Blog · Review Essay · Last reviewed June 19, 2026

Bullshit Jobs and the Automation of Pointless Work

David Graeber's Bullshit Jobs is a provocative theory of paid work that even the worker experiences as pointless, unnecessary, or harmful. Its AI-era value is not that every empirical claim survives scrutiny. It is that the book names a danger now moving into software: institutions can automate work without first asking whether the work deserves to exist.

Here, pointless-work automation means the use of software to accelerate reports, meetings, ticket flows, compliance artifacts, performance traces, email loops, or managerial rituals whose beneficiary, decision point, evidence of usefulness, and deletion rule have never been clearly named.

The Book

Bullshit Jobs: A Theory was published by Simon & Schuster on May 15, 2018. The publisher lists it at 368 pages and presents it as an expansion of Graeber's 2013 essay "On the Phenomenon of Bullshit Jobs," which had already produced a large public response.

Graeber was an anthropologist, public intellectual, and activist. Simon & Schuster identifies him as a former professor of anthropology at the London School of Economics and as the author of Debt: The First 5,000 Years and, with David Wengrow, The Dawn of Everything. LSE's own memorial material describes him as a professor of anthropology whose work connected public debate, bureaucracy, capitalism, and social possibility.

The book's argument is blunt: modern economies have created large amounts of paid work that the worker privately believes contributes little or nothing. Graeber's emphasis is not simply low pay, unpleasant work, or exploitation. His target is a stranger condition: being paid to perform, justify, administer, decorate, supervise, or repair activity that the worker believes cannot be honestly defended.

Meaningless Work

The book is useful because it separates two problems that are often collapsed. A bad job can be exhausting, underpaid, dangerous, or degrading while still clearly useful. A bullshit job, in Graeber's sense, may be comfortable, credentialed, and respectable while producing a private crisis of meaning.

The definition is subjective but not trivial. The point is not that an outside observer dislikes the occupation. The point is that the person inside the role experiences a gap between the public story of usefulness and the private knowledge of what the work actually does. A bullshit job is therefore a job whose occupant is required to keep performing the job's justification.

Graeber gives the phenomenon a taxonomy, which is part of why the book travelled. He sorts bullshit jobs into five types: flunkies, who exist to make superiors feel important; goons, whose work is aggressive or manipulative and exists mainly because rivals employ them too, such as lobbyists and telemarketers; duct tapers, who patch problems that should have been fixed at the source; box tickers, who let an organization claim it is doing something it is not; and taskmasters, who supervise people who need no supervision or invent new pointless work for others. The categories are useful because they locate the emptiness in a relationship to power rather than in the difficulty of the task itself.

That distinction matters for AI and labor. Automation debates often ask whether machines will replace useful work. Graeber asks a prior question: why is so much human time already trapped inside systems that require people to simulate usefulness? If the answer is status, hierarchy, paperwork, compliance theater, rent extraction, or managerial control, then automating the task may only make the underlying emptiness faster and harder to challenge.

The book's most durable insight is psychological rather than statistical. People do not only need income. They need a tolerable relation between effort, usefulness, recognition, and truth. A workplace that requires workers to pretend their activity matters when they believe it does not creates a special kind of institutional unreality. The employee is not merely tired. The employee is asked to participate in a fiction as a condition of survival.

Managerial Feudalism

Graeber's political explanation is "managerial feudalism": a world of retainers, status entourages, internal service roles, box-ticking, needless supervision, and administrative growth around power rather than production. Some of this argument overlaps with his later and earlier concern with bureaucracy in The Utopia of Rules, but Bullshit Jobs shifts the focus from forms to labor identity.

This makes the book a useful companion to Moral Mazes. Jackall shows how organizations teach managers to speak, perceive, and survive inside hierarchy. Graeber shows how workers can be absorbed into roles whose main product is institutional appearance: a report that proves oversight, a meeting that proves alignment, a dashboard that proves motion, a process that proves control.

The result is a legibility trap. A job may exist because its outputs are visible to management, procurement, auditors, investors, regulators, or other internal systems. The visible artifact becomes evidence of seriousness. The institution then needs people to produce more evidence. Meaning drains out, but documentation multiplies.

The AI-Age Reading

Read in 2026, Bullshit Jobs is a warning about AI productivity rhetoric. Many workplace AI systems promise to summarize meetings, draft emails, fill tickets, produce status updates, generate reports, score performance, write policy text, and maintain internal knowledge bases. Some of that work is genuinely helpful. Some of it is maintenance of a bureaucratic reality machine.

The danger is not only job loss. It is job preservation in a more automated and alienated form. A worker may become the person who prompts, reviews, routes, corrects, and signs off on machine-generated artifacts whose purpose was already unclear. The organization can then claim productivity gains while deepening the same problem Graeber diagnosed: human time organized around the production of believable institutional output.

As of June 19, 2026, the labor evidence does not support a simple story in which generative AI automatically abolishes pointless work. The ILO-NASK 2025 index reports that 25 percent of global employment falls within occupations potentially exposed to generative AI, with higher exposure in high-income countries, and emphasizes transformation more than full replacement. That is exactly where Graeber remains useful: the near-term question is often not whether a job disappears, but whether its least honest tasks are multiplied through software.

AI can also make pointless work more legible to power. If every meeting has a summary, every summary becomes searchable memory. If every employee activity becomes a workflow trace, every trace can become a metric. If every metric becomes a management surface, workers learn to produce the record that the system expects. The machine does not merely automate bullshit. It can make bullshit measurable, comparable, and permanent.

