The Human in the Loop

A reader left a comment on "Forty Percent of the Code, a Quarter of the Staff". They said they were curious about how the concept of the human in the loop relates to the argument of that piece.

We said we should write a piece to clarify it.

---

First, the Summary: Ten Theses

1. "The human in the loop" is not a design choice. It is a labour relation. The term is presented as a technical arrangement in which the human completes what the machine cannot do. In fact it describes a division of labour, a wage relation and a distribution of responsibility.
2. The loop has three stops, and a different form of labour works at each. Before the model is trained, the data worker; while the model is running, the worker of hidden intervention; after the output, the worker of approval and supervision.
3. Artificial intelligence is the automation of the division of labour. As Matteo Pasquinelli shows, machine intelligence copies the patterns of human labour once it has been broken into parts. The Babbage principle is at work again: the job is split into pieces, and each piece is paid only for the skill it requires.
4. The automation fetish is an instrument of valuation. Looking "fully automatic" raises a company's market value. Hiding human labour makes the product seem more magical than it is, and the company more valuable than it is.
5. A global division of data labour has been built. The heaviest and cheapest jobs, traumatic content moderation and mass labelling, are subcontracted from Kenya to the Philippines, from Venezuela to Türkiye. Value accumulates in the centre while the cost is paid at the periphery.
6. The composition of labour is shifting: from the cheap generalist to the expensive specialist. The rising tendency of the last two years is to hire doctors, lawyers, engineers and programmers by the hour so that they transfer their own expertise into the model. Skilled labour is training the machine that will take its place.
7. The piece-wage has returned. The form of payment Marx called the one best suited to the capitalist mode of production has become the basic wage form in data work. And it is now combined with algorithmic supervision.
8. The human who supervises is the human who carries the responsibility. The assurance that "a human makes the final decision" can turn into a mechanism that charges the cost of the error to the worker. Madeleine Clare Elish calls this a "moral crumple zone".
9. The loop is steadily narrowing. In companies' use of artificial intelligence there is a marked shift from use that complements the human to use that replaces the human. The human in the loop should also be read as the human waiting to be removed from the loop.
10. Organising has begun. The global union alliance of content moderators, the Data Labelers Association in Kenya, and the struggle of Telus workers in İzmir inside Çağrı-İş show that invisible labour is becoming visible, and that labour which is visible has begun to organise.

Let us now open the theses one by one.

---

The Man inside the Mechanical Turk

In 1770 the Hungarian inventor Wolfgang von Kempelen built a chess automaton that toured the courts of Europe. A mannequin dressed in Turkish costume sat at a table and beat its opponents. It is said to have played Napoleon and Benjamin Franklin. The machine was presented as a miracle. Inside the table, though, a chess master was hidden, moving the pieces with magnets and levers.

In 2005 Amazon gave exactly that automaton's name to the platform through which it distributed small jobs, at piece rates, to people around the world: Mechanical Turk. In the company's own words, this was "artificial artificial intelligence": having humans do the jobs the computer could not do, in such a way that the computer appeared to be doing them.

The name is not an accident. The present structure of the artificial intelligence industry is built on the art of hiding the chess master inside that table. The difference is this: inside the table there is no longer a single master, but hundreds of thousands of workers scattered across the world.

---

What Is the Loop, and Where Is the Human?

In engineering, "human in the loop" is used for designs that leave room for human intervention at particular points in an automatic system. On paper it is a reasonable principle: the machine can err, the human corrects. But the technical plainness of the term makes the labour relation behind it invisible. Let us look at the three stops of the loop.

First Stop: Before the Model, the Data Worker

A language model or an image-recognition system is trained on millions of examples. A large share of these examples is prepared, classified and corrected by humans:

• Data labelling: Marking the pedestrian, the traffic light, the tumour in a photograph. Deciding whether a sentence is positive or negative.
• Reinforcement learning from human feedback (RLHF): Choosing which of two answers the model produced is better, scoring answers, writing sample answers. Behind the "polite", "helpful" and "safe" appearance of today's chatbots lies this labour.
• Content moderation and harmful-content filtering: For the model not to produce text containing violence, abuse and hate, that kind of content first has to be read and labelled by humans.
• Red teaming: Deliberately pressing the model to find its weaknesses, and testing scenarios of misuse.

