Blog · Review Essay · Modified July 10, 2026 · Last reviewed July 10, 2026

Inhuman Power and the Capitalist Machine Mind

Nick Dyer-Witheford, Atle Mikkola Kjosen, and James Steinhoff's Inhuman Power is a severe little book about artificial intelligence as a political-economic machine. It does not ask whether AI will become conscious, friendly, or magical. It asks what happens when perception, prediction, classification, logistics, and decision support are absorbed into capital's machinery of production, management, and control.

For this review, AI capitalism means more than firms using models. It means data, compute, sensors, cloud platforms, management systems, labor markets, and ownership rights arranged so that machine cognition becomes a way to command work, price risk, route attention, and make institutions dependent on private infrastructure. The key claim is not that machines have moral subjectivity. It is that institutions can treat machine perception and prediction as authority.

The practical test is a cognitive-command record: what the system captures, how it classifies, what action or pressure follows, who owns each layer, who can contest it, and what human authority can stop the loop before measurement becomes management.

The Book

Inhuman Power: Artificial Intelligence and the Future of Capitalism was published by Pluto Press in June 2019 as part of the Digital Barricades series. Pluto lists the book at 224 paperback pages, with Nick Dyer-Witheford, Atle Mikkola Kjosen, and James Steinhoff as authors. The publisher describes it as an exploration of Marxist theory and AI, organized around concepts including surplus value, labor, the general conditions of production, class composition, and surplus population.

That description is accurate, but it undersells the book's force. Inhuman Power is not a general AI primer and not a neutral policy report. It is a Marxist argument about AI as capital's attempt to automate not only muscle, routine, and coordination, but cognition itself. Machine learning appears here as a new kind of productive apparatus: data, algorithms, compute, sensors, warehouses, platforms, and management systems joined into a system that can perceive, sort, predict, command, and replace.

That places the book near Platform Capitalism, Atlas of AI, and Feeding the Machine, but with a different center of gravity. Those books emphasize platforms, extraction, and hidden labor. Inhuman Power asks why cognition itself becomes an object of ownership, automation, and class strategy.

Mario Khreiche's review in the Journal of Digital Social Research usefully frames the book's twofold project: it surveys AI research and development while using Marxist theory to analyze a possible capitalist project that operates beyond and without human involvement. That is the core tension. AI is not just a tool inside capitalism; it may become one of the ways capitalism imagines freeing itself from dependence on unruly human workers, consumers, publics, and institutions.

Current Context

As of July 10, 2026, the book's strongest claim is easier to state without exaggeration: the issue is not that AI systems have become conscious or politically sovereign. The issue is that machine perception, prediction, language generation, ranking, routing, and tool use are being installed inside employment, procurement, logistics, customer service, finance, education, public administration, and platform governance.

Current labor evidence supports a mixed rather than apocalyptic reading. The ILO's 2025 exposure index estimates that one in four workers globally are in occupations with some generative-AI exposure and 3.3 percent of global employment falls in the highest exposure category, while emphasizing that most jobs still contain tasks requiring human input. The ILO's 2023 study likewise treated augmentation, work intensity, autonomy, gendered exposure, digital divides, consultation, and regulation as central issues rather than treating full job replacement as the only question.

The governance context has also moved. The EU AI Act is in force, and the European Commission's AI Act page, last updated July 7, 2026, says the May 7, 2026 political agreement on the AI Omnibus sets a new implementation timeline for high-risk systems: rules for high-risk areas including employment, education, migration, biometrics, and critical infrastructure apply from December 2, 2027, while systems integrated into products apply from August 2, 2028. The Act's prohibitions and AI-literacy duties have applied since February 2, 2025, GPAI obligations since August 2, 2025, and transparency rules are due in August 2026. The Platform Work Directive is also in force, but Member States must transpose it by December 2, 2026. That matters for this review because the relevant legal pattern is no longer abstract ethics; it is notice, documentation, human oversight, worker voice, human review, and the right to contest automated management.

The United States remains more fragmented, but existing law is not absent. In April 2023, the FTC, DOJ Civil Rights Division, CFPB, and EEOC issued a joint statement that existing civil-rights, consumer-protection, credit, and competition authorities apply when automated systems produce bias or discrimination. That statement is not a workplace-AI statute, but it undercuts the idea that a vendor model or opaque scoring system sits outside ordinary accountability.

