The Eye of the Master and the Labor Hidden Inside AI

The Book

Verso lists The Eye of the Master: A Social History of Artificial Intelligence as a 272-page paperback published in October 2023, ISBN 9781788730068. Ca' Foscari University identifies Pasquinelli as an associate professor of philosophy of science and principal investigator of the ERC-funded AIMODELS project. The Deutscher Memorial Prize records the book as its 2024 winner.

The book's intervention is sharper than the loose slogan that all technology is social. Pasquinelli traces how schemes for dividing labor, measuring conduct, and coordinating collective knowledge become conceptual resources for machine intelligence. Industrial machinery embodies analyzed work; statistical and cybernetic systems formalize relations among observation, prediction, correction, and command; machine learning acquires capability from organized datasets, labeling conventions, benchmarks, and feedback.

This is a genealogy, not a claim that a factory machine and a transformer are technically identical. Its value is to change the unit of analysis. The relevant system is not only model architecture. It is the model plus the work that generated its inputs, the institution that chose its target, the people who handle exceptions, and the authority that turns an output into a consequence. AI can abstract labor before deployment, depend on labor during development, and reorganize labor after deployment.

Current Context

As of August 12, 2026, European law recognizes parts of this labor-control problem, but its scopes and dates must not be collapsed. Directive (EU) 2024/2831 sets algorithmic-management requirements for digital labor platforms: limits on processing private conversations, off-duty data, emotional state, inferred sensitive characteristics, and predicted exercise of fundamental rights; information for people performing platform work and their representatives; human oversight; reasons and review for automated decisions; and human decisions for account or contractual restrictions with equivalent detriment. Member States must transpose it by December 2, 2026. It is not yet a uniform EU-wide workplace regime, and it does not cover every employer.

The AI Act has a different scope and timetable. Annex III names certain systems for recruitment, work allocation, monitoring, evaluation, promotion, and termination as high-risk, subject to classification exceptions. But Regulation (EU) 2026/1744, published July 24, 2026, moved the relevant Chapter III regime for Annex III systems to December 2, 2027. That delay includes the deployer obligations in Article 26, such as workplace notice for covered high-risk systems. The Act's general application date of August 2, 2026 therefore does not make those employment-system duties presently operative.

Evidence also requires careful verbs. A July 2026 Joint Research Centre working paper, using the AIM-WORK survey across all 27 EU Member States, reports associations between algorithmic management and lower autonomy, less ability to take breaks, and higher stress; direct algorithmic direction of task execution and pace shows the strongest associations with reduced discretion and work intensification. The study is cross-country observational evidence, not proof that every monitoring tool causes the same outcome. An ILO brief published in April 2026 makes the parallel distinction for generative AI: occupational exposure is an early-warning indicator, not a forecast of adoption, layoffs, or net employment effects. Actual wages, hours, transitions, and job quality still have to be measured.

The United States comparison is more fragmented. The GAO's revised 2025 review found that digital surveillance can support safety or operations but can also increase anxiety or injury risk, and that flawed benchmarks or incomplete task records may contribute to adverse evaluations, pay, discipline, or termination. It also reported that several federal agencies had rescinded prior resources or were reassessing them in 2025. NIST's AI Risk Management Framework 1.0 remains voluntary and is under revision. It helps structure risk work; it is not a labor-rights or worker-recourse regime.

Labor as Machine Intelligence

Pasquinelli's most productive move is to read the machine as organized knowledge with an owner and a command structure. Mechanization does not only substitute machinery for a task. It changes which parts of a practice are visible, which skills are separated, who controls timing and comparison, where exceptions go, and whose description of competent work becomes standard.

The historical hinge is the Babbage principle. In On the Economy of Machinery and Manufactures, Charles Babbage applied division of labor to mental as well as manual operations and analyzed how a process could be split among different levels of skill and cost. Pasquinelli's use of Babbage is political: decomposition is already a redistribution of knowledge and bargaining power before it becomes a technical design.

