Sterile Knowledge: Artificial Intelligence, and the Genealogy of Understanding

In pediatrics there is an instrument that makes it possible to judge whether development is normal from a single measurement: the growth curve. Its epistemic power is provided by a combination of several conditions: a known window of development, an understood mechanism, a reference population of healthy children, an agreed threshold of deviation, and a system of regular monitoring. A single point on the chart is meaningful only because there exists a curve relative to which it can be interpreted.

No comparable instrument exists for cognitive development. Meanwhile, the mass adoption of generative artificial intelligence in education creates a situation that can be described as a large-scale experiment on the development of thinking, conducted without a reference standard, without a threshold, and without monitoring. The concern that arises in this connection has come to be called “cognitive lag”: if a child hands mental work over to a machine too early and too habitually, the capacity for independent reasoning may fail to form at all. It is essential that such harm is deferred in nature: it will manifest when the window of development has already closed, and it will be evident from its consequences but unprovable from its cause.

Posing the problem in this form is convincing, but its internal logic deserves critical examination. The question is not only whether we measure cognitive development, but also what exactly is placed at risk and, consequently, what ought to be measured in the first place. The present essay traces how, as this question is refined, the very object of anxiety shifts: from effort to the model, from the model to metacognition, and finally to the genealogy of knowledge.

The original argument rests on the concept of “desirable difficulties”: material whose acquisition requires effort is retained more durably. From this follows the conclusion that artificial intelligence, by removing difficulty, also removes the condition for durable learning. Empirically, this thesis is supported by observations that growing trust in AI is accompanied by a decline in critical thinking, and that writing with the help of a chatbot is accompanied by reduced activity in the brain systems associated with attention and memory. The concept of “cognitive debt” is used to describe this effect.

Here, however, a fundamental question arises: what mental content is actually formed through effort? If what is meant is the memorization of what already exists, the argument is vulnerable, since rote learning was never regarded as a path to understanding. If what is meant is the generation of something new, the mechanism connecting retrieval from memory with creativity remains unclear.

The clarification is that “desirable difficulties” are the opposite of mechanical repetition. Retrieving an answer on one’s own, constructing an example on one’s own, and returning to material at spaced intervals are not ways of accumulating facts but ways of building mental structures that make it possible to reason, to detect error, and to transfer knowledge to a new situation. Consequently, the danger of offloading lies not in the replacement of memory but in the replacement of the effort out of which the very capacity to think is built.

But even this clarification remains incomplete until it is determined which structures exactly are being built and which of them are actually vulnerable.

A considerable part of what is transmitted in school is not knowledge in the strict sense. It is either information (dates, values, names, facts, for example the very fact that black holes exist) or intellectual tools (methods of solving problems, algorithms, search techniques, for example solving quadratic equations). Knowledge as a picture of the world, that is, an understanding of how all this is connected, where it came from, and why it exists, stands behind information and tools but is not directly contained in them.

The thesis of “cognitive lag” usually treats thinking as a single faculty and therefore conflates these levels. Meanwhile, artificial intelligence acts on them differently. Information it replaces almost entirely, and this is probably not catastrophic, since information was always replaceable and was always stored on external media. Tools it is also capable of replacing, and here the real problem begins. A picture of the world it cannot supply, since such a picture is not transmitted in ready-made form but is built up.

However, as will be shown below, this three-level scheme also proves insufficient: between the tool and the picture of the world there lies yet another, decisive layer.

The strongest objection to the alarmist position is historical. The generation that studied successively with Bradis tables, the slide rule, the calculator, and the microcomputer repeatedly handed computational work over to external devices, and this did not prevent it from producing professional physicists. If delegating arithmetic proved harmless, why should delegating reasoning be any different?

A possible answer lies in the difference between execution and problem formulation. The calculator replaced the execution of an operation, but not the decision about what to calculate, why, and how to interpret the result. Moreover, it was operationally transparent: the user knew what a logarithm was even without calculating it from a table. A generative system is opaque, in that it presents an answer but conceals the path, and, more importantly, it takes on not only execution but also the choice of strategy.

At the same time, this objection cannot be considered refuted. For someone whose picture of the world has already formed, an opaque tool can function in the same way as a calculator. What remains open is the question of those in whom that picture has not yet formed. The assertion that the consequence will be a lag is no better proven than the assertion that there will be no consequences at all. It is precisely this uncertainty that constitutes the essence of the problem of the missing measurement.