The practical question is therefore concrete: when an institution buys machine cognition, does it also buy a system of command? The answer depends on data access, model ownership, compute dependence, task allocation, monitoring, appeal, vendor lock-in, and whether workers or affected people can negotiate the system before it becomes the workplace's common sense.

Means of Cognition

The book's best concept is "means of cognition." Industrial capitalism needed means of production: factories, machines, transport, energy systems, instruments, and workers organized around them. AI capitalism adds systems that automate perception and inference. Cameras, sensors, models, data centers, recommendation engines, logistics software, fraud systems, ad exchanges, workplace dashboards, and autonomous agents become machinery for knowing and acting.

A means of cognition is not a mind in the moral sense. It is an owned apparatus for turning traces into actionable classifications: this worker is slow, this applicant is risky, this customer is persuadable, this image is suspect, this neighborhood is profitable, this claim should be denied. The danger is not only error. It is that the right to classify becomes concentrated in systems people cannot inspect, contest, or leave.

The concept is strongest when made operational. The system first captures traces: clicks, images, location, biometrics, tickets, keystrokes, purchases, calls, logs, documents, and prior work. It then classifies those traces into scores, predictions, rankings, labels, risk flags, and recommendations. Finally, it commands through prices, schedules, eligibility, prompts, task queues, account status, search visibility, managerial dashboards, and institutional defaults. Capitalist power appears when the same owner or vendor can profit from all three steps.

The capture-classify-command chain also clarifies what has to be audited. A data-protection review that stops at collection misses command. A model audit that stops at accuracy misses capture and appeal. A labor review that looks only at layoffs misses pace, scheduling, scoring, and the quiet transfer of judgment into dashboards. The whole chain is the system.

This is why the book belongs beside media theory and cyberculture rather than only economics. AI changes what institutions can see. It changes which events become measurable, which people become predictable, which risks become actionable, and which forms of life become noise. The machine mind is not a disembodied intelligence floating above society. It is embedded in warehouses, clouds, military systems, hiring software, customer-service pipelines, ad markets, and public administration.

The phrase also helps explain why fluent chatbots are only the visible edge of the system. A conversational model may feel like AI because it speaks back, but the larger shift is quieter: cognition as infrastructure. Once perception and prediction become rentable cloud services, institutions can install model-mediated judgment almost anywhere. The office, the classroom, the clinic, the border, the shop floor, and the feed become sites where machine cognition formats reality before people argue about it.

That is why AI compute, data centers, and system documentation are not side issues. They are part of the machinery that determines who can deploy cognition at scale, who can audit it, and who must live under it.

Automating the Social Factory

Inhuman Power extends automation beyond the workplace. The authors are interested in the "social factory": the wider field of life that capital depends on and increasingly instruments. Production is not only what happens at the wage site. It also includes mobility, communication, attention, consumption, reproduction, care, education, and the constant generation of data traces.

That frame is especially useful in the AI era because training and deployment blur the old boundaries. Users generate behavioral data while socializing. Workers produce training examples while completing ordinary tasks. Customers become test subjects. Moderators absorb trauma so interfaces can look clean. Drivers, warehouse workers, call-center agents, students, patients, applicants, and creators become inputs to systems that may later discipline or replace them.

The book's political claim is blunt: AI development does not merely automate isolated tasks. It can reorganize social life around the needs of accumulation. Recommendation systems capture attention; prediction systems sort risk; logistics systems compress labor; surveillance systems make behavior visible to management; generative systems convert prior cultural labor into new output; agentic systems promise to route more activity through privately governed platforms.

The political question is therefore not only whether a job disappears. It is whether more of life becomes training data, management signal, performance metric, or dependency relation. A system can leave the worker formally employed while moving judgment, timing, evaluation, and appeal into software.

That is the social-factory version of vendor lock-in. A person can still appear free while more of their work, movement, attention, and speech is formatted for a privately owned measurement layer. Exit is not just deleting an app; it is recovering the ability to work, learn, buy, sell, receive care, or participate without passing through the same cognitive infrastructure.

Inhuman Labor

The title's "inhuman" does several jobs. It names the nonhuman machinery of AI. It names the anti-human consequences of a system that treats people as replaceable or optimizable components. It also names the possibility that capital might pursue production paths where human flourishing is irrelevant except as a constraint to be reduced.

This is where the book is strongest as a labor text. It refuses both the comforting story that AI will merely create better jobs and the clean apocalypse story that everyone is simply replaced. The more disturbing pattern is mixed: some workers are substituted, some are intensified, some are deskilled, some are newly surveilled, some are pushed into hidden data work, and some are made responsible for cleaning up machine failures without gaining authority over the system.

