Blog · Review Essay · Modified August 12, 2026 · Last reviewed August 12, 2026

Artificial Whiteness and the Ideology Called AI

Yarden Katz's Artificial Whiteness is not a standard book about biased algorithms. It is a critique of artificial intelligence as a flexible institutional idea: a label that gathers funding, expertise, military ambition, corporate futurism, carceral reform, and racialized models of knowledge into a package that can be sold as technical necessity.

For this review, AI ideology is institutional coordination through an elastic label: unlike systems are bundled into one historical force while stories, metrics, budgets, expert roles, procurement categories, and interface claims make adoption appear to be a neutral technical transition. Its signature move is not mere exaggeration. It converts contested authority into presumed inevitability.

The practical test is a label warrant with three separate records: what the system can demonstrably do, what authority or resources the AI name unlocks, and who can reject the project rather than merely tune it. Accuracy cannot answer the latter two questions.

The Book

Artificial Whiteness: Politics and Ideology in Artificial Intelligence was published by Columbia University Press in November 2020. The publisher lists the paperback ISBN as 9780231194914, the hardcover ISBN as 9780231194907, the e-book ISBN as 9780231551076, and all three editions at 352 pages. JSTOR's book record shows three parts—formation, self and the social order, and alternatives—with chapters on empire, capital, epistemic forgeries, carceral-positive reform, artificial whiteness, dissenting visions, and refusal.

Katz's University of Michigan profile lists him in American Culture and Digital Studies and identifies imperialism, white supremacy, racial capitalism, science and technology studies, eugenics, exploited scientific labor, and radical social movements among his fields. That location matters. This is a book about AI history, but also about the institutions that make some definitions of intelligence useful, fundable, and politically convenient.

The review belongs beside Atlas of AI, Race After Technology, Algorithms of Oppression, Dark Matters, Surveillance Valley, and The Cultural Logic of Computation. Those books ask how computation becomes power. Katz presses harder on the premise that artificial intelligence is a coherent technical destiny at all.

Current Context

As of August 12, 2026, Katz's question is not whether governance exists but what its categories permit. The AI label now organizes procurement, compliance, research funding, public services, school policy, workplace monitoring, military planning, and vendor roadmaps. “AI governance” can name democratic control over consequential systems. It can also become a professional layer that leaves an institution's purpose intact while adding audits, dashboards, and risk language.

The U.S. record contains both possibilities. NIST Special Publication 1270 treats AI bias as sociotechnical rather than merely statistical. The voluntary AI Risk Management Framework 1.0 is being revised, but its current Manage 1.1 outcome already asks whether development or deployment should proceed. A 2023 FTC, DOJ Civil Rights Division, CFPB, and EEOC statement says existing civil-rights, consumer-protection, fair-lending, and employment authorities can reach automated systems, while expressly noting that the statement is informational, creates no new rights or obligations, and is not final agency action.

The European Union now has three different status layers that should not be collapsed. Prohibitions and AI-literacy duties have applied since February 2025; general-purpose-model duties have applied since August 2025, and the Commission's enforcement powers for those duties began August 2, 2026. Article 50 transparency obligations also began applying August 2, 2026. By contrast, binding Regulation (EU) 2026/1744, in force since July 27, delayed defined high-risk duties to December 2, 2027 for stand-alone systems and August 2, 2028 for systems embedded in regulated products. A duty in force, a duty delayed, and nonbinding implementation guidance are different kinds of evidence.

Federal procurement shows the label doing work directly. OMB M-25-22 uses AI acquisition to demand documentation, testing, monitoring, portability, pricing transparency, and protection against vendor lock-in. OMB M-26-04 turns “truth-seeking” and “ideological neutrality” into contract requirements and disclosure thresholds for agency-procured large language models. Its scope is bounded: it generally excludes national-security systems and sunsets December 11, 2027 unless replaced. It does not supply a politics-free benchmark for truth or neutrality.

These controls strengthen Katz's critique when they preserve a prior decision. A framework can identify harm and still leave room not to proceed; a procurement rule can reduce lock-in while normalizing acquisition; a fundamental-rights assessment can support refusal or convert refusal into a managed stakeholder concern. The decisive test is whether evidence can change the institution's purpose, budget, workflow, or decision to deploy—not whether the paperwork is complete.

