, and even think. These statements are not entirely wrong, but they lack an economic perspective.

Economics doesn't merely inquire about what technology can do. Its real question is: What costs does technology change, and how are the relative prices in the entire system rearranged as a result? Once relative prices change, the division of labor, organizational boundaries, value capture, and property - rights structure all change accordingly.

From this perspective, the most significant change in the AI era is the significant drop in the marginal cost of generation, rather than machines learning to generate content per se. In the past, only an author, a lawyer, a research team, or a consulting group could produce candidate texts, solutions, codes, and hypotheses. Now, machines can generate them in large quantities at extremely low costs. Generation is no longer scarce—at least, the low - level, preliminary, and enumerable forms of generation are rapidly losing their scarcity.

Scarcity has thus shifted. As the number of candidate solutions increases, what truly becomes scarce is the ability to judge whether a solution can be implemented and whether it's worth the risk: whether a contract hides risks, whether code has loopholes, whether a drug is safe and effective in the human body, whether a research design is truly exogenous, and who should bear the consequences of a strategic misjudgment. The bottleneck in the market system is shifting from insufficient generation to insufficient verification.

This judgment requires a qualification. The verification discussed here is not all types of verification. Verifications such as whether code can be compiled, whether accounts balance, whether dosages exceed limits, and whether a contract triggers a specific written rule are all formalizable. AI can also reduce the cost of these verifications, and they are precisely the areas where structured intelligence is most suitable. What is truly scarce is another type of verification: it cannot be fully formalized, and someone must bear the consequences. Whether a research hypothesis truly captures the causal mechanism, whether a business bet is worth the resources, who should bear the risk of an unwritten contract, and whether a policy recommendation can overcome the complex frictions in local implementation—these judgments cannot be fully defined by pre - established rules. They are scarce because the state space of the world is incomplete, contracts are incomplete, and consequences cannot be outsourced, regardless of how difficult they are to calculate.

Here, we need to address a possible counter - argument. Some may say that "non - formalizability" is only temporary. Ten years ago, tasks such as radiology image reading, autonomous driving road - condition judgment, and contract due diligence were considered to require human endorsement. Now, algorithms are taking over large parts of them. If the boundary of formalization is constantly retreating, then "non - formalizability" is just a synonym for "not yet conquered" and cannot be a stable source of scarcity.

This counter - argument is valid, but it actually points to the real anchor. The root of scarcity lies not in the technical attribute of "difficult to formalize" (the technological frontier is indeed advancing continuously) but in the property - rights attribute of "untransferable consequences". Even if algorithms completely take over road - condition judgment, once an accident occurs, the liability for compensation must fall on a subject capable of bearing it. No matter how high the prediction accuracy is, the question of "who is responsible" cannot be eliminated. What can be formalized is prediction, and what cannot be transferred is consequences. The scarcity of verification mentioned in this article is precisely anchored to the latter. As for why consequences are untransferable, we will answer it in the third section using incomplete - contract theory.

This scarcity is not a temporary phenomenon due to technological backwardness but a deep - seated structure revealed after the rearrangement of the price system in the AI era. AI can significantly reduce the costs of search, recombination, and expression, but it cannot reduce the costs associated with the physical body, institutions, responsibilities, and time itself. It can help us generate candidate answers more quickly, but it cannot replace the real - world screening of these answers.

Large Probabilistic Models Are Not Low - Level Intelligence

To discuss this issue, we first need to avoid two extreme views on AI technology routes.

One is the theory of the omnipotence of large models: it seems that as long as models continue to grow, corpora continue to expand, and context windows continue to lengthen, all cognitive, commercial, scientific, and organizational problems will eventually be absorbed by a single set of probabilistic prediction systems. The other is the theory of belittling large models: large probabilistic models are simplified as "predicting the next token", regarded as accumulations of statistical correlations and language - imitation machines, without a real world model, causal structure, or reliable reasoning. According to this view, true intelligence can only come from another route—world models, symbolic systems, structural reasoning, or some deeper causal machines. Both views are one - sided.

Large probabilistic models will indeed produce hallucinations, make mistakes in the long - tail cases, and mistake fluent language for factual judgments. They also cannot replace experiments, clinical trials, audits, compliance checks, causal identification, and organizational responsibility - taking. However, these flaws precisely indicate that the world in which they operate is an open and uncertain one, and they do not prove that they are low - level intelligence.

Human society is not a closed - boundary test paper. There is no complete answer to corporate strategies, consumer preferences are not stable, policy environments are full of games, the spread of ideas depends on context, and in scientific exploration, we often do not know the correct form of the problem. Many important judgments face not a well - defined problem but an incompletely unfolded state space. In these situations, the most crucial ability is to organize possibilities under incomplete information, rather than calculating a closed - ended question with zero error.

This is where the value of large probabilistic models lies. They are far more than just repeating facts: they compress the existing knowledge state of humanity in a high - dimensional semantic space, recombine the relationships scattered in texts, and quickly form candidate explanations. They extract the potential connections hidden in literature, long texts, codes, cases, institutional texts, and daily language and present them in front of new problems. They may not provide the final answer, but they greatly expand the set of answers we can consider.