Current labor evidence supports that mixed reading. The International Labour Organization's 2025 update on generative AI exposure estimates that one in four workers globally are in occupations with some exposure to generative AI and that 3.3 percent of global employment falls in its highest exposure category, while stressing that most occupations still contain tasks requiring human input. Its earlier 2023 global study likewise emphasized augmentation, job quality, work intensity, autonomy, social dialogue, and regulation rather than a simple total-replacement story.

ILO Working Paper 144, published in July 2025, makes the governance implication more concrete: worker representatives are already contesting AI over employment impacts, algorithmic management practices, and working conditions in AI value chains. That turns the book's theoretical question into an institutional one. Who has standing to negotiate the system before it becomes normal?

Standing is the hinge. A worker who can only complain after a score has been used is not governing the system. A union, works council, safety committee, regulator, or public buyer needs access before deployment: what data will be collected, what model or rule will act, which decisions are affected, which human can override, and which records survive for appeal.

That mixed pattern is already recognizable. The AI interface can look post-labor while depending on annotators, moderators, miners, chip fabs, data-center technicians, prompt evaluators, software maintainers, content creators, and verification workers. Capital's dream is a smooth machine. The actual machine is a stack of people, extraction, infrastructure, and command relations made less visible by the fluency of the output.

Against Acceleration

Inhuman Power is explicitly hostile to left accelerationism and to any politics that imagines capitalist technical development can simply be pushed faster and then repurposed cleanly. The authors' argument is that AI is not a neutral engine waiting for a better operator. It is being shaped inside ownership structures, data regimes, military investment, platform monopolies, labor markets, and fantasies of optimization.

That does not mean the book is anti-technology in a simple sense. It is anti-fatalism. It asks readers to stop treating AI capability as a free-standing historical force and start asking what kind of social relation each capability requires. Who owns the data? Who pays for the compute? Who sets the objective? Who is watched? Who verifies the output? Who is displaced? Who can refuse? Who benefits when the model appears to understand?

The accelerationist temptation is powerful because AI gives technical form to an old dream: enough computation might dissolve political conflict into optimization. The book's counterpoint is that computation often intensifies conflict by hiding it inside infrastructure. A model can allocate, recommend, classify, rank, and command without making its political theory explicit.

Governance and Safety

The safety implication is that AI capitalism should be governed at the points where cognition becomes command. Model risk, labor risk, and platform risk converge when a system can monitor workers, allocate tasks, set pace, classify applicants, trigger discipline, deny access, or make customers and citizens visible to automated management.

That frame is now visible in enacted law and official implementation planning, even where duties are still phasing in. The EU AI Act's Annex III covers specified AI systems used for recruitment, work-related decisions, task allocation, monitoring, performance evaluation, and access to self-employment. Article 26 sets deployer duties for high-risk systems, including appropriate human oversight and notice to natural persons when Annex III systems assist decisions about them. Article 50 sets transparency duties for direct AI interaction and certain synthetic or biometric/emotion-recognition uses. The Commission's July 2026 AI Act implementation page says AI Omnibus changes move the rules for high-risk areas including employment to December 2, 2027, and product-integrated high-risk systems to August 2, 2028; transparency rules remain due in August 2026.

The Platform Work Directive puts algorithmic management even more directly into labor governance, with national transposition due by December 2, 2026. It requires information about automated monitoring and decision-making systems, regular evaluation with worker-representative involvement, sufficient human resources for oversight, competent overseers with authority to override automated decisions, human decisions for account restriction or termination, rights to explanations and human review, and safety-and-health risk assessment for automated systems that may affect accidents, psychosocial risk, or ergonomic risk.

NIST's AI Risk Management Framework and its Generative AI Profile point in the same operational direction: risk management has to follow the lifecycle, not only the launch demo. The U.S. Department of Labor's 2024 AI workplace roadmap is useful as an archived worker-centered reference rather than binding or necessarily current policy; the page itself warns that some news-release information may be out of date after January 20, 2025. Its relevant contribution is still the checklist logic: governance structures, meaningful human oversight for significant employment decisions, worker transparency and input, labor rights, training, and worker-data protection.

U.S. federal enforcement is piecemeal but not blank. The 2023 joint agency statement from the FTC, DOJ Civil Rights Division, CFPB, and EEOC identifies familiar automated-system failure modes: unreliable or non-representative data, opaque models, discriminatory outcomes, and claims that automation makes a decision objective. For this book's purposes, the important point is narrow: capital cannot launder a consequential classification merely by routing it through software.

