Blog · Review Essay · Last reviewed June 25, 2026

Resisting AI and the Politics of Refusal

Dan McQuillan's Resisting AI is not another plea for fairer dashboards or friendlier model documentation. It asks a harder question: what if some automated systems should be refused because they extend the institutional habits that already make people sortable, suspicious, cheap, and governable?

In this review, AI refusal means a documented decision not to build, buy, deploy, continue, or expand an AI system because its institutional purpose, power shift, data dependency, or contestability failure cannot be made acceptable. Refusal is not technophobia. It is a governance threshold with reasons, evidence, alternatives, and a public record.

The threshold has five tests: lawful purpose, necessity, proportionality, contestability, and exit. If an institution cannot pass those tests without shifting burden onto the people being classified, the answer should be no, delay, narrow, or retire.

The useful version of refusal is procedural and political at once. It names the use case, identifies who would be governed by the system, tests whether a less automated alternative exists, and leaves an auditable reason why an institution chose not to automate a power it could not justify.

The Book

Resisting AI: An Anti-fascist Approach to Artificial Intelligence was published by Bristol University Press/Policy Press in 2022. The publisher lists the paperback at 190 pages with ISBN 978-1529213508 and publication dates of August 30, 2022 and July 15, 2022 across formats. Oxford Academic lists the Policy Press Scholarship Online record with DOI 10.1332/policypress/9781529213492.001.0001, online ISBN 9781529213539, print ISBN 9781529213492, and a July 15, 2022 publication date.

McQuillan is a lecturer in creative and social computing at Goldsmiths, University of London. The publisher biography is relevant because the book does not come from a narrow technical-audit tradition: it draws on physics, computing, care work, mental-health advocacy, asylum information projects, citizen science, Amnesty International, data justice, and political theory.

The book appeared just before the ChatGPT boom made generative AI a daily interface. That timing makes it more useful, not less. Its core target is not one chatbot product or one model family. It is the social logic of deep learning when it is adopted by institutions that already prefer measurable proxies, austerity administration, security theater, predictive suspicion, and optimization without accountability.

Current Context

As of June 25, 2026, refusal has moved from protest vocabulary into compliance, procurement, and safety review. The EU AI Act prohibits specified practices outright, classifies many biometric, education, employment, essential-service, law-enforcement, migration, and justice systems as high risk where the regulation applies, and requires covered deployers of high-risk systems to follow instructions for use, assign competent human oversight, monitor operation, keep logs under their control, and conduct fundamental-rights assessment in relevant settings. Article 86 also adds a scoped right to clear and meaningful explanations for certain Annex III high-risk AI decisions that produce legal or similarly significant adverse effects. In the United States, OMB's 2025 civilian-agency AI memoranda require covered agencies to identify high-impact AI uses, complete impact assessments before deployment, manage risk, report publicly, and write acquisition terms that address testing, monitoring, portability, government data use, and vendor lock-in.

The practical change is that "Should we use AI here?" cannot be collapsed into "Can this model be made safer?" Some systems should be blocked, narrowed, delayed, piloted only with independent evaluation, replaced with a non-AI workflow, or ended after post-deployment evidence. McQuillan's book is valuable because it gives that decision a political name, but the governance work is more exacting: identify the authority being automated, the people being exposed, the evidence being relied on, and the institution that must remain answerable.

That also makes refusal a lifecycle control, not only a launch veto. A system can become unacceptable after a vendor update, data-source expansion, new use case, changed legal authority, model drift, incident pattern, labor impact, or evidence that the affected people cannot meaningfully appeal. In that sense, refusal belongs beside procurement, inventory, audit trails, post-market monitoring, incident reporting, and decommissioning.

The Machine as Institution

Resisting AI is strongest when it treats AI as a political arrangement rather than a neutral capability. McQuillan does not begin from the usual question, "How accurate is the model?" He begins from the setting that gives the model power: the welfare office, border system, workplace, school, police program, health triage workflow, content platform, or procurement process that wants a technical object to make hard social decisions appear manageable.

This aligns the book with Automating Inequality, Race After Technology, Seeing Like a State, and The Tyranny of Metrics. In each case, the danger is not only that a tool makes an error. The deeper danger is that an institution builds a world in which the tool's categories become the terms on which people must survive.

A model can turn old records into a prediction. An agency can turn the prediction into a queue. A manager can turn the queue into a performance target. A person can then be forced to behave inside a category made from someone else's data. The loop is political before it is technical. It decides whose reality is considered evidence, whose account is treated as noise, and who gets a usable path to appeal.

