Blog · Review Essay · Last reviewed June 25, 2026

The Eye of the Master and the Labor Hidden Inside AI

Matteo Pasquinelli's The Eye of the Master: A Social History of Artificial Intelligence is a 2023 book about AI as the automation of labor, social knowledge, measurement, supervision, and collective behavior. Its central value is that it refuses the usual myth that artificial intelligence descends mainly from an effort to imitate an isolated brain. AI also descends from factories, divisions of labor, cybernetic control, image recognition, surveillance, and the long institutional habit of turning human activity into procedures a machine can repeat.

For this review, the master's eye means a control pattern: first decompose work into observable parts, then measure those parts, then use the record to command, price, replace, or discipline the worker. The AI-era question is not whether a model has become a mind. It is whether collective practice has been captured as machine capacity while workers, users, and affected publics lose power over the categories, records, and feedback loops built from them.

The Book

The Eye of the Master was published by Verso on October 10, 2023. Publisher and retail listings give the book as a 272-page paperback with the subtitle A Social History of Artificial Intelligence, ISBN 9781788730068, and ISBN-10 1788730062. Pasquinelli is a philosopher of science whose work joins political economy, media theory, and the history of automation; his Ca' Foscari profile identifies him as Associate Professor in Philosophy of Science and principal investigator of the ERC AIMODELS project. His own site and the Deutscher Memorial Prize record note that the book won the 2024 prize.

The book's argument is direct and useful: AI should not be understood only as an attempt to reproduce biological intelligence. It should also be understood as a history of encoding social practice. Algorithms emerge from patterned human activity. Industrial machines embody divisions of labor. Cybernetic and connectionist systems formalize feedback, perception, and distributed control. Machine learning captures collective traces, labels, images, habits, and classifications. What looks like machine intelligence often contains prior human organization compressed into technical form.

That makes the book a strong companion to work on automation, hidden labor, surveillance, legibility, cybernetics, algorithmic management, and the politics of AI infrastructure. Pasquinelli is not simply saying that AI exploits workers at deployment time. He is making a deeper claim: the forms of labor, supervision, and social cooperation that precede AI shape what AI is able to become.

Current Context

As of June 25, 2026, Pasquinelli's argument lands in a policy environment where labor control and AI governance are no longer separate conversations. Directive (EU) 2024/2831 on platform work is in force and, according to EUR-Lex, aims to determine correct employment status, regulate automated monitoring and decision-making systems, and increase transparency in platform work. The EU AI Act separately treats many employment, recruitment, task-allocation, monitoring, evaluation, promotion, and termination systems as high-risk, and includes worker-notice duties for high-risk AI used at work.

That context matters because the master's eye is not only a metaphor from the factory. The European Commission's Joint Research Centre defines algorithmic management as computer-programmed procedures that coordinate labour input and can assign shifts, deliver instructions, assess performance, and assign rewards or penalties. The U.S. Department of Labor's 2024 AI best-practices roadmap is nonbinding guidance, but it names worker voice, transparency, meaningful human oversight, labor-rights protection, training, and responsible worker-data use as workplace AI concerns. Those are exactly the points where Pasquinelli's history becomes operational.

NIST's AI Risk Management Framework and Generative AI Profile add the documentation layer: governance, mapping, measurement, management, provenance, testing, and incident disclosure. Partnership on AI and Fairwork add the labor-supply-chain layer by asking how data enrichment, annotation, moderation, evaluation, and outsourced human judgment are sourced and governed. The current governance lesson is plain: a system's safety case is incomplete if it documents the model but not the labor and supervisory relations that made the model possible or made its deployment authoritative.

Labor as Machine Intelligence

The most productive move in the book is its labor theory of automation. Pasquinelli reads the machine as a crystallization of collective knowledge. The factory does not merely replace muscle with mechanism. It also reorganizes attention, timing, comparison, counting, skill, error, and supervision. In that world, the machine is not outside labor. It is built from a prior analysis of labor.

The book's anchor for this claim is what is known as the Babbage principle. Charles Babbage, the same figure behind the Difference Engine, observed that the division of labor could be applied to mental as readily as to manual work, and that breaking a job into graded parts let an owner buy the exact quantity of each kind of labor a task required. Pasquinelli treats that insight as the hidden template of machine intelligence: before a machine can mechanize a task, the task has already been decomposed, measured, and priced as labor. The calculating engine inherits the workshop, not the brain.