The AI Label

The strongest move in Artificial Whiteness is to treat “AI” as a historical label with political work to do. Katz is not saying that software, machine learning, statistics, robotics, and neural networks are imaginary. The point is sharper: the label has repeatedly expanded, narrowed, disappeared, and returned as funders, universities, firms, the military, and professional experts reorganized around it. Technical discontinuity can coexist with institutional continuity.

That makes the question less mystical. Instead of asking only whether a system is truly intelligent, ask what becomes easier once it is called AI. A research program can attract defense money. A company can rebrand data extraction as futurism. A university center can become a policy authority. A policing system can present itself as objective analysis. A welfare tool can appear modern rather than punitive. A workplace dashboard can become an intelligent manager rather than a managerial choice.

This is why Katz is useful in the present moment. The AI label does not merely describe a capability. It can authorize a relationship. It lets institutions say the machine has arrived, society must adapt, experts must manage the transition, and old political arguments are now technical implementation problems.

That argument pairs naturally with AI Snake Oil, but it is aimed at a deeper layer. Narayanan and Kapoor ask what AI systems have actually proved. Katz asks why so many institutions want the label to carry authority before proof, and why the label survives even when its technical content keeps changing.

A practical reading begins with a label warrant. The capability warrant describes the deployed function, version, test conditions, uncertainty, and comparison with a non-AI baseline. The authority warrant records the power, budget, procurement category, expert status, or legal duty the label unlocks. The refusal warrant names who can reject, shrink, sunset, or replace the system if the evidence fails or the institutional purpose is wrong. A useful AI system inventory should store the functional description alongside the marketed label; otherwise a vendor's category becomes the institution's map. Without all three warrants, “AI” is doing ideological work before it is doing accountable technical work.

Whiteness as Method

The title's difficult term is doing analytical work. Katz is not using whiteness as demographic shorthand for individual engineers, or claiming that every system behaves identically. The book treats whiteness as an organizing logic: a historically situated, racialized, gendered, imperial, and capitalist standpoint presents itself as universal, neutral, placeless, and entitled to rule. “Artificial” matters because that authority is made and maintained, not because its consequences are unreal.

This maps onto AI in concrete ways. A benchmark can be treated as a universal test of intelligence even when it encodes narrow institutional priorities. A game-playing system can be treated as evidence about thought in general. A facial-recognition improvement can be framed as inclusion while extending the reach of biometric control. A prediction system can claim neutrality while turning old records into future suspicion. A model can appear to know autonomously while hiding the people, data, objectives, funding, and deployment setting that made its output possible.

The three “epistemic forgeries” make that argument precise. First, a situated system is presented as universal intelligence outside social context. Second, human thought is reduced to performance in controlled tasks so that a machine can appear to rival thought as such. Third, machine knowledge is presented as autonomous, hiding the developers, data, funders, and institutions that established the conditions of knowing. These are not simply false outputs. They are false accounts of where authority came from.

That is the book's best contribution to AI criticism. Bias is not only a defect inside a model. Bias can be a function of the institutional project that asked the model to exist.

This matters for current debates over unbiased AI because neutrality can itself become an ideological claim. A system's answer depends on source selection, benchmark design, prompt policy, model training, post-training rules, retrieval ranking, refusal behavior, and the cost assigned to different errors. A governance process that announces neutrality without exposing those choices is not outside politics. It has hidden the politics in the measurement stack.

The governance consequence is sharp: a benchmark, evaluation, or audit cannot stand in for a theory of harm. A high score on a narrow task may support a capability claim. It does not prove that the task is legitimate, that the dataset represents the affected world, or that the institution has earned authority over the people it classifies. This is the same legibility problem examined in Seeing Like a State: simplification becomes dangerous when an administrative representation gains power to reorganize the reality it omitted.

Carceral-Positive Logic

The chapter on carceral-positive logic is the hinge of the book. Katz argues that some critical AI work can end up improving the legitimacy of harmful systems by making them more technically refined. A facial-recognition system that performs more evenly across demographic groups may still expand surveillance. A risk score with better calibration may still make cages, raids, watchlists, and deprivation look like neutral administrative outputs. A fairness audit may turn an abolition question into a vendor remediation ticket.