That is the recursive reality problem. Institutions describe work through AI systems. Workers adapt to those descriptions. The adapted behavior becomes data. The data trains future tools and informs future management decisions. A hollow process can return as evidence that the process is real.

The Automation Test

Before deploying AI into a workflow, ask four questions. What human need does the task serve? What decision, repair, care, learning, or accountability action depends on it? What evidence would show that the task helped? What rule would let the institution stop doing it?

If nobody can answer, automation is likely to preserve the ritual rather than improve the work. A meeting summary is useful when it records decisions, owners, evidence, objections, and next actions. It is pointless-work acceleration when it only makes a meeting appear accountable. Drafting email is useful when it clarifies communication. It is waste when it feeds a reply loop whose purpose is visibility. A ticket is useful when it repairs a service. It is hollow when the ticket is the artifact that lets everyone avoid the underlying failure.

A responsible automation review should therefore name the beneficiary, decision point, evidence of usefulness, labor burden, privacy cost, appeal route, and deletion rule. It should ask workers which tasks would disappear if they were allowed to tell the truth about the workflow. It should also ask who benefits when the task survives.

Governance and Safety

The governance implication is practical: workplace AI should not be assessed only at the model layer. NIST's AI Risk Management Framework Core organizes AI risk work around govern, map, measure, and manage functions. Applied to Graeber's problem, that means mapping the workflow before procurement, measuring whether the workflow advances a real objective, managing surveillance and work-intensification risks, and governing who has authority to stop a system when its outputs become theater.

Employment uses require particular care because workplace AI can convert ambiguous social judgment into automated evidence. The EEOC, DOJ, CFPB, and FTC joint statement on automated systems says existing civil-rights, consumer-protection, and fair-competition laws still apply when automated systems are used. The EEOC's iTutorGroup settlement shows the point in concrete form: application software alleged to automatically reject older applicants still triggered age-discrimination enforcement. NYC's automated employment decision tool rules add a local example of required bias audits, public summaries, and notices.

The Department of Labor's 2024 AI best-practices roadmap gives a workplace checklist that fits this review: worker input, transparency to workers, meaningful human oversight for significant employment decisions, protection of labor and employment rights, training, and worker-data security. The hard part is making those controls real. Oversight is not meaningful if the reviewer lacks time, evidence, authority, or job protection to contradict the system. Worker input is not meaningful if it cannot change procurement, metrics, or deployment decisions.

The safety risk is not that a model becomes conscious or magically powerful. It is that an institution uses fluent automation to produce more reports, scores, explanations, summaries, warnings, and performance traces while reducing the number of people who can contest whether the process should exist at all.

Where the Book Needs Friction

Bullshit Jobs is at its strongest as diagnosis and provocation. It is weaker when it moves from testimony to population-level claims. Graeber leaned partly on a 2015 YouGov poll in which 37 percent of British workers said their job did not make a meaningful contribution to the world. Magdalena Soffia, Alex J. Wood, and Brendan Burchell later tested several of Graeber's claims using European Working Conditions Survey data. Their article in Work, Employment and Society found that perceived useless work was strongly associated with poorer wellbeing, but that the prevalence they measured was much lower and declining rather than rapidly increasing.

That empirical critique matters. It prevents the book from becoming an all-purpose sneer at other people's work. Many jobs look strange from the outside because their usefulness is contextual, delayed, defensive, relational, or distributed across a larger system. A reviewer, compliance worker, support engineer, moderator, administrator, assistant, or analyst may be preventing failure that only becomes visible when the role disappears.

Graeber's subjective criterion is powerful, but it is also hard to measure. Worker testimony deserves attention because the worker sees the daily machinery. It is not automatically a population estimate, and it does not settle whether a function has hidden public value. The better use of the book is therefore not to label whole occupations as fake. It is to ask where institutions force people to maintain processes they cannot honestly justify, where metrics reward appearance over contribution, and where automation is being used to accelerate a ritual that should be redesigned or abandoned.

What This Changes

The practical lesson is simple: before automating work, interrogate its claim on human life. Does the task help someone act, understand, heal, build, learn, repair, decide, care, or contest? Or does it mostly feed a hierarchy's need for proof that something is happening?

This question belongs beside recurring concerns with labor, legibility, dashboards, classification, and institutional speech. AI makes the question sharper because it lowers the cost of producing official-looking language. A weak process can now generate stronger paperwork. A vague decision can now receive a fluent explanation. A meeting can now leave behind a polished memory that nobody fully owns.

Graeber's book remains worth reading because it refuses the default assumption that all paid activity is evidence of social need. In an AI institution, that refusal becomes a governance tool. Do not optimize the ticket queue until you know why the tickets exist. Do not automate the report until someone can say who uses it and for what. Do not celebrate productivity when the output is only more convincing performance of work.

The future of labor is not only a fight over whether machines take jobs. It is a fight over whether machines will help people do work that matters, or help institutions preserve work that survives because nobody has permission to admit what it is.

Source Discipline

This review separates book evidence, empirical evidence, current labor context, and governance sources. Simon & Schuster, Graeber's official site, LSE, and contemporary reviews support bibliographic and reception claims. YouGov and Soffia, Wood, and Burchell support the dispute over prevalence and measurement. ILO-NASK supports the current generative-AI labor exposure context. NIST, DOL, NYC DCWP, EEOC, and the multi-agency joint statement support governance and workplace-enforcement claims.

The AI sections are an application of Graeber's theory to current workplace automation. They are not a claim that Graeber predicted generative AI, that every administrative role is useless, or that any present AI system is conscious, divine, or generally intelligent.

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