This matters for AI because contemporary systems often appear as if they arrive from pure mathematics or natural cognition. The training set, annotation workflow, benchmark, user trace, workplace record, image archive, customer ticket, school dataset, and moderation queue are treated as technical inputs. Pasquinelli pushes the reader to see them as social forms. They are residues of work, conflict, classification, and institutional purpose.

That view changes the labor politics of AI. If AI systems contain accumulated social intelligence, then the question is not only whether workers will be replaced by machines. It is also who owns the machines built from collective practice, who governs the categories used to train them, who can contest their outputs, and who benefits when ordinary skills are abstracted into a product.

A sharper definition follows: AI labor is not only the labor displaced after deployment. It is the labor abstracted before deployment, the labor hidden during training and evaluation, and the labor disciplined after the system becomes a dashboard, score, agent, or workflow. The same model can contain all three: prior human expression as training material, data workers as tuning and safety labor, and downstream workers as monitored subjects who generate the next round of traces.

The Master's Eye

The title points to supervision. A factory master does not need only machines that act; he needs a way to observe, compare, standardize, and command. The "eye" is the managerial gaze that turns distributed activity into something measurable and controllable. In an AI setting, that gaze can be camera, sensor, dashboard, metric, model, classifier, recommender, risk score, workflow tool, or generated summary.

This is where Pasquinelli's history connects to surveillance. Image recognition and behavioral modeling do not simply help machines perceive the world. They also help institutions decide which aspects of the world count as actionable. A face becomes a feature vector. A worker becomes productivity telemetry. A student becomes risk and performance signals. A patient becomes notes, codes, and predicted costs. A public becomes sentiment, influence, and targetable segments.

The key insight is that AI vision is rarely innocent perception. It is often perception organized for intervention. The model sees in the format required by the institution that will act on the seeing. That makes the politics of AI inseparable from the politics of what gets made visible, what remains unmeasured, and who is forced to live inside the resulting description.

This belongs beside the site's workplace-control thread because the eye becomes recursive. A dashboard defines performance, workers adapt to the dashboard, the adaptation becomes new data, and the institution treats the new data as proof that the dashboard sees reality. That loop appears in Data Driven, The Algorithm, The Quantified Worker, and The Boss Becomes a Dashboard. Pasquinelli gives the longer genealogy: the loop did not start with AI. AI inherits and accelerates it.

The AI-Age Reading

Read in 2026, The Eye of the Master is a useful antidote to two bad stories. The first is the magical story: AI as a new mind descending on society from the clouds of computation. The second is the purely instrumental story: AI as a neutral tool that becomes political only after misuse. Pasquinelli's history makes both stories too simple. AI is already political because the abstractions that make it work are drawn from organized life.

The book is especially helpful for thinking about large language models and generative systems. These systems are trained on vast deposits of human expression, institutional documentation, software, scholarship, forums, manuals, art, and ordinary writing. Their fluency can feel like synthetic cognition. But much of that fluency is accumulated linguistic labor: writing, editing, teaching, debating, documenting, joking, troubleshooting, translating, and explaining. The model answers back because a culture has already spoken into the archive.

That does not make AI fake. It makes it social. The operational question becomes: what kind of social intelligence has been captured, under what permissions, through which infrastructures, for whose profit, and with what ability for people to refuse, inspect, repair, or share control?

Agentic systems make the point sharper. A workflow agent can appear to plan, route, summarize, check, and act. But the action space is already a decomposed workplace: tickets, files, permissions, playbooks, issue labels, tests, calendars, procurement rules, dashboards, and escalation paths. The agent does not float above labor. It runs through labor's prior grammar. Governance has to ask who wrote that grammar, whose work it encodes, what logs it leaves, and who can pause or revise the workflow when the agent turns old control into faster control.

Governance and Safety

The practical instrument is a labor-control audit. Before deploying AI in a workplace, platform, school, clinic, public agency, or supply chain, an institution should identify the labor being decomposed, the observations being captured, the categories being imposed, the model or rule that converts traces into action, and the authority path that lets a person contest the result.

The audit should answer six questions. What prior work was converted into data, examples, labels, policies, benchmarks, or workflow rules? Who performed the data work, evaluation work, moderation work, safety work, or exception handling? What decisions will the system influence: hiring, assignment, pricing, scheduling, scoring, discipline, access, benefits, education, care, or policing? What records will it create, who can inspect them, and how long are they retained? What human oversight exists with real authority to change the outcome? What incident path exists when the system harms a worker, user, or affected public?