This is not an argument against measuring harms. It is an argument against letting measurement define the moral horizon. If the institution is doing violent work, making the classifier more accurate may make the violence more durable. If the problem is that police, prisons, borders, landlords, employers, insurers, or schools have too much unaccountable power, then the ethical question cannot stop at whether the software is less biased than its previous version.

The Information & Culture review highlights Katz's attention to the Stop LAPD Spying Coalition, whose organizing against Los Angeles predictive-policing systems becomes an example of community research refusing the premises of data-driven policing. WIRED's reporting on Operation LASER and PredPol gives the operational context: historical police data, scoring, hotspots, and targeted attention can route more policing toward communities already exposed to police contact.

That is the recursive danger. The system records enforcement contact, interprets the record as risk, sends more attention, creates more records, and then treats the new trail as independent evidence. Calling the loop AI makes it feel as if intelligence discovered danger. In practice, an institution may have automated its own suspicion while changing the data-generating process it later uses to validate the system.

The Los Angeles example should be read historically and structurally. The LAPD Office of the Inspector General's 2019 review treated Operation LASER and PredPol as data-driven policing programs and documented oversight and data-quality concerns; the National Academies later summarized that review as finding weak evidence that LASER reduced crime and noted civil-rights concerns around the program. The durable lesson is not that every jurisdiction uses the same tool. It is that a reform vocabulary can keep the carceral premise alive unless affected communities have power over whether the system should exist.

The control implication is measurable. A review should distinguish reports of harm from records of enforcement, document where labels originated, compare the system with a nondeployment baseline, measure added surveillance exposure as well as predictive error, and stop when deployment changes the input distribution faster than reviewers can disentangle prediction from police activity. For policing, prisons, borders, and biometric surveillance, non-use is not a philosophical extra. It is a safety control. A system that can only be tuned, audited, or made more demographically even, but never withdrawn from a coercive workflow, has already excluded the people most likely to bear the risk.

Recursive Reality

Artificial Whiteness is especially useful for thinking about recursive reality because it shows how a category becomes infrastructure. "AI" starts as a research label, becomes a funding magnet, becomes an expert identity, becomes a policy object, becomes a procurement category, becomes a public fear, becomes an ethics industry, and then becomes evidence that society must organize around AI.

The loop is not only discursive. It changes budgets, careers, conferences, standards, grant programs, vendor roadmaps, police tools, university centers, classroom assignments, military planning, and public vocabulary. Once an institution has an AI office, AI strategy, AI committee, AI procurement line, and AI ethics policy, the premise has already won a great deal. The world has been rearranged so that AI appears to be the thing everyone must respond to.

That pattern should be familiar from other machine-readable systems. A ranking creates ranking behavior. A dashboard creates dashboard work. A benchmark creates benchmark training. A risk model creates risk records. An answer engine creates source behavior around answer extraction. A label creates the institution that then proves the label was real.

This is category lock-in: budgets, offices, contracts, careers, and compliance duties accumulate around a label until renaming or rejecting the project becomes institutionally expensive. Once “trustworthy AI,” “high-risk AI,” “unbiased AI,” or “AI readiness” becomes the official frame, organizations produce artifacts that satisfy it: inventories, model cards, procurement files, impact assessments, audits, and training modules. Those artifacts can be valuable. They can also stabilize the premise that the institution's real question is how to manage AI, rather than whether this classification, surveillance, or automation project should exist.

A public AI register can resist that recursion only if it records more than systems adopted. It should also preserve rejected proposals, non-AI alternatives, reasons for nondeployment, expired warrants, serious incidents, and withdrawals. Otherwise the register remembers adoption and forgets refusal, producing a history in which automation always looked like the available path.

Katz adds a harder political question: who benefits when the label becomes real, and who loses the ability to name the underlying institution?