That is why refusal has to name the whole deployment, not just the model. The object to evaluate is the model plus the institution, data source, policy rule, worker role, vendor contract, appeal path, monitoring plan, and decommissioning route. A model that is tolerable in research can become intolerable when embedded in a benefits denial, hiring screen, school discipline system, or border workflow.

A refusal review therefore starts with the institutional duty being changed. If a public office owes due process, service access, nondiscrimination, records retention, and human review, the AI system has to support those duties rather than route around them. If a company owes workers safety, fair pay, schedule predictability, and a meaningful grievance path, a scoring system that makes management less accountable should fail the review even when the model benchmark looks respectable.

The recurring test is authority, burden, and repair. Who gains authority from the system? Who carries the burden of its mistakes, surveillance, or proof demands? Who has practical power to repair the record when the system is wrong? If the answers are hidden in vendor terms, dashboards, or managerial discretion, the deployment has not earned trust.

Optimization and Bureaucracy

One of the book's useful moves is to connect AI to bureaucracy rather than treating it as a break from bureaucracy. The LSE review by Milan Stürmer and Mark Carrigan emphasizes this point: McQuillan reads AI less as a science-fiction rupture than as an upgrade to existing administrative order. That is the right lens for many real deployments.

Optimization sounds technical, but in institutions it often means choosing which value will be maximized and which costs will be made external. A benefits system may optimize for fraud reduction and push risk onto families. A workplace platform may optimize for throughput and push instability onto workers. A school may optimize for detection and push suspicion onto students. A border system may optimize for risk and turn a life story into a machine-readable case file.

This is why "AI ethics" can become too narrow. Bias audits, transparency statements, and fairness metrics can help in some cases, but they cannot by themselves answer whether a system should exist, whether the institution using it deserves the power it gains, or whether the people affected can actually contest the output. A more legible cage is still a cage.

The bureaucratic question is therefore prior to the technical one: what decision is being accelerated, and who loses discretion when the workflow hardens? A refusal process should force the agency or company to say whether it is improving service, rationing scarcity, shifting liability to a vendor, disciplining workers, or manufacturing the appearance of objectivity.

The Labor and Matter of AI

McQuillan's chapter abstracts make clear that his account of AI includes data dependence, neural-network opacity, carbon emissions, centralization of control, underpaid human labor, and the anti-worker politics that can accompany automation. That places the book beside Atlas of AI, The Eye of the Master, and Ghost Work.

The point is not that AI is fake because workers are involved. The point is that the public story of automation routinely erases the workers, data subjects, maintainers, moderators, labelers, miners, and energy systems that make automation possible. A system marketed as intelligence can function as a redistribution machine: attention, risk, exhaustion, and environmental cost move downward and outward while authority moves upward into platforms and agencies.

That matters for human-machine cognition. A model is not just a mathematical artifact sitting apart from society. It is a way of organizing perception and action across many people and machines. When the organizational design is extractive, the resulting intelligence inherits that design. It can learn from care while making care workers disposable. It can learn from public culture while enclosing the benefits. It can learn from collective activity while giving the collective no governing role.

Refusal is strongest when it follows that material chain. A buyer should ask whether the system depends on contested training data, hidden annotation labor, workplace surveillance, environmental costs shifted to another community, or a cloud dependency that the public cannot audit or exit. If those dependencies cannot be governed, the system is not merely incomplete. It is asking for authority while withholding the conditions of accountability.

The Force of the Title

The title is easy to caricature. The useful reading is not that every statistical model is fascist. It is that AI, as currently organized, can mesh dangerously with authoritarian, eugenic, and austerity-driven habits: ranking lives, normalizing exception, sorting populations, treating care as cost, and presenting exclusion as objective necessity.

This is where the book's polemical force helps. Many AI debates stay inside reform language: make the dataset better, add a human in the loop, improve explainability, publish a principle, audit the vendor. McQuillan presses on the prior question. If an institution is using AI to intensify scarcity, automate suspicion, or evade responsibility, then the problem is not a missing fairness patch. The problem is the project.

That does not mean the strongest argument is always the most maximal one. The word "fascist" should remain tied to concrete mechanisms: emergency powers, hierarchy, racialized and ableist valuation, forced legibility, policing, border control, labor discipline, and the shrinking of democratic space. Used carefully, the term names a pattern. Used lazily, it can flatten important differences between systems and weaken the case for refusal.

Refusal as Governance

The book's most valuable contribution is its insistence that refusal is not the absence of governance. Refusal can be governance. A community can say that a predictive policing system should not be purchased. Workers can reject an algorithmic management system that turns the job into surveillance. Patients can resist a triage system that hides rationing behind a score. Public agencies can decide that some forms of automated eligibility, risk assessment, or biometric sorting are incompatible with democratic service.

