Blog · Review Essay · Last reviewed June 25, 2026

Artificial Unintelligence and the Politics of Technochauvinism

Meredith Broussard's Artificial Unintelligence is a practical antidote to machine enchantment. Its central lesson is not that computers are useless. It is that institutions become dangerous when they treat computation as proof of superior judgment.

For this review, technochauvinism means the institutional habit of treating a technical solution as superior before the problem, evidence, affected people, maintenance burden, and nontechnical alternatives have been inspected. The danger is not only bad software. It is the transfer of authority from situated judgment to systems that make the world legible by compressing it.

The safety test is concrete: what social duty is being converted into an interface, what evidence shows that the conversion improves the duty, who can contest the output, and who has authority to stop the system when the machine-readable version of the world begins to replace the world it was meant to serve?

The counterpractice is right-tool governance: start with the duty and the affected people, then decide whether software, staffing, repair, policy change, public records, or no automation is the responsible instrument.

The Book

Artificial Unintelligence: How Computers Misunderstand the World was published by MIT Press in 2018. MIT Press lists the hardcover publication date as April 27, 2018, the paperback as January 29, 2019, and the book at 248 pages. The Press describes it as a guide to the inner workings and outer limits of technology, written by a software developer and journalist who argues against assuming that computers reliably get things right.

Broussard is a professor at NYU's Arthur L. Carter Journalism Institute and research director at the NYU Alliance for Public Interest Technology. Her background matters because the book is not a rejection of technical knowledge from outside the field. It is a programmer's argument for better judgment about what computation can do, what it cannot do, and what social damage follows when people confuse automation with intelligence.

The book won the 2019 PROSE Award in Computing and Information Sciences, according to the PROSE Awards list and MIT Press's awards page. Its durability comes from a plain diagnosis: many failures sold as AI breakthroughs are actually failures of institutions, incentives, data, assumptions, maintenance, and accountability.

Current Context

As of June 25, 2026, Broussard's warning has become more operational, not less. Generative AI, retrieval systems, automated decision tools, and agentic workflows have made computation feel more conversational and more capable, but the old problem remains: a system can be fluent, scalable, and expensive while still being pointed at the wrong task, trained on partial data, wrapped in weak oversight, or used to avoid a political decision.

The current governance vocabulary now names parts of that problem. NIST's AI Risk Management Framework gives organizations a voluntary lifecycle structure for managing AI risks to people, organizations, and society. ISO/IEC 42005:2025 gives guidance for AI system impact assessments focused on how systems and foreseeable applications may affect individuals, groups, and society. OMB Memorandum M-25-21 requires U.S. federal agencies to apply minimum risk-management practices for high-impact AI uses, including pre-deployment testing, AI impact assessments, ongoing monitoring, human oversight, and discontinuation when minimum practices cannot be met. The FTC, DOJ, CFPB, and EEOC joint statement on automated systems makes the same point from enforcement: existing civil-rights, consumer-protection, competition, and equal-opportunity laws still apply when a decision is automated.

The EU AI Act adds a legal version for many high-risk systems, including uses in education, employment, essential services, law enforcement, migration, justice, and democratic processes. Its duties around risk management, transparency, logging, human oversight, deployer obligations, complaint paths, and explanation routes are not a complete answer to technochauvinism. They are a refusal to let the technical surface erase the institutional record.

The practical update is that "does the model work?" is too small a question. A 2026 review has to ask whether the system is the right instrument for the problem; whether the affected people can inspect, refuse, or appeal; whether the institution can monitor drift and harm; and whether a budget, staffing, policy, or care intervention would solve the real problem more directly than software.

That is the current edge of Broussard's argument. Impact assessments, inventories, audits, and procurement files are useful only if they preserve the possibility that the answer is not automation. Otherwise governance becomes a better way to buy the wrong tool.

Technochauvinism

Broussard's key term is technochauvinism: the belief that technological solutions are inherently superior. That belief is less crude than it sounds. It often appears as a management habit, a funding preference, a procurement slogan, a product demo, or a reform plan that treats social complexity as a backlog item waiting for software.

This is why the book belongs beside To Save Everything, Click Here, The Technological Society, and Weapons of Math Destruction. Each asks what happens when a technical method receives authority before the surrounding institution has earned trust.

Broussard's tone is useful because it avoids mystical fear. Computers are powerful symbolic machines. They are also literal machines with brittle inputs, narrow representations, maintenance burdens, biased histories, security problems, and human organizations around them. The myth begins when people look at the first fact and stop inspecting the rest.

A sharper definition helps: technochauvinism is not ordinary enthusiasm for tools. It is a burden shift. The institution no longer has to prove that automation fits the problem; skeptics, workers, teachers, patients, applicants, journalists, or residents are asked to prove why the machine should not be used. Once that shift happens, the tool has already won power before its evidence has been tested.

The opposite of technochauvinism is not nostalgia or blanket refusal. It is instrument discipline: the willingness to use computation where it fits, limit it where it distorts, and choose nontechnical repair where the real deficit is money, staffing, care, legal authority, trust, or democratic legitimacy.