The AI Reading

Read in 2026, Artificial Whiteness is a warning about the governance language around foundation models, agents, and automated decision systems. The phrase "AI governance" can mean democratic control over technical systems. It can also become a way for technical experts, vendors, consultants, universities, and state agencies to professionalize the management of systems whose deeper institutional purposes remain untouched.

The problem is visible whenever a product turns a political choice into an AI-readiness question. Should a school surveil students? Should a court score defendants? Should a welfare office automate suspicion? Should a workplace instrument every keystroke? Should a city fuse cameras, license-plate readers, emergency calls, and predictive maps? Should a border agency make asylum seekers machine-readable before their stories are heard?

The weak version of AI ethics asks whether the model performs fairly enough. Katz pushes toward the prior question: why is this institution seeking this form of machine authority at all?

That question does not make technical work irrelevant. It gives each technical result a bounded job. Evaluation tests a capability under stated conditions; red teaming searches for failures; documentation preserves claims and limits; an impact assessment connects a deployment to people and rights. None supplies political authority by itself. These instruments become governance only when their findings can stop, shrink, redesign, or refuse the system.

In that sense, AI ideology is not only vendor hype. It is a governance interface. Laws, standards, procurement memos, dashboards, and ethics policies decide what counts as unbiased, trustworthy, high-risk, transparent, safe, or ready for adoption. Katz's book is useful because it keeps those categories from looking like neutral containers.

Governance and Safety

A Katz-informed governance file starts one step before model review. It asks for the institutional project: who wants the system, what power it extends, which people become more legible, which budget or grant line supports it, which vendor or university gains authority, which existing process is made to look obsolete, and what would count as a reason not to deploy.

That file should then connect to controls with exact status and scope. NIST's sociotechnical framing supports review of systemic, human, computational, and statistical sources of harm, and AI RMF Manage 1.1 makes proceeding a decision rather than an assumption. Under the EU AI Act, Article 10 data-governance requirements and Article 27 fundamental-rights assessments apply only to defined high-risk systems and actors on the amended timetable; they are not present universal duties for anything marketed as AI. OMB M-25-22 applies to covered federal acquisition. The label warrant should record which rule actually attaches, why, and when.

The safety implication is refusal capacity: an assigned right and workable process to delay, narrow, suspend, reject, or retire a system. Individual recourse and collective refusal are different controls. An appeal may correct one denial while leaving the workflow intact; refusal challenges the dataset, category, vendor, or workflow itself. For carceral, employment, welfare, education, housing, border, health, or surveillance uses, both routes need owners, deadlines, evidence access, anti-retaliation protection, and consequences.

A minimum project warrant should identify the task, institution, funding source, vendor, affected communities, legal basis, data sources, racialized or protected-class pathways, labor impacts, evaluation limits, appeal route, community-review authority, nonautomation baseline, sunset date, and owner empowered to stop the system. It should connect to procurement files, public registers, audit rights, complaint and incident channels, data-retention and deletion rules, and contract terms that preserve records, portability, audit access, model-change notice, and exit without loss of public control.

Before deployment, require an articulated public purpose, scoped data, independent evidence, funded affected-community review, and a real choice among rejection, non-AI redesign, bounded pilot, and deployment with renewal conditions. During use, require logs, monitoring, appeal, incident review, stop triggers, and versioned records. At renewal, test public outcomes and distribution of burden, not only model metrics, and require proof that withdrawal still works. The point is not to convert abolition into a form. It is to stop the form from deciding in advance that adoption is the only legitimate outcome.

Where the Book Needs Friction

Artificial Whiteness is polemical, and the polemic is both its force and its risk. It is strongest when it shows how AI talk can hide institutional violence behind technical inevitability. It is weaker if read as a complete taxonomy of every technical system that has ever traveled under the AI label.

Technical differences still matter. Expert systems, statistical machine learning, search ranking, recommender systems, facial recognition, language models, robotics, optimization tools, and agent frameworks do not all work the same way. Their failure modes, dependencies, labor politics, energy costs, legal duties, and governance levers differ. A critique of the AI label should not flatten those differences after showing how the label itself flattens politics.

