The Limit of Extraction: Why Behavioral Data Is Not Equivalent to Expertise

In the spring of 2026, a major technology corporation installed on the work computers of its American employees a system of continuous logging: cursor movements, clicks, keystrokes, periodic screenshots of the screen across hundreds of applications — from the development environment to the corporate messenger. The stated goal is to collect training data for AI agents that are meant to perform everyday tasks at the computer. There is no way to opt out on a corporate device. European offices have been excluded from the program: European regulation does not recognize consent given within an employment relationship as freely given if the alternative is dismissal. In the United States, where federal regulation of employee monitoring is minimal and consent obtained at hiring is formally sufficient, the program is legally impeccable — and it is precisely this impeccability that makes the case analytically interesting.

The case is convenient to use as a point where several problems converge at once: what exactly is being taken from the worker, whether this “something” can be extracted at all in principle, and whether what is extracted retains its value after generalization. These are different questions, and conflating them is the main source of confusion in the public discussion.

What is being taken: process instead of result

The employment contract, in its classical form, distributes rights to the result: code, text, a drawing, a report. How the worker arrived at the result — the accumulated habit of solving problems, the sequence of trials and retreats — was historically considered a personal, portable skill. It is precisely this layer that becomes the object of the new collection systems: not the artifact, but the procedure of its production.

This logic has a long genealogy. Early-twentieth-century Taylorism attempted to extract the knowledge of the skilled worker and rewrite it as an instruction accessible to the unskilled one. Later this operation was described as the separation of conception and execution: conception concentrates in the hands of the owner of the means of production, while execution is devalued. Contemporary economists call the digital version of the same operation “theft of expertise,” while legal scholars document the migration of algorithmic management practices from the gig economy into ordinary employment. The difference from Taylorism, however, is fundamental: the time-and-motion man stood beside his stopwatch and saw a limited sample of actions; the logging system sees everything, continuously, and the result of the observation is not an instruction but a training dataset for a model.

The parallel with financial markets shows that the mechanism is not new — it has simply crossed an industry boundary. Brokers and exchanges have long used the flow of client orders for sentiment analytics, for forecasting the clustering of stop-orders, for training models. The legal foundation is the same: data arising on the company’s infrastructure in the course of performing a service is, by default, considered its asset. The question that remains open in both cases is where the performance of the service ends and the extraction of expertise as a separate commodity begins. The difference between the trader and the engineer lies not in the mechanism but in expectations: the market participant knows from the outset that they are being observed; the office worker, for a while, does not.

Channels of perception and the invisibility of understanding

What follows is the less obvious part of the problem — the cognitive part. Pedagogical practice has long recorded that people differ in the channel through which meaning enters the mind. For one person it is enough to see an indicator on the dashboard — the picture already contains the history of what happened. Another needs to hear a report: a sequence of events set out in words. A third will understand neither the sign nor the account until they have gone through the whole path with their own hands. These are not learning styles in the popular sense, nor different degrees of trust in information — these are different architectures of understanding.

And they have a direct consequence for logging systems: such a system is capable of registering only the motor channel. The moment of understanding for the visual type — a pause of a fraction of a second and one precise click — is indistinguishable in the logs from chance. Understanding that arises in the listener during an oral report leaves no behavioral trace at all in the relevant application. Only the trajectory of the “hands-on” worker — the long path of clicks through the innards of the system — looks substantive in the data, and it is precisely this trajectory that is most likely to be classified as inefficiency and optimized away.

Three people who have solved the same problem will produce three different click trajectories, leading to one result for three different reasons. They cannot be generalized: averaging the trajectories does not produce an averaged understanding, because understanding was absent from the data to begin with.

Averaging as destruction: the example of two cooks

The problem of averaging is conveniently shown with a kitchen example. The first cook works by the recipe: onion, then sauce from scratch, then boiling — linearly. The second works by the kitchen: he knows that ready-made sauce is already in the refrigerator and that the water takes seven minutes to heat, so he puts the water on first and cuts the onion in the window this creates. The second strategy is faster and requires more experience — but it occurs less often and matches the “textbook” sequence less well.

An algorithm trained on the averaging of both trajectories will end up with neither of them. Half the observations put the water on at the start, half in the middle; half cook the sauce, half take it ready-made. The statistical average of these actions is a sequence that no one followed and that does not work. Real models, of course, do not average literally: they select the most frequent pattern as the norm. But this does not solve the problem, it merely shifts it: the novice’s strategy becomes the norm, while the experienced cook’s technique — sauce from the refrigerator, the parallel use of windows of time — gets flagged as an anomaly. The trained agent will reproduce competence in its most standard, least valuable form.

The limitation can be formulated more strictly. Any meaningful action has three levels: what is being done (I am cutting an onion), for what purpose (I am cooking a dish according to a recipe), why precisely now (while the water is heating, my hands are free). Logging records the first level exhaustively. The second can be attempted to be reconstructed from context — the name of the task, the open document — with limited accuracy: one and the same instance of cutting an onion can serve different purposes given identical logs. The third level — the management of time and resources, the sense of the moment — is absent from behavioral traces in principle, because what carries it is not an action but a representation of the state of the system: the water is heating, this window needs to be filled.

And it is precisely the third level that distinguishes the expert. And it, too, is not the same across different people: one decides it’s time to cut the onion upon seeing the water start to boil; another, upon hearing the kettle’s noise; a third, upon feeling that their hands are idle. Even where the reason for an action exists, it arrives at different workers through different channels — and it passes through the log by none of them.

None of this implies that collecting behavioral data is useless: a model trained on thousands of trajectories will know where people usually click and will cope with routine procedures. What it does imply is something else: the value of what is extracted is systematically overestimated by those who extract it, since the visible layer of the action is taken for the action as a whole. This consideration also changes the question of fair compensation: if what is being taken is not the full expertise but its motor surface, then the worker’s claim, too, should concern precisely this layer — but ownership of it must still be a matter of agreement, not of default silence.

The minimal structure of such an agreement is fairly foreseeable: informational clarity (what is collected and where it is directed), structural freedom (the possibility of refusal without sanctions), managerial participation (workers’ control over the use of the data), compensatory symmetry (a share in the value that the company itself declares, as it commits tens of billions to training infrastructure). None of the elements is exotic: all four exist, in one form or another, in European regulation and in research-consent practices. Their absence in this case is not a technical necessity but the result of the fact that American labor law allows one to do without them.

The employment contract was designed for an exchange of “time and effort for remuneration, result for the employer’s property right.” It did not envisage a transaction whose object is the very manner of acting and judging. The attempt to carry out such a transaction through formal consent at hiring runs up against two independent limitations. The first is institutional: consent given under the threat of losing one’s job is hard to regard as consent — something European law recognizes and American law does not. The second is cognitive, and it is the stronger of the two: even the fully voluntary transfer of behavioral logs conveys not expertise but its motor imprint.

Understanding lives in purposes and in the choice of the moment, arrives through different channels for different people, and cannot be averaged without being destroyed. Extraction systems get exactly what they are capable of getting — the surface of the procedure — while the gap between this surface and what makes the worker competent remains technically unclosed and legally unresolved. It is precisely in this gap, and not in the fact of data collection as such, that the subject of future disputes lies.