How Computers Misunderstand

The title's force is in the word "misunderstand." Computers do not misunderstand in the human sense. They process formal representations. The mistake is ours when we feed messy human situations into narrow computational forms and then treat the output as if it contained the missing context.

The book's examples move across driverless cars, machine learning, standardized tests, campaign finance data, and programming practice. MIT Press's summary notes that Broussard tests AI against standardized-test problems, uses machine learning on the Titanic survival dataset, and attempts to build software for campaign-finance investigation. Those examples make the same point from different angles: impressive computation can still fail when the problem is underspecified, the data is partial, or the world does not match the model's assumptions.

Her sharpest case study is the Philadelphia public schools, where she set out to use software and data to answer an apparently simple question: whether students had what they needed to pass the state's standardized tests. The investigation kept colliding with material reality, missing and unreliable records of what each school actually held, and beneath that, classrooms lacking the basic books and resources the tests assumed students would have. The lesson was not that the algorithm was badly written. It was that a data-and-software fix had been aimed at the wrong question. You cannot compute your way to higher test scores in a district that cannot afford the textbooks, and the technochauvinist reflex is exactly what lets an institution reach for the model instead of the budget.

That lesson is easy to forget in the generative AI era because interfaces now talk back fluently. A chatbot can make misunderstanding look like comprehension. A system can summarize, advise, rank, or explain without having a grounded grasp of the social situation it is being asked to govern.

The AI-age version is not simply that models hallucinate. It is that institutions can hallucinate with them. A school can turn scarcity into an analytics problem. A public agency can turn access to benefits into a triage problem. A workplace can turn supervision into a dashboard problem. In each case, the computer's formal representation is narrow, but the institution's willingness to accept that narrowness is the larger failure.

The Institutional Problem

Artificial Unintelligence is strongest when read as institutional criticism. The danger is not only that a model makes mistakes. People make mistakes too. The danger is that automated mistakes can be wrapped in procurement authority, scaled through workflows, hidden behind vendor secrecy, and normalized as progress.

A public agency can buy a system before it knows how to audit it. A school can adopt a platform because dashboards look more objective than teacher judgment. A newsroom can chase automation while weakening the reporting labor that would check the data. A company can frame deskilling as modernization. In each case, the computer is not acting alone. It is lending prestige to an institution that wants speed, certainty, savings, or legitimacy.

This is the bridge to legibility. A computer system needs the world to arrive as fields, labels, scores, documents, images, prompts, clicks, and logs. Once an institution depends on those formats, people learn to become machine-readable. They optimize resumes for filters, claims for portals, identities for forms, speech for moderation systems, and work for dashboards. The map does not simply describe the person. It trains the person to survive the map.

That is where the book connects to recursive reality. A system encodes a partial view of a domain. The institution acts as if the partial view is complete. People adapt to the interface, the categories, and the incentives. Their adaptation becomes new data. The next system then appears more accurate because the world has been bent toward the categories that were easiest to compute.

The AI-Age Reading

Read in 2026, the book feels less like a period piece than a precondition for AI literacy. Many current AI arguments begin too late, at model capability. Broussard begins earlier, with the cultural permission structure that lets organizations reach for computation before asking whether computation is the right form of care, judgment, repair, or governance.

Large language models make technochauvinism more seductive because they move from number-shaped authority to language-shaped authority. Older automated systems often looked bureaucratic: scores, ranks, flags, eligibility notices. New systems can sound patient, nuanced, and self-aware. That does not eliminate the old risks. It makes them easier to accept.

The practical test is simple: what human capacity is the system replacing, and what institutional duty is it helping avoid? If an AI tutor supplements a teacher, the details matter. If it becomes an excuse to underfund teachers, the product has become policy. If an AI assistant helps a benefits worker find information, that is one thing. If it becomes an unappealable filter between a person and support, the interface is now an administrative authority.

Broussard's book gives readers permission to be specific. Do not ask whether AI is good or bad in the abstract. Ask whether the data fits the case, whether the objective fits the social purpose, whether affected people can contest the output, whether the tool improves institutional responsibility, and whether a nontechnical solution would work better.

The same specificity matters for AI agents. A model that drafts a memo is different from a system that sends the memo, updates the record, denies the request, schedules the shift, or changes access. Tool use turns text into action. Technochauvinism becomes dangerous when a fluent interface makes that action feel inevitable before permissions, logs, review authority, and appeal paths are in place.

Governance and Safety

The governance lesson is to make the right-tool question mandatory before procurement. A technochauvinist system begins with the answer: software. A safety-oriented institution begins with the duty: educate, care, hire, allocate, investigate, inform, repair, protect, or decide. Only after the duty is named should the organization ask whether computation improves the duty enough to justify the costs and risks.

The practical artifact is a right-tool warrant. It should be signed before purchase or deployment, not after a vendor has framed the choice. The warrant asks what duty is being served, what nontechnical alternatives were considered, what evidence would defeat automation, what affected people need to inspect or appeal, what resources the system may be masking, and who can cancel the project if the tool narrows the duty instead of improving it.

