Blog · Review Essay · Last reviewed June 25, 2026

The AI Con and the Hype Machine

Emily M. Bender and Alex Hanna's The AI Con is a book about language as a control surface. Its central claim is not that every machine-learning system is fake. It is that the public label "AI" often converts uncertainty into deference: automation becomes inevitability, extraction becomes progress, and ordinary managerial choices become the supposed demands of the future.

For this review, the AI con is a claim-laundering loop: a vague label turns a demo into authority, authority into procurement, procurement into dependency, and dependency into evidence that the system was inevitable. The test is not whether software is impressive. It is whether the claim survives task-specific evidence, affected-person rights, and an accountable path to stop using it.

The hype machine is the stack around that loop: naming, benchmark choice, launch demo, press coverage, investor story, procurement language, integration work, institutional dependency, and post hoc justification. A serious review asks where each step can be traced to evidence, who can contest it, and who has authority to reverse it.

The Book

The AI Con: How to Fight Big Tech's Hype and Create the Future We Want was published by Harper on May 13, 2025. Amazon lists the hardcover at 288 pages, with ISBN-10 0063418568 and ISBN-13 978-0063418561. HarperCollins describes the book as technology criticism about systems sold as artificial intelligence, the drawbacks of systems sold under that banner, and the way hype can cover concentrated power.

The authors bring complementary forms of authority to the argument. Bender is a University of Washington linguist and computational linguistics professor whose earlier work helped make scale, data, language, and meaning central questions in large-language-model criticism; the site's Emily M. Bender wiki page tracks that broader record. Hanna is a sociologist of technology, labor, and politics and Director of Research at the Distributed AI Research Institute. The book grows from that pairing: language critique joined to institutional critique.

The title's "con" should be read narrowly and carefully. The claim is not that statistics, machine learning, pattern recognition, or automation are imaginary. The con is the substitution of a brand category for a demonstrated system: a word that bundles unlike tools, hides tradeoffs, transfers authority to vendors, and asks the public to accept deployment before the task, evidence, limits, and affected people are visible.

Current Context

As of June 25, 2026, this is no longer only a media-literacy problem. AI hype has become a procurement, consumer-protection, civil-rights, public-record, and safety-review problem. Regulators and standards bodies are not asking whether a product feels futuristic. They are asking what the system does, what claim is being made, what evidence supports that claim, who is affected, and what records exist when the system fails.

The Federal Trade Commission's Operation AI Comply, announced on September 25, 2024, treated deceptive AI claims and AI-enabled deception as ordinary enforcement targets, including fake-review tools, "AI Lawyer" marketing, and AI-flavored business-opportunity schemes. The FTC's 2025 Air AI case continued that line into agentic-productivity claims: the agency alleged deceptive business-growth, earnings-potential, and refund claims around AI sales tools. A 2023 joint statement from the FTC, DOJ, CFPB, and EEOC made the same point for automated systems more broadly: existing consumer-protection, competition, civil-rights, and equal-opportunity laws still apply when the tool is sold as AI.

NIST's AI Risk Management Framework and its Generative AI Profile give this critique an operational vocabulary: govern, map, measure, and manage risks across the lifecycle, including design, development, deployment, operation, and decommissioning. OMB Memorandum M-25-21 is federal-agency policy, not a general public law, but it is a useful benchmark because it requires minimum risk-management practices for high-impact federal AI uses, impact assessments, pre-deployment testing, ongoing accountability, and discontinuation when safeguards cannot be met. The EU AI Act adds phased legal duties around high-risk systems, transparency for certain AI interactions and synthetic outputs, and information that lets deployers interpret outputs and use systems appropriately; the European Commission timeline lists Article 50 transparency rules as starting to apply on August 2, 2026, and the Commission's June 2026 transparency code page frames those duties around risks of deception and manipulation in the information ecosystem.