Using the language of the philosophy of science, this point will be clearer. Probability does not have only a frequency - based interpretation. In Keynes' 1921 A Treatise on Probability, probability is understood as a logical relationship between propositions: given a set of evidence, to what extent a conclusion is worthy of reasonable belief. It describes the logical form of reasoning by a rational subject when evidence is incomplete, rather than the statistics of event frequencies. Keynes especially emphasized that many probabilities cannot be represented by numbers or even compared with each other. In the face of true uncertainty, what we can do is not to calculate an exact posterior but to arrange the relative positions of various possibilities as consistently as possible within the scope allowed by the existing evidence. This tradition of logical interpretation, through Jeffreys and Cox, was formulated by E. T. Jaynes as "probability as extended logic": probability is the only consistent extension of logic in a world of incomplete information, and the principle of maximum entropy tells us that under given constraints, the rational prior should be the one that is the least arbitrary and spreads the possibilities the most widely.

This logical, or cognitive, view of probability precisely describes the working mode and comparative advantage of large probabilistic models. In an open world with an incomplete state space, rational operation does not mean approaching the error to zero (which belongs to the closed - world, frequency - interpretation, and structural - estimation scenarios) but organizing possibilities as consistently and fully as possible under the constraints of known information. Large probabilistic models present "the most worthy semantic extensions under the state of human knowledge" in the conditional distribution of a large corpus, which is essentially an operation of "spreading possibilities under constraints". Their core virtue is therefore to generate a well - structured candidate set in open uncertainty, rather than pursuing zero errors.

It is worth noting that Keynes, who held this view of probability, is also the Keynes who placed "fundamental uncertainty" at the center of economics. His famous statement in 1937, "We simply do not know", refers precisely to this world with an incomplete state space that cannot be exhaustively described in advance. Logical probability theory and Keynesian fundamental uncertainty are two sides of the same philosophical stance, and the open world described by this stance is precisely where large probabilistic models have a comparative advantage. Belittling large probabilistic models actually means misunderstanding the proper form of intelligence in an open world.

The other technology route is also important. Structured intelligence, represented by world models, symbolic reasoning, rule - constrained systems, and specialized small models, pursues another ability: achieving a low error rate and high controllability in an environment with clear boundaries, clear goals, and stable feedback. Whether code can run, whether accounts balance, whether processes comply with regulations, whether production parameters are abnormal, whether drug candidates meet established constraints—these tasks require convergence, verification, execution, and risk control, rather than open - ended generation. If we use economics as an analogy, structured intelligence is more like structural model estimation: it aims to find more stable constraints, mechanisms, and parameters in a well - defined task space, not being satisfied with the shadow of correlation. Its virtues lie in low variance, verifiability, auditability, and repeatability.

Let's imagine two situations. The world faced by a settlement clerk behind the counter is closed, with only right and wrong answers, allowing no ambiguity. He needs a structured verification that never makes mistakes. However, the world faced by a CEO is never a black - and - white test paper: there is no standard answer for strategies, preferences are not stable, policies are full of games, and many judgments point to an incompletely unfolded state space. He needs to organize possibilities under incomplete information, which is exactly the form of large probabilistic models. It should be noted that the difference in status between a CEO and a clerk can be misleading: more often than not, they are not two different types of people but two successive steps in the same process, as we will see in the next section.

Large probabilistic models and structured intelligence therefore face different uncertainty structures. There is no superiority or inferiority between them, nor a simple substitution relationship. The former is suitable for the open world, generating and expanding possibilities in chaos; the latter is suitable for closed tasks, converging and ensuring execution within boundaries. The real question is how the two can form a new division of labor in the economic system, rather than which route will ultimately win.

Four Stages and the Row - Wise Impact of AI

Many discussions on AI applications like to divide tasks into two categories: open - ended tasks are assigned to large probabilistic models, and closed - ended tasks are assigned to structured systems. This classification is clear in analysis but does not match the real operation of enterprises. In reality, what is more common is the vertical relay within the same task, rather than the horizontal separation of two types of tasks.

Let's imagine the entire economy as a matrix. Its columns represent \(N\) types of tasks in economic activities—drafting a contract, writing a piece of code, researching and developing a new drug, making a strategic judgment, and completing a paper. Its rows represent four stages that almost every task has to go through: Generation (proposing candidate texts, solutions, codes, and hypotheses), Verification (judging whether these candidates are valid), Execution (implementing the approved solutions in the real world), and Endorsement and Responsibility - Taking (holding the residual claim and bearing the consequences for the final result). Any task is a column in this matrix, passing through these four stages from top to bottom.

AI reduces the cost of an entire row in this matrix—the generation row—across all columns, making it close to zero. I call this the Row - Wise Impact of AI: the impact falls on an entire row, rather than on a few columns. The collapse of this row will then affect the formalizable part of the verification row (the verification of whether code can be compiled and whether accounts balance will also become cheaper), but it cannot touch the non - formalizable core of verification, let alone execution and endorsement.