That argument produces a more useful definition of automation. The sequence is practice → decomposition → instrumentation → model or rule → delegated decision → residual work → feedback. A failure at any step can be hidden by the word automation. A system may automate only the recorded cases while workers repair exceptions; it may accelerate a task without removing a job; it may generate a recommendation while a supervisor rubber-stamps it; or it may replace one role while creating lower-paid checking, escalation, or customer-recovery work elsewhere.

An AI labor ledger should therefore distinguish at least five groups: people whose prior writing, images, code, records, or workplace routines become inputs; people paid to collect, annotate, moderate, evaluate, or red-team data and outputs; domain workers whose judgment is translated into rules and benchmarks; downstream workers allocated, scored, assisted, or displaced by the system; and exception workers who repair what the formal workflow cannot handle. Calling all five simply users or human feedback erases different contracts, risks, and claims.

The ledger does not settle ownership by itself. Collective contribution does not automatically make every dataset common property, nor does a model's dependence on human work prove that every output belongs to every contributor. Copyright, privacy, confidentiality, employment law, contract, collective bargaining, and sector rules answer different parts of the question. Pasquinelli's achievement is prior to that allocation: he makes it harder to pretend that capability appeared without a social production process.

The political test then shifts from Will AI take jobs? to a set of separable questions. Which tasks are technically exposed? Has the system actually been adopted? Which decisions has it acquired authority to shape? Does it change staffing, wages, hours, pace, discretion, safety, or skill? Who receives any productivity gain? And can workers contest the resulting record? Exposure, deployment, displacement, and work intensification are related, but none is a proxy for the others.

The Master's Eye

The title points to supervisory asymmetry. The eye is not merely a camera or a manager watching a screen. It is a relation in which one party defines the units of performance, observes many people at scale, combines records across time, and can turn the comparison into assignment, pay, promotion, discipline, or exclusion. The observed person sees only a fragment and may have no power to change the category or correct the file.

AI expands that relation by making institutional perception cheap and continuous. A camera becomes pose or identity features; work becomes productivity telemetry; a conversation becomes a quality score; a schedule becomes a forecast of availability; an exception becomes a risk flag. The system sees only what its instrumentation and labels allow, but the institution can still act as though the description were the person.

The decisive governance question is therefore not whether a model perceives accurately in the abstract. It is perception for what intervention? A false positive used for a low-stakes prompt differs from the same error used to cut shifts. An accurate measure of keystrokes can still be an invalid measure of contribution. A generated summary can reproduce every sentence correctly and still omit the contextual work that should block a disciplinary inference.

Authority is usually distributed across a chain: a vendor selects features, procurement accepts defaults, an employer supplies data, a model produces a score, a workflow routes it, and a supervisor acts. Each participant can claim that someone else made the decision. A useful record names the provider, deployer, labor contractor or client, data owner, operational decision-maker, reviewer, and remedy owner separately. Otherwise distributed production becomes distributed deniability.

The eye also changes what it watches. A dashboard defines performance; workers adapt to the dashboard; adaptation becomes the next dataset; management reads that dataset as confirmation that the metric captured reality. Data Driven, The Quantified Worker, and The Boss Becomes a Dashboard show this recursive workplace loop. Pasquinelli supplies its longer history: AI did not invent the managerial gaze, but it can make the gaze faster, more portable, and harder to locate.

The AI-Age Reading

Read in 2026, The Eye of the Master corrects two symmetrical errors. One treats AI as a self-originating intelligence arriving from outside social life. The other treats it as a neutral instrument whose politics begin only with a bad use. Pasquinelli shows why the politics begins earlier: choosing what to record, how to divide a task, which label counts as success, and whose behavior becomes training material already allocates visibility and power.

Large language models make that history vivid but should not be described as simple containers of culture. Their capabilities depend on transformed statistical patterns learned from corpora, plus data selection, cleaning, annotation, preference work, evaluation, infrastructure, and repeated user interaction. No model response can be mapped neatly to one worker or source. Even so, the inability to trace a sentence to a contributor does not erase the labor, rights, or institutional choices in the production chain.

The useful question is not whether the model is secretly a person. It is which human capabilities and records were made available to it, under what permissions; which evaluators and operators keep it usable; who can convert its output into an authoritative record; and whether affected people can refuse, inspect, correct, appeal, or share in the benefit. These are governance questions that remain answerable without making claims about consciousness or AGI.