For procurement, that means asking for more than a model card. The buyer needs a labor ledger, data-provenance record, system inventory entry, human-oversight plan, worker-notice plan, appeal path, monitoring plan, and stop condition. If a vendor cannot say where human judgment entered the system, how workers are protected, what data categories are used for supervision, or how a person can correct the machine-shaped record, the system is not merely opaque. It is administratively unfinished.

The safety risk is not that AI becomes a conscious master. The risk is that managerial command becomes embedded in infrastructure while appearing as neutral intelligence. A tool can intensify work, hide responsibility, make appeals harder, punish lawful breaks, score affect, flatten expertise, or turn generated summaries into personnel records without ever making a final autonomous decision. That is why human oversight must mean paid time, source access, authority, anti-retaliation protection, and records fit for later review.

Where the Book Needs Care

The book is ambitious and theoretically dense. Its reach is part of the value, but it can also make the argument feel like a chain of historical workshops rather than a single continuous social history. Reviews in Critical Inquiry and Media Theory both recognize the book's intervention in critical AI studies, while the Critical Inquiry review also notes tension between a claimed history from below and an intellectual history centered on political epistemology, Marx, Hayek, Babbage, cybernetics, and neural networks.

That caveat is important. The book explains how labor and social relations shape machine intelligence, but readers looking for detailed workplace ethnography, data-labeling testimony, content-moderation accounts, or procurement-level governance will need companion texts. Ghost Work, Atlas of AI, Behind the Screen, Automating Inequality, and Data Driven supply more granular views of contemporary labor and surveillance systems.

The book can also underplay the technical specificity of different AI systems if read too quickly. A Babbage engine, a perceptron, a recommender system, a computer-vision model, and a transformer do not encode social practice in identical ways. The strongest reading treats Pasquinelli's argument as a genealogy of abstraction and control, not as a substitute for technical or institutional analysis.

There is also a governance limit. A labor theory of automation does not by itself tell an employer, regulator, union, school, or public agency which controls are enough for a specific system. That requires sector law, validation evidence, security review, privacy analysis, worker consultation, accessibility testing, incident response, and procurement terms. The book supplies the suspicion: look for labor and command inside the machine. It does not supply the whole compliance file.

What This Changes

The reason to add The Eye of the Master is that it names the hidden conversion by which social reality becomes machine reality. A workplace, archive, platform, city, classroom, or bureaucracy is first decomposed into tasks, categories, counts, images, labels, and feedback. Then those forms are fed into systems that appear to perceive, decide, write, or optimize. The machine seems to have intelligence because human and institutional intelligence has already been arranged for extraction.

This is the bridge between labor and recursive reality. Once a model is trained on institutional traces, it can begin producing new traces in the same format: summaries, rankings, classifications, generated reports, recommendations, alerts, and synthetic records. Those outputs are then read by people and organizations as evidence. The system does not merely automate a world. It helps produce the next version of the world it will later learn from.

Pasquinelli's best contribution is to make AI less mysterious without making it less dangerous. The danger is not that a machine has become a free-floating mind. The danger is that old arrangements of labor, command, surveillance, and ownership can return as apparently intelligent infrastructure. If the master's eye is embedded in the model, then AI governance has to ask who is doing the seeing, who is being seen, what social knowledge has been captured, and whether the people inside the loop have any real power over the machine built from them.

The practical reading is a source-trail discipline. When a system claims intelligence, ask what work was decomposed. When it claims perception, ask what institution will act on what it sees. When it claims autonomy, ask whose prior routines, examples, permissions, and exceptions make autonomy possible. When it claims efficiency, ask who receives the gains and who inherits the intensified record. That habit connects this page to the site's work on audit trails, vendor governance, AI bills of materials, workplace scoring, and hidden labor.

Source Discipline

This review separates bibliographic facts, author biography, published reviews, law and standards, and interpretation. Verso, Pasquinelli's site, Ca' Foscari, the Deutscher Memorial Prize, Amazon, and Google Books support publication, author, prize, ISBN, page-count, and table-of-contents claims. Critical Inquiry and Media Theory support reception and interpretive context. EUR-Lex, the European Commission Joint Research Centre, NIST, the U.S. Department of Labor, Partnership on AI, Fairwork, and the International Labour Organization support current governance and labor-context claims.

The interpretive claim is deliberately bounded. Pasquinelli's book is used here as a genealogy of abstraction, labor, and control. This page does not claim that any AI system is conscious, divine, or AGI. It treats AI systems as sociotechnical production and supervision systems: models, data, workers, labels, users, managers, vendors, records, law, and feedback loops.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books