That current context sharpens the book's point. The practical question is not "is it AI?" The question is whether the buyer, vendor, agency, school, employer, or platform can produce a dated record of the exact system, task, baseline, evidence, affected population, workflow, human role, appeal path, incident process, monitoring plan, rollback point, and shutdown authority. Hype tries to collapse those details into a brand. Governance has to unpack them before deployment hardens into dependency.

A useful definition follows: hype is not enthusiasm. Hype is the substitution of a future-facing story for present-tense evidence. The system may still be useful; the defect is that the claim asks to be believed before its capability, scope, cost, failure mode, and remedy have been made inspectable.

Hype as Infrastructure

The book's strongest move is to treat hype as infrastructure rather than noise. A bad AI claim is not merely a mistaken sentence in a press release. It can help a company attract capital, discipline workers, influence procurement, weaken public skepticism, and make a contested deployment feel unavoidable. Once the slogan is installed, the product inherits borrowed authority before anyone has checked what the system actually does.

The mechanism is concrete. A demo becomes a news story; the news story becomes a budget line; the budget line becomes a procurement requirement; the procurement requirement becomes a dashboard; the dashboard becomes the institution's memory of what happened. By the time harmed users, workers, teachers, patients, applicants, or citizens meet the system, the argument has already moved from "does this work?" to "how do we adapt to it?" That is the same loop described in the site's review of The Hype Machine and in When the Benchmark Becomes the Curriculum: attention, measurement, and incentive systems train the reality they claim only to report.

A hype machine has recognizable parts: an elastic name, a cherry-picked task, a launch demo, a metric without a baseline, a comparison class that quietly shifts, a promise of labor savings, and a missing account of data, evaluation, failure, appeal, and decommissioning. The red flags are equally concrete: "human-level" without a metric, "autonomous" without a permission boundary, "AI-powered" without a versioned system, savings without workload evidence, safety claims without incident records, and productivity claims that ignore the labor of correcting the output.

That is why The AI Con belongs beside AI Snake Oil, Empire of AI, and Atlas of AI. Each book attacks a different layer of the same machine. Narayanan and Kapoor ask whether the claim is empirically supported. Hao follows the institution and its supply chain. Crawford maps the extraction behind the model. Bender and Hanna focus on the rhetorical gate that lets the others proceed: the moment a product becomes "AI" and starts receiving deference it has not earned.

Language, Labor, and Authority

Bender and Hanna are especially useful on anthropomorphic language. Calling a system a colleague, tutor, doctor, lawyer, artist, or intelligence does not only decorate it. The label changes the user's expectations, the buyer's tolerance for opacity, and the worker's place in the workflow. A generated answer starts to look like judgment. A statistical association starts to look like insight. A labor-saving device starts to look like a moral upgrade.

This matters in contracts. A school does not buy an "AI tutor"; it buys a software system that generates educational responses under specified conditions. A law firm does not buy an "AI lawyer"; it buys a tool that may produce legal-looking text inside a professional workflow. A clinic does not buy an "AI doctor"; it buys decision-support software whose intended use, validation, liability chain, and human-review authority must be explicit. The anthropomorphic noun can move responsibility away from vendor and deployer at the exact moment responsibility needs to be assigned.

The OECD's current definition is useful here because it is operational rather than mystical: an AI system is a machine-based system that infers from inputs how to generate outputs such as predictions, content, recommendations, or decisions that can influence environments. That definition does not settle whether a deployment is wise, fair, lawful, or worth buying. It pulls the conversation back to inputs, outputs, autonomy, adaptiveness, and context. Bender and Hanna's language critique does the same cultural work: keep the noun small enough that evidence can catch it.

The labor argument is equally important. Hype often presents automation as if it arrives from nowhere, but actual AI products rely on datasets, moderation, annotation, evaluation, customer-service scripts, benchmark construction, repair work, and the institutional labor of fitting a system into a workplace. When the product is framed as autonomous intelligence, the people who made it usable disappear twice: first behind the interface, then behind the story that the interface is replacing them. The companion reviews of Feeding the Machine and Ghost Work fill in that hidden production chain.