For high-stakes uses, the minimum record should include the problem statement, nontechnical alternatives considered, data provenance, intended use, affected population, validation evidence, accessibility effects, subgroup performance, human oversight role, appeal route, incident process, monitoring plan, vendor obligations, retirement criteria, and the person with authority to pause or stop the system. That file belongs beside an AI system inventory, algorithmic impact assessment, AI audit trail, and procurement record.

NIST's govern-map-measure-manage structure is useful because it prevents a narrow evaluation from standing in for accountability. Govern asks who owns the decision and risk. Map asks whether the tool fits the context and affected people. Measure asks whether evidence actually tests the claim. Manage asks what changes when the tool fails, drifts, harms, or no longer serves the mission.

Meaningful human oversight is not a checkbox. The reviewer needs time, training, evidence, independence, and authority to change the result. A "human in the loop" who is expected to rubber-stamp a model is not oversight; that person is a liability buffer. Broussard's argument makes the design principle plain: automation should make responsibility easier to locate, not easier to deny.

The safety checklist is blunt: no automation where the real problem is missing resources; no score without a declared purpose and validation domain; no generated explanation without an evidence trail; no dashboard that hides uncertainty; no vendor claim without audit access; no human review without authority; no consequential decision without notice and appeal; and no deployment that cannot be paused when the social cost outruns the technical convenience.

For agents and generative interfaces, add one more boundary: no tool action should be treated as an answer. Drafting, sending, updating a record, denying a request, scheduling a worker, or changing access are different powers. The right-tool warrant should separate read, write, send, approve, spend, delete, and permission-change authority before the interface makes them feel like one smooth task.

Where the Book Needs Updating

The book predates ChatGPT, consumer-scale generative AI, tool-using agents, synthetic media pipelines, prompt injection, model cards as public governance artifacts, and the current compute race. It does not fully address foundation-model supply chains, large-scale data extraction, or the infrastructure politics of training and deployment.

Its examples are deliberately accessible, which is a strength for public literacy and a limit for specialist readers. People looking for detailed technical audits of transformer models, reinforcement learning, interpretability, or frontier-model risk will need other books and papers.

The other update is power. Since 2018, the politics of AI have become inseparable from cloud concentration, data-center buildout, copyright disputes, labor extraction, public-sector procurement, and private vendor dependence. A critique of technochauvinism should not stop at individual bad judgment. It should ask who profits when the technical solution becomes the default solution.

There is also a risk of turning technochauvinism into an accusation that ends the conversation too early. Some software genuinely improves access, coordination, evidence review, safety monitoring, translation, disability accommodation, or public-service delivery. Broussard's better lesson is to keep the burden of proof attached to the specific task and institution, not to treat every tool as guilty by category or innocent by demo.

Still, the book's central warning survives the model shift. A society can be technically sophisticated and still be naive about institutions. It can build better models while making worse choices about where authority should live.

What This Changes

The lesson is to test every machine claim against the world it compresses.

When an institution adopts AI, the central question is not only accuracy. It is what has been made legible, what has been erased, who must adapt to the interface, who can refuse, who can appeal, what labor disappears from view, and which human obligations are being converted into technical outputs.

Recursive reality begins when a representation starts training the reality it represents. A model misreads a domain. An institution acts through that misreading. People adapt to the system's categories. Their adaptation becomes new data. The next version looks more confident because the world has been bent toward the instrument.

Broussard's contribution is a useful discipline of disenchantment. Treat computers as tools with limits, not oracles with dashboards. Demand working systems, honest evidence, responsible institutions, and the humility to leave some problems in human hands.

The positive version is not anti-technology. It is right-sized technology: tools that serve an inspected duty, preserve contestability, disclose their limits, and remain subordinate to accountable human institutions. If the system cannot answer who is helped, who is burdened, who can appeal, and what nontechnical repair was ignored, the smart move may be to stop building.

Source Discipline

This review separates book facts, author context, awards, interpretation, and current governance claims. MIT Press and MIT Press Direct support bibliographic details and the publisher's description of the book. PROSE and MIT Press award pages support the award claim. Broussard's author page and related institutional profiles support current author context. NIST, ISO, OMB, FTC/DOJ/CFPB/EEOC, and EUR-Lex support the current risk-management and regulatory vocabulary.

The interpretive claim is narrower than the sources: Artificial Unintelligence is used as a theory of institutional overconfidence, not as proof that every AI deployment is harmful or that computation has no public value. Current standards and laws are used for scope and vocabulary, not as evidence that compliance equals justice.

This page does not treat any AI system as conscious, divine, or AGI. It treats AI systems as engineered tools embedded in organizations, contracts, interfaces, data flows, incentives, and legal duties. The question is where authority goes when the tool enters the institution.

Current claims were rechecked on June 25, 2026. Dates and scope matter: OMB memoranda apply to covered federal agencies, ISO standards are not statutes, NIST frameworks are voluntary, the EU AI Act applies by role and risk category, and U.S. agencies enforce through existing legal authorities rather than one general federal AI law.

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