McQuillan's alternatives include workers' councils, people's councils, mutual aid, commons, feminist care, decolonial knowledge, and forms of "decomputing" that constrain the reach of automation. These proposals will frustrate readers looking for a procurement checklist. They are not vendor-neutral best practices. They are a demand that affected people gain power over whether a system is built, bought, deployed, repaired, or shut down.

That demand belongs in AI governance because otherwise governance becomes a paperwork layer around inevitability. If the only choices are "deploy now" or "deploy after compliance review," the public has already lost the most important decision. A real governance regime must preserve the word no, but it should also preserve the record of why no was the right answer.

A mature refusal process should distinguish five outcomes. Prohibited means the system falls inside a legal or institutional ban. Unnecessary means a less intrusive non-AI workflow can meet the legitimate purpose. Unready means evidence, logs, human oversight, accessibility, or appeal is not yet sufficient. Conditional means the system may proceed only under a narrower scope, pilot, fallback, or independent evaluation. Retired means post-deployment evidence has shown the system should stop.

As of June 25, 2026, this is no longer only an activist vocabulary. The EU AI Act has been in force since August 1, 2024, and its prohibited-practice rules and AI-literacy obligations have applied from February 2, 2025. Article 5 bars several practices that belong near McQuillan's refusal line, including social scoring, vulnerability exploitation that causes or is reasonably likely to cause significant harm, untargeted scraping to create or expand facial-recognition databases, certain criminal-risk assessments based solely on profiling, workplace and education emotion inference except for medical or safety reasons, and biometric categorization to infer protected traits. Annex III also treats many uses in education, employment, essential public services and benefits, law enforcement, migration, and justice as high-risk where the regulation applies; Article 27 adds a fundamental-rights impact assessment duty for specified deployers of high-risk systems where and when that obligation applies.

U.S. federal policy has a narrower but practical version of the same point. OMB M-25-21 requires federal agencies to identify high-impact AI uses, implement minimum risk-management practices, complete AI impact assessments before deployment, and safely discontinue high-impact AI functionality that does not comply. OMB M-25-22 adds acquisition controls for testing, monitoring, portability, government data use, and vendor lock-in. These rules do not enact McQuillan's politics. They do confirm the governance grammar: some systems may be allowed, conditioned, piloted, redesigned, paused, or refused.

Refusal also has gradations. The decision may be do not build, do not buy, do not deploy in this domain, do not automate this decision, require a non-AI fallback, require affected-party opt-out, require worker or community review before expansion, suspend pending evidence, or decommission after harm. Treating all of those as one dramatic "no" makes the politics less useful. The hard part is matching the refusal to the power being exercised.

The safety issue is administrative momentum. Once a system becomes case intake, queue ranking, eligibility triage, worker scoring, or enforcement recommendation, the burden of proof often shifts onto the affected person. A refusal threshold should trigger when data provenance is weak, appeal is ceremonial, the purpose is rationing or surveillance, protected-class effects cannot be measured or mitigated, the vendor blocks audit access, or a less intrusive alternative can meet the same legitimate need.

The Refusal Record

The concrete artifact is a refusal record. It should state the proposed system, affected people, decision domain, institutional purpose, legal basis, data provenance, labor and environmental dependencies, foreseeable harms, available alternatives, appeal path, consultation record, decision owner, and reconsideration date. It should also name the less automated option. A refused AI system should not leave only silence; it should leave evidence that the institution chose not to automate a power it could not justify.

That record should connect to the same evidence spine as an approved system: the procurement file, AI system inventory, impact assessment, audit-trail design, data-retention plan, vendor terms, incident review, and change-management history. Otherwise refusal becomes a private judgment instead of institutional memory. The next team can quietly buy the same system under a new name.

Where the Book Needs Friction

The book's strength is also its weakness. It is an urgent political argument, and urgency can compress technical variation. Predictive scoring, recommender systems, facial recognition, fraud detection, generative models, robotic systems, and decision-support tools do not all work the same way or create the same harms. Serious resistance needs specificity.

The LSE reviewers make a related criticism: they credit the book for raising serious concerns about bias and misuse, but argue that it lacks enough detailed technical analysis to support some of its social claims. That critique should be taken seriously. The more severe the charge, the more important it is to show the mechanism, the deployment path, the affected people, and the institutional incentives.