That makes The AI Con a book about belief formation. The con is not that every machine-learning system is useless. The con is that a flexible public word can collapse unlike systems into one aura of inevitability. Search ranking, synthetic media, resume screening, customer chatbots, welfare triage, code assistants, and agentic workflow tools become one mythic object, and the myth then speaks on behalf of each product.

Evidence Discipline

The most practical response is a claim ledger. Every AI claim should name the task, deployment context, affected population, data source, human labor, benchmark or field evidence, baseline comparison, failure mode, appeal path, security boundary, monitoring plan, accountable owner, review date, and shutdown trigger. Without that ledger, "AI-powered" is not a capability statement. It is a request for exemption from ordinary proof.

This is where the book connects to claim hygiene, AI evaluations, model and system cards, and algorithmic transparency. A benchmark score can be evidence, but it is not a governance answer by itself. A product demo can be useful, but it does not show distribution shift, edge cases, access barriers, labor displacement, appeal burden, or long-term institutional dependence. A responsible review asks how the system behaves when incentives, users, data, and oversight differ from the launch video.

A useful artifact is an AI claim receipt. It records the marketing sentence, the exact product version, the demo or benchmark used to support it, the field-validation date, subgroup results, known failure modes, data provenance, human labor dependencies, security permissions, sign-off authority, renewal date, and the person who owns the claim after purchase. It belongs beside an AI system inventory, AI procurement file, system card, and post-deployment incident log. Without that receipt, procurement can turn a vendor's adjective into an institution's policy.

The evidence ladder should be explicit. A marketing claim can justify curiosity. A lab evaluation can justify a bounded pilot. Field validation can justify a narrow deployment. Operational monitoring can justify continued use. Incident records, appeals, and rollback tests determine whether trust is being earned or merely asserted. Hype skips rungs; governance keeps the rungs visible.

The Governance Reading

The governance value of the book is practical: slow the word down. Before accepting an AI claim, ask what is being automated, what inputs are used, who labeled or supplied them, what output is produced, what evidence shows it works, who benefits, who can appeal, and what happens when the system is wrong. Those questions line up with the site's recurring concern that machine-readable authority needs contestable records, not stage magic.

Public institutions have begun naming adjacent problems. NIST's Generative AI Profile treats generative-AI risk as a lifecycle problem for design, development, use, evaluation, and decommissioning, not as a matter of impressive output alone. The FTC's Operation AI Comply made deceptive AI claims and AI-enabled deception an enforcement target in 2024. The EU AI Act, published in the Official Journal in 2024, requires high-risk systems to provide information that lets deployers interpret outputs and use systems appropriately under Article 13, and its Article 50 transparency rules for certain AI interactions and synthetic outputs are part of the Act's phased implementation. Those documents do not reproduce Bender and Hanna's politics, but they confirm the central premise: hype is now a risk surface.

The safety implication is not only consumer protection. Inflated claims can move authority into systems before human oversight, incident review, security boundaries, public records, worker voice, and appeal rights exist. That matters for hiring, education, public benefits, law, health, finance, journalism, and workplace management. Procurement should treat an unsubstantiated AI claim as a defect, not as sales color: the buyer should require evidence, deployment constraints, monitoring, recourse, incident reporting, audit rights, rollback tests, and exit terms before the label is allowed into a contract.

The governance controls are concrete. Require claim substantiation before purchase. Maintain an AI inventory that includes vendor, version, use case, affected population, and owner. Put model cards, system cards, logging, audit rights, notice, appeal, human-review authority, incident reporting, and non-use decisions into the same record. Make the person who can approve the deployment also name the person who can pause or retire it. The point is not to slow every tool equally; it is to prevent weak language from outrunning the evidence required by the risk.

