Blog · Review Essay · Last reviewed June 25, 2026

Ghost Work and the Hidden Labor of AI

Mary L. Gray and Siddharth Suri's Ghost Work is a necessary correction to the fantasy of clean automation. The book shows that many systems sold as artificial intelligence, seamless platforms, or frictionless services depend on human workers who are routed through software, kept at a distance from customers, and made difficult to recognize as coworkers in the first place.

Ghost work, in this review, means human judgment hidden behind an apparently automated interface: labeling, verification, moderation, transcription, quality control, feedback, safety testing, escalation, and exception handling. The problem is not that machines need people. The problem is that the arrangement can hide labor from credit, bargaining power, liability, and care.

The sharper definition is a labor relation, not a mood: ghost work exists when a service captures human discretion as if it were machine capacity while denying the worker the visibility, rights, and institutional status that would normally attach to that discretion.

The Book

Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass appeared in 2019. Microsoft Research lists publication on May 7, 2019, and the book's official site presents it as an account of the human workforce that makes web services and AI-mediated products appear smooth. That framing is still useful because it treats automation as an institutional arrangement, not only as a model capability.

Gray is an anthropologist and Suri is a computer scientist. That combination matters. The book is not only a labor ethnography and not only a technical critique. It asks how work is reorganized when software platforms break judgment into small tasks, route those tasks through APIs and queues, and then present the completed service as if the machine handled it.

The book draws on a five-year study of workers in the United States and India. Harvard Law School's report on a 2019 event with Gray describes the research as an ethnographic and computational study of people doing this work: the people who transcribe, label, verify, moderate, edit, sort, and repair the gaps left by automated systems.

Current Context

As of June 25, 2026, Ghost Work sits inside a more explicit labor-governance environment than the one around its 2019 publication. Directive (EU) 2024/2831 on platform work is in force, with Member State transposition due by December 2, 2026. EUR-Lex describes its aim as improving working conditions, determining correct employment status, regulating automated monitoring and decision-making systems, and increasing transparency in platform work. That directive does not cover every annotator, moderator, evaluator, support worker, or subcontractor behind AI systems, but it gives a legal vocabulary for the same problem Gray and Suri name: software-mediated work does not stop being work.

The EU AI Act adds a second layer. Annex III treats many AI systems used for recruitment, worker management, task allocation, monitoring, evaluation, promotion, and termination as high-risk. Article 26 requires employers deploying high-risk AI systems at work to inform workers' representatives and affected workers before use. Article 14 requires human oversight for high-risk systems, but the practical lesson from Ghost Work is that oversight itself is labor: it needs time, authority, training, protection from retaliation, and a route to change the system.

In the United States, the Department of Labor's 2024 AI best-practices roadmap remains guidance rather than binding law, but it is useful source material because it names worker empowerment, transparency, worker voice, human oversight, labor-rights protection, training, and responsible worker-data use as workplace AI concerns. NIST's AI Risk Management Framework and Generative AI Profile are voluntary, but they push AI governance toward lifecycle records: provenance, documentation, stakeholder input, measurement, incident response, and monitoring. Those records are incomplete if the people producing labels, feedback, moderation decisions, and safety evidence disappear from the file.

Buyer-side practice has also matured. Partnership on AI's responsible-sourcing work treats data labelers, data cleaners, and other contributors of human judgment as critical to AI development, and Fairwork's AI supply-chain work turns labor claims into auditable questions about pay, conditions, contracts, management, and representation. The governance question is no longer only whether a model works. It is whether the labor chain that made the model work can survive scrutiny.

Automation's Last Mile

The central concept is the "last mile" of automation. Some tasks can be formalized and scaled through software. Others require situated judgment, cultural context, visual recognition, language sense, discretion, or responsibility for ambiguous cases. Instead of admitting that humans remain inside the system, companies often route these tasks to a distributed labor force and keep the human contribution out of the product story.