The book also asks a lot of readers. Its argument depends on critical race theory, histories of empire, histories of AI, science studies, abolitionist critique, and political economy. Joshua K. Smith's Prometheus review is useful here because it praises the book's value while pressing on the practical difficulty of Katz's refusal politics. That is a real tension. Refusal can be clarifying, but institutions also need concrete ways to redirect money, shut down harmful systems, preserve useful tools, protect workers, and give affected communities operational power.

Those limits do not make the book less important. They make its best use clearer. Read Katz as a diagnostic for the authority that gathers around AI, not as a substitute for technical analysis, organizing strategy, procurement rules, labor policy, or democratic institution-building.

The book was published before today's foundation-model services, agent frameworks, OMB memoranda, and amended EU timetable. Applying it to those systems is this review's extension, not evidence that Katz tested their performance or predicted their legal treatment. A deployment review still has to identify the system version, data source, workflow, jurisdiction, affected population, and remedy. The useful position is both: do not let technical detail hide ideology, and do not let ideological critique erase differences that determine how a system can be tested or stopped.

The current fight over ideological bias in language models needs the same discipline. Katz gives a strong frame for asking how claims of neutrality become institutional power; he does not remove the need for empirical testing. A claim that a model is biased, unbiased, neutral, or truth-seeking still needs a versioned test protocol, source corpus, sampling method, evaluation rubric, deployment context, uncertainty account, error costs, and remedy path. Otherwise the critique becomes another slogan competing with vendor and government slogans.

What This Changes

The practical lesson is to audit the institutional project and the technical system as separate objects, in that order.

When a system is presented as AI, ask what the word is doing. Does it attract money? Deflect scrutiny? Convert a political problem into a technical problem? Create a new expert class? Make an old institution look modern? Turn coercion into service delivery? Make refusal seem irresponsible because the future has supposedly arrived?

Then ask what would count as success for the people most exposed to the system. Better accuracy may not be success. More representative training data may not be success. A cleaner dashboard may not be success. Success may mean fewer surveillance points, fewer automated denials, fewer carceral pathways, stronger appeal rights, smaller datasets, public ownership, worker control, community veto power, or no system at all.

A usable label warrant can be short but must be concrete. Name the institution, marketed label, functional description, version, claimed capability, non-AI baseline, affected population, data path, authority gained, legal or procurement trigger, refusal route, individual appeal route, stop owner, renewal evidence, and expiration date. Record rejected and nonautomated alternatives as carefully as the chosen system. If those fields are missing, the AI label is probably carrying more authority than the record can support.

Artificial Whiteness matters because it refuses to let AI stand as a natural event. The machine does not simply arrive. It is named, funded, narrated, deployed, repaired, defended, and normalized by institutions. Once that is visible, the question changes from how to adapt to AI to who is using the idea of AI, against whom, and for what world.

Source Discipline

This review separates book claims, author context, reception, policing evidence, law, implementation guidance, agency statements, and procurement memoranda. Columbia University Press and JSTOR support bibliographic and chapter-structure claims; the University of Michigan supports current author context. Reviews support interpretation, not proof that every deployment instantiates Katz's theory. WIRED, the LAPD Office of the Inspector General, and the National Academies support different parts of the Los Angeles policing account.

For current governance, Regulation (EU) 2024/1689 and amending Regulation (EU) 2026/1744 are binding legal texts; Commission pages explain implementation but do not replace them. NIST AI RMF 1.0 is voluntary and under revision. The 2023 multi-agency discrimination statement describes enforcement positions under existing authorities but expressly creates no new rights or obligations. OMB M-25-22 and M-26-04 are federal acquisition memoranda with bounded scope, not general scientific standards for AI, truth, or neutrality.

Do not use Katz as a shortcut for saying every technical system is identical. The book's point is that the AI label can flatten politics; a careful review should not reproduce that flattening. Claims about a system should preserve the specific capability, institution, dataset, vendor, affected people, legal regime, and review date.

The feedback-loop controls, label warrant, category-lock-in analysis, and refusal-capacity design are this review's operational inferences from the book and cited sources. They are not statutory requirements or empirical findings. The page paraphrases the book, uses only short chapter terms, and makes no claim that any AI system is conscious, divine, or AGI. Current book, governance, procurement, policing, and policy claims were rechecked on August 12, 2026.

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