An additional difficulty arises in connection with the thesis that the human ability to reason and to verify is necessary for the meaningful use of AI results. A counterargument can be advanced against it: for decades this ability did not lead to the solution of a number of mathematical problems, whereas machine systems obtain results that no one anticipated. On what basis will a person verify a solution that exceeds their own understanding?

Here it is appropriate to distinguish three cases. In formal systems, verification is as a rule simpler than search, since a proof admits mechanical verification. In the empirical sciences, verification is carried out through predictions, which is slow and not always possible but feasible in principle. Finally, there are domains where verification is as difficult as discovery; here a person can neither verify nor evaluate the result, and their attitude toward it comes down to trust or distrust.

Thus, the anxiety that children will “unlearn how to think” is supplemented by a more serious anxiety: perhaps society as a whole will lose the ability to understand what the machine is doing and to establish whether it is right. Measuring cognitive development in such a situation risks turning out to be the measurement of what we do not understand, with an instrument we do not have, against a standard that does not exist.

The question of what measurement itself serves deserves separate consideration. The history of mass intelligence testing, including tests of the Eysenck type, suggests that their social function often consisted not in identifying the gifted or the lagging, but in the selection and normalization of the middle, that is, in bringing people into conformity with the demands of mass production, the army, and administration. This logic long ago went beyond the bounds of psychometrics: the minimalism of interfaces, standardized furniture, and unified design are all designed for the average user.

In this light, artificial intelligence appears not as a neutral tool but as a continuation of the same tendency. Trained on a corpus of human texts, it reproduces what is statistically prevalent; by averaging and relying on established methods, it is structurally inclined to produce “mediocrities.” Hence the redefinition of the diagnosis: the principal risk consists not so much in lag (ending up below the norm) as in homogenization (ending up exactly at the norm). This is more dangerous because the intellectual development of a civilization is determined not by average values but by the “tails” of the distribution, by those who do not fit in.

From this likewise follows the duality of any future “cognitive growth chart.” The absence of the instrument is not a neutral gap but a choice not to know, one that is advantageous to those who conduct the experiment. But the presence of the instrument likewise does not guarantee its liberating function. Decisive are the questions of whose interests the measurement serves, to whom it is accountable, and what is considered a good result. Moreover, if the ability to go beyond the average is recognized as the value, then a standard of such an ability is in principle impossible: were it a standard, it would have become the average.

The three-level ontology proposed above contains an error: it assumes that the picture of the world is built directly through tools. In reality, between the tool and the picture of the world lies the model, a representation that specifies which entities exist, which mechanisms are admissible, and what is generally considered possible. The tool serves the model but does not search for a new one.

Instructive is the example of supermassive black holes in the early Universe. The established model regarded a black hole as the final stage of the evolution of a massive star; consequently, objects of such great mass should not exist in the early Universe, since for stellar origin and subsequent growth there is simply not enough time. Observations contradicted this expectation. The essential shift in explanation occurred not owing to a new computational tool, but owing to a revision of the model: the possibility was proposed that black hole seeds form through the direct collapse of a gas cloud, bypassing the stellar stage. This is not a computation but a redefinition of the domain of the possible.

No tool is capable of performing such an operation on its own: a telescope can show that the former model does not work, but it cannot suggest that the mechanism is different. A new model requires abandoning an assumption that previously had not even been formulated as an assumption.

With regard to artificial intelligence, this means the following. Learning from texts, the system acquires not only facts but the models themselves: accepted ways of explanation, admissible mechanisms, and legitimate moves of thought. It is capable of combining, interpolating, and optimizing within this space, but departure beyond its limits cannot be derived from the data. Consequently, homogenization extends not only to answers but to models: what is established is replicated, since it is precisely this that dominates the training corpus. The genuine cognitive lag in this perspective is not the inability to calculate without a calculator, but the inability to doubt a model, to see in it an assumption rather than nature.

One might suppose that AI is capable at least of serving as a generator of hypotheses within a given space. However, this formulation too overestimates its role. It would be more accurate to describe it as a collector of data, or at best as an implementation of Zwicky’s morphological “box”: the problem is decomposed into parameters, combinations are enumerated, and those that do not work are filtered out. Such enumeration can indeed accelerate the emergence of a new model, but only within the framework of the established paradigm.