That is why the book is more interesting than a simple complaint about bad gig jobs. Gray and Suri show a structural pattern: as machines improve, they do not simply replace human work. They create new frontiers of exception handling, training, labeling, testing, content review, fraud detection, customer support, data cleaning, and edge-case repair. The boundary between machine work and human work keeps moving, and the people near that boundary are often treated as temporary, interchangeable, and invisible.

Last-mile work is therefore not a residue left over until the model gets smarter. It is the handoff zone where the system meets ambiguity: the mislabeled image, offensive joke, regional phrase, borderline policy case, faulty transcript, suspect account, low-confidence prediction, or customer whose life does not fit the menu. If that handoff zone is staffed by people with no stability or appeal, the system's technical reliability is being purchased through labor instability.

The authors' own essay on automation's last mile describes ghost work as people working with programmers through APIs to fuel AI and internet automation. That description is important because it refuses the usual opposition between "human" and "machine." The system is human-machine labor. The problem is that only the machine receives the credit, valuation, and institutional protection.

A stronger definition follows: ghost work is not merely hidden labor near a machine. It is hidden discretion converted into infrastructure. The worker interprets an edge case, absorbs uncertainty, applies a policy, resolves a contradiction, or cleans a record; the platform records the result as throughput, quality, alignment, safety, or automation. The judgment remains human, but the institutional credit moves to the system.

The Interface Hides the Worker

Ghost Work is strongest when it shows invisibility as a design achievement. The customer sees a platform, app, recommendation, verification result, caption, cleaned dataset, moderated feed, or responsive service. The worker sees a queue, a task, a timer, a rejection risk, a thin instruction set, and often no clear path to contest nonpayment or platform discipline.

This is not accidental. The interface is built to make the service feel automatic. A user who believes the system is automated is less likely to ask who did the work, what they were paid, what risks they absorbed, or whether the task required care, skill, and judgment. The platform turns labor into a backstage process while keeping the polished front end available for investment, marketing, and user trust.

A useful visibility test is simple: would the user, buyer, regulator, or auditor judge the product differently if the interface disclosed the human work behind the answer? If the answer is yes, the automation claim is not only incomplete. It is shaping consent, procurement, and accountability.

That pattern belongs beside work on content moderation, data extraction, algorithmic management, and platform capitalism. A moderation queue, labeling platform, delivery app, customer-service tool, or AI evaluation pipeline can present human labor as a temporary patch on the road to full automation. But the patch becomes an operating model.

The interface also hides the chain of command. A rejected task, account suspension, quality score, or denied payment can be distributed across a requester, platform rule, vendor manager, automated fraud check, task rubric, and appeal queue. That diffusion matters because it makes responsibility hard to reach. The worker experiences management; the system presents procedure.

The AI-Age Reading

Read in 2026, the book has become more relevant, not less. Generative AI has expanded the public sense that software can write, see, classify, converse, summarize, evaluate, and act. But those capabilities still depend on data work, policy work, feedback work, safety testing, content review, red teaming, preference labeling, post-training evaluation, and customer support. The hidden workforce has not disappeared. It has moved deeper into the stack.

The AI-era danger is double invisibility. First, workers are hidden behind the interface. Then their work is hidden again behind the model. A chatbot answer, image classifier, fraud system, transcription service, or agentic workflow may feel like a direct expression of machine intelligence. In reality, it may rest on years of labeled data, crowd judgments, moderation policies, outsourced review, escalation labor, and human corrections that never appear in the product's story about intelligence.

That double invisibility matters for safety because human feedback is not just moral decoration. It is part of the control system. If output ratings, red-team findings, policy labels, and escalation decisions are produced under rushed, opaque, or punitive conditions, the resulting "alignment" record may be thin evidence for public trust. The worker's conditions become a model-quality issue.

Not every worker in this chain has the same status or exposure. A full-time researcher writing an evaluation protocol, a contractor rating chatbot responses, a vendor employee moderating graphic content, and a crowdworker labeling images are not interchangeable. The common pattern is that their work can be folded into a model or metric while their authority over the system remains small.