Agentic workflows sharpen the point because they join prediction to permission. An agent can draft, route, rank, schedule, send, purchase, or change a file only because an institution has already decomposed work into tickets, tools, roles, credentials, schemas, calendars, and escalation paths. The agent runs through labor's prior grammar. Safety therefore depends on bounded tool scopes, separation between proposing and committing consequential actions, identity-bound logs, revocation, rollback, and a named human who can halt the workflow. The Agent Tool Permission Protocol addresses that action layer.

Pasquinelli's lens also prevents an easy category error. A model that performs many tasks does not establish that an occupation will disappear, and an occupation that remains does not establish that workers were unharmed. Employers can use the same capability to reduce drudgery, intensify pace, deskill a role, expand output without sharing gains, eliminate positions, or create new exception work. Those are organizational choices to be measured, not properties read directly from a benchmark.

Governance and Safety

The practical instrument is a labor-control file maintained before procurement and through operation. It should cover any system that observes work or materially shapes hiring, assignment, scheduling, pace, pay, evaluation, promotion, discipline, safety, accommodation, or termination, whether the component is marketed as AI, analytics, optimization, fraud prevention, or productivity software.

• Purpose and decision map: the problem, affected population, decisions influenced, intended benefit, prohibited uses, legal basis, and the counterfactual process against which the system will be judged.
• Labor and provenance ledger: upstream source contributors; annotation, moderation, evaluation, and red-team labor; domain expertise translated into rules; downstream affected workers; exception and repair work; contractors and subcontractors; compensation and working-condition controls; datasets, versions, and retention limits.
• Authority map: provider, deployer, employer, client, labor contractor, data owner, supervisor, reviewer, worker representative, incident owner, and external remedy. Record who may inspect features, change thresholds, override an output, repair a personnel file, suspend use, and compensate harm.
• Impact measures: wages, hours, staffing, pace, schedule volatility, breaks, injuries and psychosocial load, discretion, skill use, accommodations, false flags, override rates, appeal outcomes, subgroup error, exception backlog, productivity gains, and how those gains are distributed. Compare by task and workplace; do not substitute an occupational-exposure score.
• Notice and recourse: plain-language notice before use, worker and representative consultation, the data and rules relevant to a decision, a reason specific enough to challenge, paid time to seek review, a competent reviewer outside the original automated chain, anti-retaliation protection, correction of downstream records, and deadlines for remedy.
• Boundaries and stop conditions: bans or strict limits on off-duty collection, private communications, emotion inference, union or protected-activity prediction, inferred sensitive attributes, covert secondary use, indefinite retention, and unvalidated repurposing; thresholds for pausing deployment; rollback and deletion procedures; and post-incident notice.

Meaningful human oversight is an operating condition, not the presence of a person near the workflow. The reviewer needs paid time, adequate staffing, training, access to the evidence and applicable policy, authority to depart from the system, protection for doing so, and a way to repair every record or downstream action that inherited the error. If the human is evaluated for following the score, cannot see its basis, or cannot restore pay or access, the human is an accountability buffer rather than an overseer.

Procurement must bind vendors to that file. A model card cannot reveal whether local managers use a score to deny shifts; an employer's impact assessment cannot establish how a subcontractor treated data workers. Contracts should require relevant documentation, change notice, audit access, retention and deletion controls, support for explanations and appeals, incident cooperation, subcontractor flow-down, security obligations, and exit assistance. The Vendor and Platform Governance, AI System Inventory, and AI Audit Trails pages provide compatible records.

Worker participation is a control because workers know where formal tasks diverge from actual practice, which metrics punish safe adaptation, and which exceptions keep the process functioning. Consultation should occur while a system can still be changed, with enough information and independent expertise to evaluate it. It does not replace collective bargaining, health-and-safety duties, privacy rights, equality law, or sector-specific obligations.

The safety case is incomplete until it covers both model behavior and employment power. A technically accurate system can be unsafe if the target is invalid, the pace injurious, the surveillance disproportionate, the appeal ineffective, or productivity gains purchased through uncompensated exception work. Conversely, automation can reduce hazardous or repetitive work when deployment is evaluated against a real baseline and workers share authority over its design and consequences.