The investigation Billy Perrigo of TIME published in January 2023 showed, sharply, what this stop means. To train ChatGPT's harmful-content filter, OpenAI employed workers through the subcontracting company Sama in Kenya. At under two dollars an hour, the workers read and labelled texts describing sexual abuse, violence and torture. Many of them said they suffered serious psychological harm.

Second Stop: While the Model Runs, the Worker of Hidden Intervention

Some systems are sold as "automatic", but they are supported by humans in real time. When the model is not sure, the job is quietly routed to a person. The user does not know. We also discussed this old secret of the history of science, the habit of hiding the labour behind the laboratory, in "Mythos, Model 2 and the Secret Framework".

The most discussed example is Amazon's cashierless store system, "Just Walk Out". According to a 2024 report by The Information, the system relied on about a thousand people working in India who watched and verified the shopping. Amazon denied the story, saying these workers were only marking data in order to improve the model. But the dispute itself showed something: for many services sold under the label "artificial intelligence", it cannot be known from the outside how much is machine and how much is human. The sector even has a name for this: "fake automation".

Third Stop: After the Model, the Worker of Approval and Supervision

The model produces an output; a human reviews it, approves it or corrects it. The radiologist reading a medical image with the support of artificial intelligence, the lawyer checking a contract the model wrote, the call-centre worker supervising automatic replies, the programmer reviewing code that artificial intelligence produced.

Two transformations happen together here. First, the content of the job changes: supervising instead of producing. The programmer reads code instead of writing it; the translator corrects machine translation instead of translating. This work is usually valued at a lower wage, because "the machine does the real job". The Microsoft piece that prompted this is exactly that: agents now open more than 40 percent of pull requests on GitHub, and the programmer is turning into a supervisor who reviews what the agents wrote.

Second, the distribution of responsibility changes. In "Moral Crumple Zones", published in 2019, Madeleine Clare Elish showed that when something goes wrong in automatic systems, the blame is laid not on the company that designed the system but on the human in the loop. Like the crumple zone in a car, the human in the loop becomes the zone that absorbs the force of the impact. When a driverless car crashes, the safety driver at the wheel is prosecuted; when an automatic diagnosis is wrong, the physician who approved it is put on trial. This is both a transfer of legal responsibility and a form of the intensification of labour: the worker has to watch the machine and be responsible for the machine's errors. We took up why the claim that companies can assume this responsibility by auditing themselves is empty in "Can Companies Audit Themselves?".

---

Theoretical Frame: Artificial Intelligence Is the Automation of the Division of Labour

The theoretical frame that ties these three stops together is Marx's analysis of machinery and large-scale industry.

In On the Economy of Machinery and Manufactures, published in 1832, Charles Babbage set out a principle that would later take his name: when a job is divided into parts according to different levels of skill, the employer pays, for each part, only for the skill that part requires. Instead of having a master do the whole job, it is far more profitable to have the master do the hard part and a cheap worker the easy part. Babbage applied the same logic to mental labour: he split the calculation of mathematical tables into parts and distributed them to cheap "computers", the human calculators of the time. This division of labour was also the basis of his design for a calculating machine.

In The Eye of the Master: A Social History of Artificial Intelligence, published in 2023, Matteo Pasquinelli reads the history of artificial intelligence along this line. For him, artificial intelligence is born not from the imitation of the human mind but from the copying of the patterns of the social division of labour. The machine first extracts a model of human labour that has been broken into parts, standardised and measured, and then takes the place of that labour.