This matters for belief formation. When a system looks autonomous, people over-attribute agency to the machine and under-attribute responsibility to the institution. The worker disappears, but so does the manager, vendor, procurement officer, platform designer, policy writer, and executive decision-maker. Automation becomes a fog in which nobody seems answerable.

Agentic systems sharpen the issue. A workflow agent may be marketed as acting for a user, but workers still write task instructions, construct tool taxonomies, review action traces, label failures, handle exceptions, and repair bad records. If that corrective labor is invisible, the agent looks more competent than the organization can actually support.

Institutions and Labor Rights

The book's political force is its insistence that ghost work is work. It is not a hobby, glitch, side effect, or residue that will vanish when models improve. It is part of the production system. That means ordinary institutional questions apply: pay, benefits, scheduling, collective voice, appeals, health, safety, training, dignity, and accountability.

Kirkus summarized the book as an exploration of the hidden human labor force working with AI to power popular websites and apps, noting platform workers who handle judgment-call tasks under uncertain conditions. The ILR Review record identifies Benjamin Shestakofsky's 2019 review as engaging the book as a labor-relations text, which is the right shelf. This is not only about technology ethics. It is about labor organization under computational management.

For public agencies, schools, hospitals, nonprofits, and companies adopting AI, the practical lesson is to ask where the human work has gone. If a vendor claims an automated service, the buyer should ask who labels, reviews, escalates, audits, repairs, and absorbs harmful content or abusive users. If a model promises efficiency, the institution should ask whether it is creating a new class of hidden workers with no meaningful recourse.

That question should become a procurement requirement. A buyer should be able to request a labor bill of materials: annotation vendors, moderation vendors, evaluation workers, red-team contractors, subcontracting layers, worker exposure controls, pay and contract terms where available, grievance paths, and the person accountable when the supply chain fails. The site's AI bill-of-materials map and vendor governance protocol turn that habit into a checklist.

Partnership on AI's data-enrichment sourcing work points to the same buyer-side responsibility: fair compensation, clear instructions, worker feedback channels, attention to worker experience, and supply-chain transparency are not soft values outside the technical file. They are evidence about whether the system was built through reliable and accountable human judgment.

A labor bill of materials should protect workers while still making the system accountable. It can disclose role categories, task types, jurisdictions, vendor layers, aggregate pay practices, support mechanisms, quality-review rules, and grievance channels without naming individual workers or exposing sensitive moderation material. Privacy for workers should not become secrecy for buyers.

Governance and Safety

By June 25, 2026, the book's labor argument had become a governance issue. A serious AI review should include a hidden-labor register: who collected, labeled, cleaned, moderated, ranked, red-teamed, evaluated, escalated, corrected, or appealed the work; whether they were employees, contractors, crowdworkers, vendor staff, domain experts, unpaid users, or affected workers; and which protections, instructions, pay practices, support systems, and grievance channels governed the task.

The safety implication is simple: a person in the loop is not a safety control by itself. Human review becomes another layer of ghost work if the reviewer has no time, authority, training, job security, exposure protection, or right to challenge the system. NIST's AI Risk Management Framework and its Generative AI Profile are voluntary, but they point in the right direction by treating AI risk as a lifecycle problem involving design, development, use, evaluation, documentation, measurement, and governance.

For workplace and public-service deployments, hidden labor should therefore appear in the risk register. Who touched the data? Who reviewed the outputs? Who absorbs traumatic or abusive material? Who handles failed automation? Who can appeal a score, task allocation, account suspension, nonpayment, or disciplinary action? If those questions cannot be answered, the system is not merely opaque. It is administratively unfinished.

Procurement should bind that register to remedies. Contracts should require subcontractor disclosure at the right level of aggregation, worker-support obligations for harmful material, appeal and nonpayment processes, data-retention limits, audit rights, change notices when labor pipelines move, and incident reporting when worker conditions affect system quality or safety. A vendor's claim that "humans review outputs" should trigger more questions, not end the review.