Where the Book Needs Care

The book's scale is both its strength and its risk. It connects political economy, epistemology, industrial organization, cybernetics, and neural networks across two centuries. Marc Kohlbry's Critical Inquiry review identifies a tension between the book's commitment to recovering labor and an intellectual history still organized around prominent theorists and technical programs. Workplace testimony, shop-floor conflict, colonial extraction, gendered service work, and contemporary data labor do not receive equal depth.

The genealogy can also be overextended. A Babbage engine, a statistical classifier, a computer-vision system, a recommender, and a transformer have different architectures, objectives, training processes, error modes, and institutional dependencies. The Babbage principle illuminates decomposition and cost control; it does not causally explain every design choice in modern machine learning. Technical analysis still has to show exactly where data, labels, objectives, evaluation, and human intervention enter a given system.

Nor does the category of collective intelligence settle competing claims. A patient's confidential record, a worker's telemetry, an artist's image, a programmer's repository, a public-domain text, and an annotator's judgment do not carry the same permissions or remedies. A labor account can reveal whose contribution is missing from the balance sheet, but law and collective institutions must still decide consent, compensation, confidentiality, attribution, bargaining, and public benefit.

The critique should not romanticize the unautomated workplace. Human managers can be inconsistent, discriminatory, coercive, or unsafe; paperwork can hide responsibility as effectively as software. Some automation reduces dangerous or repetitive work, improves accessibility, or makes decisions more consistent. The right comparison is the documented baseline and a distributional outcome, not an assumption that either human discretion or machine regularity is inherently fair.

Finally, the book does not supply a deployment standard. Employers, unions, regulators, schools, clinics, and public agencies still need sector law, task-validity evidence, security and privacy analysis, worker consultation, accessibility testing, procurement terms, incident response, and enforceable remedy. The book supplies a durable diagnostic: look for analyzed labor and command inside the machine. It does not replace the compliance file.

What This Changes

The Eye of the Master clarifies the conversion by which a social world becomes a machine-readable one. A workplace is decomposed into tasks; tasks become fields, labels, targets, and permissions; a model or rule turns those records into decisions; people adapt to the decisions; and the adaptation becomes the next record. This is the bridge between labor history and recursive reality.

The recursion matters because the model does more than describe prior work. It can produce schedules, rankings, summaries, warnings, evaluations, and synthetic personnel records that alter the work it will later observe. A productivity target can remove the contextual acts that do not score; a generated review can become evidence in the next evaluation; an exception routed out of sight can make the formal process appear more reliable than it is. The system partly manufactures the reality used to validate it.

Responsibility must follow that loop. The data team is responsible for provenance and measurement limits; the vendor for documented behavior and changes; the employer or deployer for purpose, local validation, and working conditions; the supervisor for consequential use; procurement for enforceable access and exit terms; and a named reviewer for correction and remedy. None can discharge its role by pointing to the model or to a nominal human at the end.

The durable diagnostic is concrete. When a system claims intelligence, identify the work that became data or procedure. When it claims perception, identify the intervention the perception authorizes. When it claims autonomy, map the prior routines, permissions, and exception handlers that make action possible. When it claims efficiency, measure who receives the gains and who absorbs the residual work, risk, or lost discretion. When it claims oversight, test whether a person can actually stop, reverse, and repair.

Pasquinelli makes AI less mysterious without making it harmless. The risk is not a conscious machine master. It is a historically familiar arrangement of observation, ownership, and command embedded in infrastructure and presented as neutral competence. The answer is not a counter-myth about technology. It is shared power over the categories, records, decisions, and benefits produced from collective work.

Source Discipline

This review separates Pasquinelli's thesis, historical evidence, current empirical findings, law, voluntary guidance, and this page's recommendations. Verso supports publication details and the publisher's account of the argument; Ca' Foscari and the Deutscher Memorial Prize support author and award claims; Babbage's public-domain text supports the historical principle. The labor-control file, five-part labor ledger, and supervisory-asymmetry definition are this review's synthesis, not passages or named frameworks from the book.