In the famous "Fragment on Machines" in the Grundrisse, Marx wrote that the knowledge and skill accumulated by society, the "general intellect", is objectified in fixed capital, in the machine. We discussed this earlier as the expropriation of the general intellect. The concept of the human in the loop shows the living mechanism of that expropriation: the general intellect does not flow into the machine by itself. It is transferred into the machine by the daily, piece-rated, measured labour of millions of workers. Every label, every score, every correction is a piece of the labour crystallising in the machine. We discussed the form this same expropriation takes in scientific production, through OpenAI's 722 mathematics papers, in "The Closed Laboratory of the General Intellect", and the roots of these questions in philosophy in "Old Questions, New Laboratory".

There is a contradiction here. This labour does not own the value it produces. The knowledge transferred into the model becomes the company's intellectual property. The worker transfers their own skill to the machine, and as a result of that transfer their own skill is devalued. Marx's analysis of alienation appears here in its plainest form: the worker's product rises up against the worker as an alien power that directs them and takes their place.

---

The Ideological Function of Invisibility

It is necessary to ask why the human in the loop is hidden. Three reasons stand out.

Valuation. The market value of artificial intelligence companies rests on the assumption that their dependence on human labour is slight. A software company is valued at a high multiple because it can write a product once and sell it without limit. If there are thousands of workers behind the product who have to be paid continuously, the company is not a software company but a service company, and it is valued at a far lower multiple. We discussed this logic of valuation when writing on the AI bubble; hiding the human in the loop is one of the instruments by which the bubble is inflated. We had already seen how the same logic of valuation is written into a prospectus, in the piece on Anthropic's IPO, and how it is sold on stage, in "What Are the Artificial Intelligence Summits Selling?" and in the piece on Jack Clark's speech at the IMF.

Fetishism. Marx's analysis of commodity fetishism takes a new form here. The commodity hides the social relations of labour that produce it, and appears as an object that carries value by itself. Artificial intelligence likewise hides the labour that produces it, and appears as a subject that thinks by itself. The sentences "artificial intelligence wrote it" and "artificial intelligence decided" actually mean "a system trained on the labour of thousands of workers produced this, under rules set by a company".

Wage pressure. The invisibility of the job creates the impression that it is worthless. Descriptions such as "simple click work", "side income" and "flexible work" belittle the attention, the knowledge and the psychological load this labour requires. Invisible labour is easily turned into cheap labour.

---

The Global Division of Data Labour

The labour of artificial intelligence forms a new layer of capitalism's international division of labour. Companies in the centre countries subcontract the heaviest and cheapest jobs to the peripheral countries.

Data Workers' Inquiry, a joint project of the DAIR Institute and the Technical University of Berlin, inspired by the "Workers' Inquiry" Marx prepared in 1880, published reports in which data workers wrote their own experiences in their own hand. Content moderators in Kenya described working to a quota of more than one item a minute and about five hundred items a day, and suffering psychological collapse in the face of images of violence and abuse. A worker labelling data in Syria to cover the costs of education described vague task definitions and demanding clients. Layered subcontracting made it impossible to find out where the responsibility lay. The summary in TechCrunch conveys the project's method and findings well.

Türkiye is part of this division of labour too. At Telus Digital's facility in İzmir, workers doing content moderation for TikTok organised in the Çağrı-İş union. According to UNI Europa's statement of March 2025, while collective bargaining was under way the company first dismissed fifteen, then about thirty workers who supported the union. The union defined this directly as an attack on its existence as a union, and held a protest in front of the company in İzmir.

This picture is not an accident. It is a system. The export of capital, and the exploitation of the cheap labour of the peripheral countries, which Lenin described in his analysis of imperialism, is being reproduced today in data work. The difference is that what is exploited is no longer the mine or the agricultural product, but the human mind and human attention. We continue to follow the weekly trace of this division of labour in the series "The Week's Class Developments in Technology".

---

The Shift in the Composition of Labour: From the Cheap Generalist to the Expensive Specialist

An important transformation has been under way in data work over the last two years. Because the models have improved enough on basic tasks, companies now need expert knowledge far more than mass, cheap labelling. For the model to develop in fields such as law, medicine, engineering, mathematics and software, the knowledge of the specialists in those fields has to be transferred into the model.