The standard description of automation in economics—the task framework of Acemoglu and Autor—is presented column - by - column. It views production as a collection of a series of tasks and asks which tasks will be transferred from humans to machines. In the language of this matrix, it compresses each task into a single row, so automation becomes a Column - Wise Impact: the impact selects certain columns and takes the entire column away from humans. The popular question "Which jobs will be replaced?" is a popular version of this scenario.

After splitting tasks vertically into four rows of generation, verification, execution, and endorsement, the direction of the impact changes. AI no longer selects certain columns but reduces the cost of an entire row across all columns—this is a row - wise impact. The column - wise impact deletes tasks, while the row - wise impact rearranges tasks. The former asks which columns will disappear, while the latter asks how the relative prices of the four stages within each column will be re - adjusted. After the collapse of generation costs, the constraint shifts from generation to downstream verification, execution, and endorsement, so resources, manpower, and rents flow from the generation stage to downstream stages within the task. I believe that this re - allocation within tasks is more common and more profound than the entire - column substitution between tasks.

The weight of the generation stage varies in different tasks—in some tasks, it is the main cost, while in others, it only accounts for a small part. Therefore, the same row - wise impact will cause different degrees of rearrangement in different columns. The higher the weight of generation in a task, the more drastic the change in internal relative prices; tasks where generation is not a bottleneck will be relatively less affected. AI does not neatly divide the world into "replaced" and "not replaced" parts. It rearranges the internal structure of each task with different intensities.

Completing a column in the matrix from top to bottom is a relay: generation in an open state space passes the baton to verification, execution, and endorsement that face accountable consequences. I call this vertical chain the Generation - Verification Relay, and I use "verification" to collectively refer to the stages after generation that someone must be responsible for. The difference between this and an ordinary assembly line lies in the transformation of the uncertainty structure, not just the sequence of processes: the first part faces an open state space and is responsible for spreading possibilities; the second part faces untransferable consequences and is responsible for convergence, verification, and endorsement.

This framework helps us understand the real - world business organizations in the AI era. AI will not just lead to enterprises hiring fewer people. A deeper change is that the production function within enterprises is being re - arranged. In the past, many middle - level managers were responsible for information transfer, aggregation, translation, and format conversion: senior management put forward vague strategies, middle - level managers decomposed them into reports and tasks, lower - level employees executed them and then reported back step by step, and middle - level managers translated the results for senior management. As a result, a large number of middle - level links responsible for translation functions accumulated in the organization. Large probabilistic models reduce the cost of this part—they can directly understand complex contexts, generate task decompositions, draft materials, integrate information, and propose options; structured systems can also complete a large number of standardized verifications and process controls in the background. Therefore, the part of the traditional middle - level management with low information content will be squeezed.

However, the middle - level management will not disappear. The real middle - level managers who remain will transform from information monopolists to handlers of real - world frictions: they need to deal with long - tail exceptions that are difficult for AI to cover, local knowledge, internal organizational trust, cross - departmental coordination, and the fulfillment of commitments. Enterprises will shift from a multi - level information - transfer structure to a flatter system centered around generation - verification - real - world coordination.

There is an interesting isomorphism here. The bottleneck in the idea market is shifting from insufficient explanation to insufficient verification, and the bottleneck in corporate organizations is also shifting from insufficient execution to insufficient endorsement: candidate solutions are no longer scarce, and the stages that can endorse solutions and bear the consequences are scarce. Functions such as risk control, compliance, testing, clinical trials, auditing, signing, and regulatory interfaces, which previously seemed to be back - end functions, will become the key nodes in the value chain. This is exactly what price theory has repeatedly reminded us—when the cost of one stage decreases, the relative value of the complementary stage will increase: the cheaper the generation, the more expensive the verification; the more powerful the tools, the scarcer the responsibility.

The Shift of the Coasean Boundary

Coase once posed a simple yet profound question: Since the market can coordinate resources, why do we still need enterprises? The answer lies in transaction costs. Market transactions require searching, negotiating, monitoring, and enforcing contracts. Enterprises internalize some transactions because internal coordination is cheaper in some cases. What AI changes are precisely these transaction costs.

When the costs of semantic retrieval, information recombination, text and code generation, solution generation, and preliminary analysis significantly decrease, many cognitive tasks will be pushed to the market. Enterprises can purchase AI services and outsource a large number of low - consequence, standardized, and replaceable generation stages. Many preliminary tasks that previously required an in - house team can be replaced by external tools, platforms, and model services.

However, not all tasks will be pushed to the market. Some tasks actually need to be internalized even more. This time, the reason lies in consequence attribution and the enforceability of contracts, rather than information - processing capabilities. Any stage that requires bearing consequences, difficult - to - articulate coordination, embedding in a specific institutional environment, internal organizational trust, and long - term commitments is difficult to simply outsource: if a strategic judgment fails, the model service provider cannot bear the enterprise's losses; if a clinical trial fails, the generator of candidate molecules cannot bear all the consequences; if a compliance judgment is wrong, the automatic summarization system cannot take on the legal liability.

Therefore, the statement "AI assists in judgment but