Current claims use dated primary sources reviewed August 12, 2026. The page distinguishes a directive awaiting national transposition from a uniformly operative rule, the AI Act's general application date from its delayed Annex III high-risk regime, a voluntary NIST framework from law, an observational association from a causal finding, and occupational exposure from adoption or job loss. Those distinctions are substantive: removing them would exaggerate both protection and harm.

Vendor claims, model benchmarks, and exposure indices are leads, not workplace outcomes. A defensible deployment claim needs a named population, task, baseline, period, decision threshold, error and override record, wage and hours evidence, safety indicators, subgroup analysis where lawful and appropriate, appeal results, and known limitations. Generated summaries and managerial impressions do not replace those records.

This page makes no claim that an AI system is conscious, divine, or AGI. It analyzes sociotechnical production and supervision: models, data, workers, managers, vendors, contracts, records, law, and feedback loops.

Related Pages

• Hidden production: Ghost Work, Feeding the Machine, Behind the Screen, and Data Enrichment Labor.
• Workplace control and feedback: Data Driven, The Quantified Worker, The Boss Becomes a Dashboard, and Algorithmic Management.
• Operational rights and records: AI in Employment, Human Oversight, Notice and Appeal, AI Audit Trails, and AI Bill of Materials.
• Institutional controls: Vendor and Platform Governance, Agent Tool Permission Protocol, Privacy and Data, and Research Integrity.

Sources

• Verso Books, The Eye of the Master: A Social History of Artificial Intelligence, publisher listing for title, author, ISBN, page count, publication month, and the publisher's synopsis, reviewed August 12, 2026.
• Ca' Foscari University of Venice, Matteo Pasquinelli curriculum, academic position, department, AIMODELS project, and research profile, reviewed August 12, 2026.
• The Deutscher Memorial Prize, Past recipients, 2024 award record for The Eye of the Master, reviewed August 12, 2026.
• Charles Babbage, On the Economy of Machinery and Manufactures, 1832 public-domain text on division of manual and mental labor, reviewed August 12, 2026.
• Marc Kohlbry, review of The Eye of the Master, Critical Inquiry, May 30, 2024, critical reception and historiographic limits, reviewed August 12, 2026.
• European Union, Directive (EU) 2024/2831 on improving working conditions in platform work, Articles 7–12 and 29 on data limits, transparency, human oversight, review, safety, and the December 2, 2026 transposition deadline, reviewed August 12, 2026.
• European Union, Regulation (EU) 2024/1689, the Artificial Intelligence Act, Annex III employment and worker-management uses and Article 26 deployer duties, read with the 2026 amendment, reviewed August 12, 2026.
• European Union, Regulation (EU) 2026/1744, July 24, 2026 official text amending the AI Act and moving the relevant Annex III Chapter III application date to December 2, 2027, reviewed August 12, 2026.
• European Commission Joint Research Centre, Alvaro Mariscal de Gante Martin and Davide Villani, "Algorithmic management and working conditions in Europe", JRC Working Papers Series on Labour, Education and Technology 2026/05, July 6, 2026, survey scope and reported associations, reviewed August 12, 2026.
• International Labour Organization and NASK, "Generative AI and Jobs: A Refined Global Index of Occupational Exposure", ILO Working Paper 140, May 20, 2025, task-level exposure methodology and limits, reviewed August 12, 2026.
• International Labour Organization, "New ILO brief explains what AI exposure indicators reveal about jobs", April 17, 2026, distinction among exposure, adoption, and observed labor-market outcomes, reviewed August 12, 2026.
• U.S. Government Accountability Office, Digital Surveillance: Potential Effects on Workers and Roles of Federal Agencies, GAO-25-107126, published September 2 and revised December 10, 2025, evidence review and federal-oversight status, reviewed August 12, 2026.
• National Institute of Standards and Technology, AI Risk Management Framework, voluntary status, current revision notice, and lifecycle risk-management context, reviewed August 12, 2026.
• National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024, model, use-case, value-chain, provenance, and human–AI configuration context, reviewed August 12, 2026.

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

• Amazon, The Eye of the Master by Matteo Pasquinelli, affiliate listing, reviewed August 12, 2026.

Return to Blog · Return to Books