This need gave birth to a new kind of company. Mercor, founded in 2023, reached a valuation of ten billion dollars, according to a Business Insider report of November 2025, manages more than thirty thousand contract workers, and pays out more than one and a half million dollars a day. The company is reported to work with six of the technology giants known as the "Magnificent Seven". Its rivals Scale AI and Surge AI work on a similar model; Scale AI speaks of more than seven hundred thousand contributors. In 2025 Meta invested 14.3 billion dollars in Scale AI.

At first glance this sector gives the impression that data work has become "respectable". Hourly wages are high, and the workers are educated. But the same report shows that this new section of workers is also in the grip of insecurity. Mercor suddenly ended the "Musen" project it was running for Meta, on which more than five thousand people were working at its peak, although the project had been said to run until the end of the year, and offered the workers a similar project at five dollars less an hour, at sixteen dollars an hour. One worker said that an earlier project had likewise been replaced by a similar one at ten dollars less an hour. In the same period xAI dismissed more than five hundred data labellers, and Scale AI disbanded a team of generalists.

The class meaning of this transformation is this: artificial intelligence carries proletarianisation into the uppermost layers of mental labour. Sections once counted as "professional", "expert" and "middle class" are selling their own expertise on a project basis, by the hour, on contract. And what they sell is the training of the machine that threatens the future of their own occupation. A lawyer is training the model that may one day take the place of lawyers; a programmer is improving the code assistant that may make programmers unnecessary.

In the first volume of Capital, Marx wrote that the machine turns the worker into "an appendage of the machine". What is happening today is one step beyond that: the worker becomes the machine's teacher, and then its rival. The process we described when writing on the misery of the white collar takes its barest form here. We took up the unravelling of the share-and-bonus order that binds the technology worker to the company in "As the Golden Handcuffs Come Undone".

---

The Return of the Piece-Wage

The dominant wage form in data work is the piece-wage, to which Marx devoted a chapter of its own. In Chapter 21 of the first volume of Capital, Marx writes that the piece-wage is a converted form of the time-wage, and that it is the wage form best suited to the capitalist mode of production. Because the piece-wage:

• Makes the worker supervise themselves; the employer is spared the cost of supervision.
• Increases the intensity of labour; the worker works faster in order to earn more.
• Stirs competition among workers, and makes collective organising harder.
• Leaves the quality standard open to being set by the employer alone; "defective" work is not paid.

On today's data platforms these four features are combined with algorithmic management. Every click of the worker is measured, their speed and accuracy are scored, and the worker whose score falls has their access to work restricted. In documents shared by the Data Labelers Association, founded in Kenya, a worker is seen to have received no payment at all for about twenty hours of labour: the work had been rejected on the grounds that it "did not meet the quality standard".

This is the most intense form of the process we defined as "mental Taylorism" in the gig economy dossier. Taylor measured and standardised every movement of the worker in the factory. In data work this measurement goes down to every step of the worker's thought. We examined the other faces of algorithmic supervision in the workplace in "When AI Becomes the Boss's Eye", the chipped tracking at Astor and the KVKK's email decision.

---

The Loop Is Narrowing

The human in the loop is not a permanent position. The figures on how companies use artificial intelligence show that the human is steadily being taken out of the loop.

According to the third Economic Index report Anthropic published in September 2025, the share of "directive" use, in which users hand a job over to artificial intelligence from start to finish, rose from 27 percent to 39 percent in eight months. For the first time, use for automation overtook use that complements the human. In companies' use through the API, the share of automation reached 77 percent; among individual users the rate was about 50 percent. A summary of the report is here.

This figure shows something important: while the individual worker uses artificial intelligence as an assistant, companies use it as a substitute. For capital, the human in the loop is a cost item that has to be removed from the loop. We told why the dream of the "lights-out factory" strikes capitalism's own contradiction in our reply to Alphan Manas, and how staffs are melted without dismissals in "No One Was Fired, the Positions Melted Away".