The practical test is whether a system can be audited backward into data, compute, labor, ownership, subcontractors, model versions, prompts, policies, and logs, and forward into appeal, human authority, incident response, and exit rights. For each deployment, the control record should name the decision point, affected people, data inputs, vendor dependencies, objective function, oversight role, override authority, worker or public consultation, appeal path, shutdown criteria, and incident owner.

A serious control record should also separate assistance from assessment. If a copilot helps a worker write, summarize, code, route tickets, answer calls, or plan tasks, its traces should not quietly become a performance dossier without fresh notice, validation, worker consultation, and appeal. Otherwise the helpful interface becomes the measurement arm of management.

Where the Book Needs Friction

The book's severity is also its limitation. It is short, polemical, and written from a declared Marxist position. Readers looking for a balanced tour of beneficial AI applications, technical architectures, model evaluation, or mainstream governance proposals will need companion sources. The book's job is not to summarize the whole field. Its job is to make one hard argument difficult to ignore.

Some of its sharpest formulations can feel over-totalizing. Capital is not the only force shaping AI: scientific curiosity, public research, disability access, medical use, education, creative practice, open-source communities, safety work, and democratic regulation all matter. A serious reading should preserve those distinctions rather than flatten every deployment into the same story.

The better critique is not that every AI use is capitalist domination in miniature. It is that every consequential deployment has a political economy: funding, data rights, labor conditions, infrastructure dependence, appeal rights, and institutional authority. Public-interest uses are real, but they still need records proving that public benefit is not being bought with hidden extraction or unappealable control.

Still, the book's pessimism has aged better than many smoother accounts from the same period. Since 2019, AI has become more visibly entangled with cloud concentration, copyright conflict, data-center buildouts, workplace monitoring, military procurement, platform dependency, data labor, and speculative AGI rhetoric. The details have changed, but the direction of the warning remains legible.

What This Changes

The review's central use is to break the spell of the friendly interface. AI products often arrive as helpful surfaces: answer boxes, copilots, assistants, tutors, agents, dashboards, companions. Inhuman Power asks what these surfaces connect to. A conversational system may feel personal, but behind it sits an institutional machine for capturing data, concentrating capability, reorganizing labor, and renting cognition back to the world.

This matters for belief formation as much as for labor. A society that treats machine cognition as neutral infrastructure will start building institutions around its outputs. Applicants will be screened by it, workers measured by it, students tutored by it, publics persuaded by it, and managers reassured by it. The model becomes a way of seeing, and the way of seeing becomes a way of governing.

The strongest lesson is not that AI must be rejected. It is that every AI system needs a political anatomy. Follow the output backward into labor, data, compute, ownership, energy, surveillance, and dependency. Follow it forward into deskilling, authority, appeal, and institutional memory. In procurement terms, start with an AI bill of materials, a data sheet supply-chain map, an AI system inventory, and explicit rules for algorithmic management, human oversight, recourse, and vendor governance. Intelligence is never just intelligence once it becomes infrastructure.

Source Discipline

Current legal and labor claims were rechecked on July 10, 2026. This review separates the book's Marxist thesis from the evidence used to update it. Pluto Press and BiblioVault establish bibliographic facts, contents, and publisher framing. Scholarly reviews help locate the book's reception, but they do not independently prove every political-economic claim. ILO sources support exposure and social-dialogue claims. EUR-Lex supports legal text. The European Commission's AI Act page supports the current post-Omnibus implementation timeline. NIST supports voluntary risk-management language. The U.S. Department of Labor source is cited as an October 2024 worker-centered roadmap with an explicit current-policy caveat, not as binding law. The 2023 U.S. joint agency statement is enforcement posture under existing authorities, not a new statute.

Use the sources by type: publisher pages for metadata, scholarly reviews for reception, labor agencies for exposure and workplace evidence, regulators for legal duties, and standards bodies for operational risk language. The argument is strongest when those categories are not mixed into one general aura of authority.

The bounded claim is therefore narrower than the book's most apocalyptic rhetoric. AI does not need to become conscious, divine, or generally intelligent to shift power. The already consequential issue is that machine perception, prediction, and language generation are being installed inside work, infrastructure, markets, public administration, and platform governance. That is enough to require labor documentation, vendor disclosure, worker voice, human oversight with real authority, and appeal 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