Governance also has to track downstream ghost work. If an AI assistant makes local staff spend unpaid time correcting outputs, if a public chatbot pushes residents into hidden call-center queues, if a model creates new moderation burdens, or if a workplace system turns managers into rubber-stamp reviewers, the deployment has not removed labor. It has moved it into a less visible accounting line.

Where the Book Needs Care

Ghost Work was published before the current foundation-model boom. It does not cover today's full stack of data annotation firms, reinforcement learning from human feedback, AI safety contractors, synthetic-data operations, model-evaluation vendors, or global content-policy supply chains. Readers should treat it as a foundational map of the labor relation, not a finished account of every AI labor market that now exists.

It also should not flatten every form of AI labor into the same moral category. Some workers are employees with teams and benefits; others are crowdworkers paid per task; others are subcontracted through business-process outsourcing firms; others are unpaid users whose corrections become training traces. Jurisdiction, contract status, exposure risk, organizing power, and local labor law matter.

The book also has a reformist confidence that deserves scrutiny. Better platform design, worker communication, benefit systems, shared workspaces, and legal recognition matter. But the deeper conflict is that the economic value of ghost work often depends on keeping labor cheap, flexible, deniable, and difficult to organize. Reform has to confront the business model, not only the user interface.

Still, the book's empirical discipline is its strength. It does not lean on speculative futures. It starts with workers already doing the tasks that make the automated world appear smarter than it is. That makes it useful in an AI culture prone to abstraction, hype, and metaphysical arguments about whether machines will replace everyone.

What This Changes

The deepest lesson is that "AI" is often a social arrangement before it is a technical object. A model, platform, or agent can look like a single intelligence while actually coordinating many people, policies, datasets, metrics, incentives, and institutional decisions. The interface compresses that arrangement into a button, answer, score, or task completion.

Once that compression becomes ordinary, people start mistaking legibility for reality. The worker becomes a row in a dashboard. The judgment becomes a label. The label becomes training data. The training data becomes a model behavior. The model behavior becomes a product feature. The feature becomes evidence that the machine did the work. Each step makes the human contribution harder to see and easier to govern without consent.

The corrective is simple but demanding: when a system appears intelligent, look for the labor, the institution, and the appeal path. Ask who trained it, who repairs it, who reviews its edge cases, who is exposed to harm, who can refuse, who can organize, and who gets named when the system succeeds. Ghost Work gives that audit habit a human face.

This also changes how claims should be written. Instead of saying that "AI verified," "AI moderated," or "AI decided," name the workchain when it is known: model, vendor, worker, policy, threshold, escalation, and appeal. The site's claim hygiene protocol is built for that discipline. A clear sentence should not erase the people who made the sentence true.

That same discipline belongs in documentation. A useful model card or system card should not stop at benchmarks and intended use. It should say what categories of human labor shaped the dataset, tuning, evaluation, moderation, and deployment, and what remains unknown about subcontracting or worker conditions.

Source Discipline

Use the official book site, Microsoft Research, HarperCollins, and publisher or catalog records for book metadata. Use ethnography, peer-reviewed labor research, worker-led inquiry, and labor-rights reporting for working conditions. Use statutes, regulators, and standards bodies for legal or governance obligations. Do not use a vendor's ethics language as proof that workers have fair pay, safe conditions, bargaining power, or meaningful appeal.

For current AI deployments, source discipline also means separating the model from the labor chain. A model card can describe training and evaluation; it rarely proves the pay, safety, contract status, or subcontracting conditions of the people whose work shaped the system. Where those facts are unknown, say unknown and make the uncertainty part of the procurement risk.

This page makes no claim that any current AI system is conscious, divine, or AGI. It treats AI as a sociotechnical production system: models, vendors, interfaces, workers, managers, contracts, data, institutions, and accountability paths.

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