The study Generative AI and Jobs: A Refined Global Index of Occupational Exposure, published by the ILO in May 2025, calculates that one in four workers worldwide works in an occupation at risk of being affected by generative artificial intelligence. In high-income countries this rate rises to 34 percent. In the highest-risk category, women are represented at nearly twice the rate of men: in high-income countries 9.6 percent of women's employment is in this category, against 3.5 percent for men. Clerical and secretarial jobs, the fields in which women's labour has historically been concentrated, are among the fields affected most.

---

A Comparative Table

The company's story	The class reality
"Artificial intelligence learns by itself"	It is trained by the labels, scores and corrections of millions of workers
"The human in the loop is there for safety"	The human in the loop is the worker who carries the cost of the error
"Flexible, free, side income"	Piece-rated, insecure labour under algorithmic supervision
"A fully automatic service"	Often the real-time labour of humans who have been made invisible
"High pay for specialists"	Specialists train, on a project basis, the machine that will make their own occupation unnecessary
"A rise in productivity"	The intensification of labour, and the seizure of the cheap labour of the peripheral countries
"Technology that complements the human"	In company use, substitution predominates
"A neutral technical term"	A division of labour, a wage relation and a distribution of responsibility

---

Organising: From Invisible to Visible

This table may look bleak, but an important development has taken place in the last two years: data workers have begun to organise.

• The Global Content Moderators' Union Alliance was founded in Nairobi on 30 April 2025, with workers from nine countries taking part. According to UNI Global Union's statement, the alliance's demands include a living wage, stable employment, safe working conditions, real mental-health support during working hours, union representation, and holding the technology giants responsible for the low-paid, highly traumatic jobs in their own value chains. A Telus worker from Türkiye also spoke at the meeting.
• The Data Labelers Association was founded in Kenya and gathered more than three hundred members in its first week. Among its aims are a fair wage, legal support, and concluding collective agreements with the intermediary companies.
• The Çağrı-İş union is organising the Telus workers in İzmir and waging one of the first serious union struggles in this sector in Türkiye.

These experiences confirm a principle we defend in The Computing Worker's Handbook: the computing worker is not only the person who writes code. Everyone who labels data, moderates content and scores model outputs is part of this field of labour, and should have a place in the same organisation. We gathered our proposals on the forms in which dispersed labour can organise in "Organising 2.0".

---

Concrete Tasks

1. Break the invisibility. Produce research, news and testimony that make visible the labour behind the sentence "artificial intelligence did it". Start work in Türkiye, like Data Workers' Inquiry, that brings workers' own voices to the front.
2. Map the data workers in Türkiye. Which companies have content moderation, data labelling and model evaluation done in Türkiye? How many people, under what conditions, at what wages? This knowledge is the precondition of organising.
3. Build links with specialist data workers. Organise discussions, with engineers, lawyers, physicians and academics working project by project on platforms like Mercor, on the future of their own occupations. This is a section that does not see itself as working class, but is objectively being proletarianised. We discussed the place of professional organisations in this argument in the piece we wrote on the TMMOB's artificial intelligence report.
4. Develop demands against the piece-wage. A minimum hourly wage, a guarantee of payment for rejected jobs, transparency of quality standards, and a right to object to algorithmic scoring.
5. Load the responsibility back onto capital. Demand legal arrangements which hold that, in "human in the loop" mechanisms, legal responsibility belongs not to the worker but to the company that designed the system and makes the profit.
6. Define mental health as a matter of occupational health and safety. Have content moderation recognised as work that carries the risk of an occupational disease; demand independent psychological support during working hours, a limit on the duration of exposure, and rotation.
7. Enlarge international solidarity. Build links between the Global Content Moderators' Union Alliance, the organisations in Kenya, and the unions in Türkiye. In a sector where the same companies exploit the same workers in different countries, organising has to cross borders too.
8. Discuss a public alternative. Defend a model in which data and models are subject not to private property but to public and democratic control. We discussed this in detail in our pieces on public artificial intelligence and the Socialist AI Manifesto.