Still, the answer is not to retreat into technical modesty. The better reading is to pair McQuillan's political theory of refusal with the evidentiary discipline of books like AI Snake Oil, the administrative realism of Recoding America, and the power analysis of The Tech Coup. Refusal is strongest when it can name exactly what is being refused and why.

Refusal also needs its own accountability. A ban, moratorium, or procurement veto can be wise; it can also be symbolic, captured, or used to protect incumbents. The standard should be public reasons, affected-party participation, evidence of alternatives, periodic review, and a path to revisit the decision if the institution, technical design, or material conditions change.

That caveat matters because refusal can become anti-evidence theater if it performs moral seriousness while leaving the underlying institution untouched. A benefits agency that rejects one vendor while preserving punitive eligibility rules has not solved the political problem. A school that bans a detector while keeping a culture of suspicion has not solved the educational one. The refusal has to change the workflow, not just the procurement outcome.

There is a second failure mode: refusal can protect the institution from scrutiny. An agency may reject "AI" while continuing the same harmful sorting through rules engines, manual spreadsheets, contractor discretion, or opaque dashboards. The honest question is not whether a vendor calls the system AI. It is whether people are still being classified, burdened, or denied through a process they cannot inspect or contest.

The AI-Age Reading

Read after the generative-AI boom, Resisting AI becomes a warning about agentic bureaucracy. Chatbots and copilots enter organizations as helpful surfaces: summarizers, triage assistants, writing tools, coding agents, tutors, casework aids, customer-service bots, and policy interpreters. The interface may feel conversational, but the system often inherits the organization's incentives.

If a benefits office is underfunded, an AI assistant may make scarcity look smoother. If a company is trying to deskill support labor, a chatbot may make degraded service look innovative. If a school does not trust students, an AI detector may make suspicion feel scientific. If a platform wants cheaper moderation, a model may hide the remaining human cost. The danger is not that machines become evil. The danger is that institutions become more effective at bad habits.

The practical lesson is simple: ask what power the system extends. Does it give affected people more voice, appeal, context, and collective leverage? Or does it make them more measurable, more replaceable, more governable, and easier to ignore? Does the model open a decision to democratic contest, or does it turn contest into an exception-handling ticket?

For generative and agentic systems, the refusal test should include handoff. A chatbot that only answers may still mislead, but an agent that files forms, updates records, sends notices, ranks cases, or recommends enforcement changes the evidence trail. The more a system can act, the more it needs identity, logging, tool permissions, human authority, incident reporting, appeal, and shutdown criteria before deployment.

Refusal is also a way to defend human capacity. If an AI system weakens worker expertise, closes the channel for affected-person testimony, or replaces public service with deflection metrics, it may pass narrow accuracy tests while damaging the institution's ability to know what is happening. A system that erodes the feedback loop needed for repair should be treated as unsafe even before it produces a headline incident.

McQuillan's book is not the final word on AI. It is a necessary irritant. It refuses the assumption that every social problem should be translated into prediction, every public function into optimization, every worker into a metric source, and every democratic disagreement into a technical implementation issue. In an AI culture addicted to capability talk, that refusal is a form of realism.

Source Discipline

This review separates book metadata, McQuillan's political argument, secondary criticism, current legal and policy context, and this site's interpretation. Bristol University Press and Oxford Academic support the bibliographic claims. McQuillan's chapter abstracts support the summary of the book's internal structure. LSE and International Journal of Communication sources support the review context. EUR-Lex, the European Commission, the EU AI Act Service Desk, OMB, and NIST support current governance claims checked on June 25, 2026.

Legal claims are jurisdiction-specific and time-sensitive. The EU AI Act, OMB M-25-21, OMB M-25-22, and the NIST AI Risk Management Framework are different kinds of authority: binding EU regulation, U.S. federal-agency policy, procurement guidance, and voluntary risk-management framework. A prohibited AI practice is not the same as a high-risk classification; an impact assessment is not community consent; a vendor test is not a democratic mandate; and a voluntary risk framework is not an enforceable right.

Evidence claims should also distinguish refusal from rejection rhetoric. A serious refusal cites the use case, version, affected population, data source, authority, risk evidence, consultation record, alternatives, and review date. A general declaration that "AI is harmful" may be politically important, but it cannot substitute for a record that lets later reviewers see what was actually refused and why.

The word "anti-fascist" is treated here as McQuillan's framing and the publisher's subtitle, not as a claim that every AI system is fascist. Strong refusal arguments should identify concrete mechanisms: profiling, coercive legibility, emergency logic, racialized or ableist valuation, workplace discipline, border control, welfare suspicion, loss of appeal, or extraction without governance. This page makes no claim that any AI system is conscious, divine, or AGI.

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