Whose labour is the machine absorbing? A call for labour process theory in HCI

The question most often repeated in the argument about artificial intelligence is this: "will AI take my job?"

A paper posted to arXiv in February and accepted for CHI 2026 contends that this is the wrong question.

The right one is more uncomfortable:

Even if artificial intelligence does not take my job, what is it turning my job into?

The field's blind spot#

The authors argue that when human-computer interaction (HCI) studies workplaces with artificial intelligence in them, it works within a fixed frame: usability, efficiency, trust, acceptance.

Within that frame, artificial intelligence is a tool. The questions are put accordingly. Does the user trust it? How quickly does it get the work done? Is the interface intelligible?

For the authors, this frame has a systematic blind spot. It does not ask who owns the tool, whose service it works in, or who holds control of the work process.

Their proposal is straightforward: bring labour process theory into HCI.

The theory's pedigree#

Labour process theory was built by three thinkers in succession.

Marx. In the first volume of Capital comes the observation that what is bought from the worker is not labour but labour power. The employer buys a capacity for a given stretch of time; how much actual labour will be got out of that capacity is another matter entirely. From this the question of control over the production process arises. The machine is one of the instruments of that control: what resides in the worker's knowledge is transferred into the structure of the machine.

Harry Braverman. Labor and Monopoly Capital, published in 1974, applied that observation to the twentieth-century workplace and established the thesis of deskilling. The function of Taylorism is to move the knowledge in the worker's head into the hands of management: conception is separated from execution, and the one who executes no longer knows why.

Michael Burawoy. With Manufacturing Consent in 1979 he complicated the picture. Control is not established by coercion alone; the workers' consent is manufactured. Game-like elements, targets and small zones of autonomy are built into the job, and the worker takes an active part in his own exploitation.

Carry that trio over into the argument about artificial intelligence and the questions that emerge stop being technical.

Which theory is artificial intelligence an example of?#

The reading the paper proposes shows today's artificial intelligence systems running all three mechanisms at once.

Transfer of knowledge. A language model is trained on text written by millions of people. A code model, on the human labour deposited in open-source repositories. A translation model, on the bilingual text professional translators produced over years. Braverman's observation that the knowledge in the worker's head is transferred into the machine is here technically true — the knowledge really has been transferred.

Control. Algorithmic management hands the pace of work, the choice of route and the measurement of performance over to a piece of software. This is a form of supervision standing in for the foreman; the difference is that it is continuous, quantified and closed to negotiation.

Consent. Artificial-intelligence tools are for the most part presented as assistants. And they genuinely do assist: they speed the work up, take on the tedious part, and give the user a feeling of power. It is exactly here that Burawoy's point takes hold — a system's being genuinely useful does not stop it from also being an instrument of control.

Deskilling, or reskilling?#

This is the oldest argument within the theory, and the point at which the paper too leaves matters open.

The classic objection to Braverman's thesis runs like this: technology does not only destroy skills, it creates new ones. With automation, some jobs disappear and others are born.

In the case of artificial intelligence, both are observed. A software engineer's practice of writing code is changing: he writes fewer lines and spends more time inspecting and verifying them. Is that a loss of skill, or a displacement of it?

The criterion the theory offers lifts the answer out of ambiguity: the question is not whether the work has become easier, but who holds control of it.

Can a software engineer refuse the model? Does he have time to assess its output? Or does the expectation of speed reduce inspection to a formality?

Why it matters for Turkey#

The argument is not an abstract one. Algorithmic dispatching in courier and transport apps, automated performance measurement in call centres, the spread of AI-assisted processes in translation and text production — all are instances of the same structure.

And none of them fits under the headline "artificial intelligence is taking our jobs". The jobs are still there. What has changed is the share of the decision left inside the work.

Reading Marx in 2026#

The paper is directly connected to our other piece today, on Karl Marx. The way Marx is read in the twenty-first century is moving steadily away from "theory of revolution" and towards critical analysis of work and technology.

As we noted in our piece this week on Chantal Mouffe, one wing of contemporary left thought puts conflict at the centre of politics. Labour process theory looks for the same conflict at the point of production.

The assumption the two share is this: the appearance of agreement does not mean that conflict has been resolved; it may mean only that it has become invisible.

Conclusion#

The paper's call looks modest: a theory proposed to a field.

Its consequences are not modest. To assess an artificial-intelligence system in terms of "user experience" and to assess it in terms of "whose labour process is controlled by whom" are two different research programmes.

And the two programmes, looking at the same interface, see different things.