---

A Source Guide for Our Reader

Let us return to our reader's question. We have gathered the main sources worth reading in this field under four headings.

Theoretical frame

• Matteo Pasquinelli, The Eye of the Master: A Social History of Artificial Intelligence (Verso, 2023). A book that reads artificial intelligence, through Babbage and Marx, as the automation of the division of labour, and that can be a starting point in this field.
• Phil Jones, Work Without the Worker: Labour in the Age of Platform Capitalism (Verso, 2021). Takes up microwork along the axis of the reserve army of labour and the global surplus population.
• Antonio A. Casilli, Waiting for Robots: The Hired Hands of Automation (University of Chicago Press, 2025; the French original is En attendant les robots, 2019). A comprehensive study of the "digital labour" behind automation.
• Hamid R. Ekbia and Bonnie Nardi, Heteromation, and Other Stories of Computing and Capitalism (MIT Press, 2017). Discusses, with the concept of "heteromation", the transfer of work not to the machine but to unpaid or low-paid human labour.
• Madeleine Clare Elish, "Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction", Engaging Science, Technology, and Society (2019). On the burden of responsibility carried by the human who supervises.

Field research and workers' testimony

• Mary L. Gray and Siddharth Suri, Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (2019). A broad field study of workers doing microwork in the United States and India.
• Milagros Miceli and Julian Posada, "The Data-Production Dispositif", Proceedings of the ACM on Human-Computer Interaction (CSCW, 2022). On the relations of power in the process of producing data.
• DAIR and TU Berlin, Data Workers' Inquiry (from 2024 on). Reports written by data workers in their own hand, in the tradition of Marx's Workers' Inquiry.
• The Kenya data workers' pay survey by the Kenya Data Labelers Association and the WageIndicator Foundation.

Institutional reports

• ILO, Generative AI and Jobs: A Refined Global Index of Occupational Exposure (Gmyrek et al., 2025).
• ILO, World Employment and Social Outlook 2021: The Role of Digital Labour Platforms in Transforming the World of Work.
• Oxford Internet Institute, reports of the Fairwork project: assessments of platform and cloud work.
• Partnership on AI, Responsible Sourcing of Data Enrichment Services (2021). A document in which the sector admits its own problems; it should be read critically.
• Anthropic, Economic Index reports (from 2025 on). To be treated with caution because the data is the company's own, but important because it offers rare figures on patterns of use.

Journalism

• Billy Perrigo, "OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic", TIME (January 2023).
• Josh Dzieza, "AI Is a Lot of Work", The Verge / New York Magazine (June 2023).
• Business Insider, on Mercor's Meta project and the wage cuts (November 2025).
• Rest of World, on the TikTok moderators in Türkiye (2025).

---

The Worker inside the Table

Kempelen's chess automaton was exposed in the end. When the master inside the table was seen, the miracle disappeared, and what remained was a human being's labour.

Today's artificial intelligence is waiting for the same exposure. But this time, inside the table, there is not a single master but a working class stretching from Nairobi to İzmir, from Manila to Caracas. As these workers become aware of their own labour, come to know one another and organise, it will be seen more clearly that what is called "artificial intelligence" is in fact a social product of human labour. Then the real question can be asked: why is this intelligence, produced socially, private property?

Dear Young Comrades,

When you speak to a chatbot, you think that what faces you is a machine. Yet behind every polite sentence of that machine, behind every correct answer, there is a worker of your age, perhaps a graduate of the same department as you, perhaps doing click work at midnight to cover the cost of their education. In a country where eighty-two out of a hundred students have to work, this worker may well be your classmate. Instead of looking at technology with admiration or with fear, learn to see the labour behind it; we discussed the cost of handing the burden of thinking over to the machine in "Whom Is the University For?". It is very likely that tomorrow you too will stand inside that loop, as a labeller, a supervisor or a specialist trainer. When that day comes, know that you are not alone, and that you can organise.

We train the machine; we should be the owners of the machine too.

---