A strong procurement file should also preserve negative decisions. If a vendor cannot substantiate a productivity number, replacement claim, bias claim, security claim, or autonomy claim, the refusal should remain in the record. Otherwise failed hype disappears, while successful hype becomes infrastructure. That file belongs beside AI audit trails, AI Bills of Materials, and public registers where the use case warrants public memory.

The useful governance frame is layered. AI governance defines the authority chain; AI audits and assurance test records against claims; human oversight asks whether review has power; notice and appeal keeps affected people from becoming silent data points; duty of care for AI platforms asks what obligations follow from scale; and safety cases force a release decision to carry reasons instead of vibes.

Where the Book Needs Care

The book is polemical by design, and that sharpness is part of its usefulness. Still, readers should avoid turning its critique into a reflexive refusal to evaluate real capability. Some AI systems are useful in narrow settings. Some accessibility tools, scientific workflows, translation aids, code assistants, and pattern-recognition systems can help when they are bounded, tested, documented, and accountable. The target is not computation. The target is the story that lets computation outrun evidence and consent.

That distinction matters because anti-hype can become counter-hype if it makes every deployment sound identical. A résumé screener, image classifier, chatbot, foundation-model API, recommender, coding assistant, and agentic workflow do not carry the same evidence burden in every setting. The book is strongest when its skepticism becomes classification: what system, what task, what data, what authority, what harm, what proof, and what exit?

The other limit is audience. The AI Con is built to arm citizens, workers, readers, and policymakers against manipulative language. It is not a technical manual for model evaluation, a full labor history of data work, a security framework for agentic systems, or a complete policy design for every AI sector. Its best use is as a front-door discipline: if the claim cannot survive plain questions about task, evidence, power, labor, and recourse, it should not get to hide behind the word "AI."

What This Changes

The site studies how interfaces become institutions and how institutions train belief. The AI Con gives that study a compact warning: the first interface is often the label. Before a user ever meets a chatbot or agent, they meet a story about intelligence. That story can invite care, fear, reverence, obedience, purchase, resignation, or policy panic.

A responsible AI culture would not begin by asking whether the system is impressive. It would ask what social power is being rearranged by calling it intelligent. It would keep language ordinary until evidence earns stronger words. It would refuse the move from demo to destiny. Most of all, it would keep human labor, institutional incentives, data extraction, appeal rights, and shutdown authority visible at the moment hype tries to make them vanish.

That makes the book useful beyond the current AI news cycle. The recurring pattern is older than any model release: a technical vocabulary becomes a social credential; the credential becomes a budget; the budget becomes infrastructure; the infrastructure becomes hard to question because people now depend on it. The antidote is not cynicism. It is the discipline of writing down claims while they are still reversible.

Source Discipline

This review separates source types. HarperCollins and Amazon support publication metadata. The official book site supports the authors' public framing. University of Washington, Alex Hanna's profile, and the Stochastic Parrots resource page support author context. OECD, NIST, FTC, OMB, and EU sources support current governance and regulatory context. Internal links provide site vocabulary and adjacent reading, not independent proof of external facts.

The verbs matter. A publisher lists; an author argues; a regulator alleges or enforces; a standard recommends; a memorandum directs agencies; a law requires or phases in obligations; a vendor claims; a benchmark measures a bounded task. Mixing those verbs is how hype becomes false certainty.

The same discipline applies to this review's own language. "AI con" names a claim-laundering pattern, not a finding that every machine-learning system is fraudulent. "Hype" names evidentiary substitution, not ordinary optimism. "Transparency" means little unless it names who receives what information, at what point, and with what power to refuse, appeal, audit, or exit.

This page does not treat any AI system as conscious, divine, or AGI, and it does not infer inevitable social transformation from fluent output. It treats AI systems as engineered products, workflows, interfaces, and institutional decision aids whose authority should be limited by evidence, contestability, and the ability to stop.

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