More than a Glitch and the Systemic Bias Machine

The Book

More than a Glitch was published by MIT Press in 2023, with a paperback in 2024. MIT Press lists the hardcover, ebook, and paperback at 248 pages, with the hardcover published March 14, 2023 and the paperback published April 2, 2024. The publisher also lists the book as a 2023 getAbstract International Book Award winner for Business Impact, a 2024 PROSE Award finalist in Popular Science and Mathematics, and a Financial Times Best Summer Books of 2023 technology selection.

Broussard is a data journalist and professor at New York University's Arthur L. Carter Journalism Institute, and research director at the NYU Alliance for Public Interest Technology. The book follows her earlier Artificial Unintelligence, but it narrows the attack. Where the earlier book named technochauvinism as the assumption that computational solutions are superior, More than a Glitch asks what happens when that assumption is carried into systems that classify bodies, allocate opportunity, interpret health, grade students, police neighborhoods, and define default users.

The book is structured as a broad public-interest technology argument rather than a single-sector case study. Sources reviewing the book identify its recurring cases as machine learning, facial recognition, criminal justice, algorithmic grading, disability and accessibility, gender, racism, medicine, algorithmic auditing, and policy responses. That breadth is the point: the glitch frame fails because the same pattern recurs across domains.

Current Context

As of June 25, 2026, Broussard's argument has moved from critique into the ordinary language of AI governance. NIST Special Publication 1270 treats AI bias as a problem across the design, development, and use of systems, not merely as a statistical defect in a finished model. The NIST AI Risk Management Framework similarly frames AI risk as a matter for organizational governance, mapping, measurement, and management, including risks to individuals and society. ISO/IEC 42005:2025 adds impact-assessment guidance for identifying effects on individuals, groups, and society before and during use.

The EU AI Act gives that diagnosis a regulatory form for high-risk systems. Article 10 requires data governance for training, validation, and testing data, including choices about collection, preparation, assumptions, representativeness, and bias detection. Article 27 requires fundamental-rights impact assessments for certain deployers of high-risk AI systems, with attention to affected groups, risks, human oversight, and mitigation. Article 113 phases in application dates, so a careful review keeps claims jurisdiction-specific and date-specific.

U.S. enforcement sources point in the same direction without creating one general AI statute. The 2023 joint statement from the FTC, DOJ Civil Rights Division, CFPB, and EEOC said their existing authorities apply to automated systems. The EEOC's September 11, 2023 iTutorGroup settlement treated allegedly automated age and sex filtering as an employment discrimination matter. CFPB Circular 2022-03 says complex algorithms do not excuse creditors from giving specific and accurate adverse-action reasons. For federal agencies, OMB Memorandum M-25-21 separately requires minimum risk-management practices for high-impact AI, including documented impact assessments, testing, monitoring, and timely human review or appeals where appropriate.

Accessibility has also become a live governance issue, not a courtesy. The Justice Department's 2024 Title II ADA web and mobile accessibility rule sets requirements for state and local government digital services, with compliance timing affected by a 2026 Federal Register extension. The rule references WCAG 2.1 Level AA, while W3C's WCAG 2.2 remains a current web-accessibility recommendation. For Broussard's purposes, that matters because a biased system can exclude people before a model score is ever produced: at the form, identity check, voice interface, proctoring screen, appeal portal, or service counter. The access path is part of the system boundary.

The Glitch Excuse

A glitch is a temporary failure inside a system that otherwise deserves confidence. It suggests that the architecture is sound, the harm is accidental, and the fix is local. Broussard's title refuses that comfort. Many biased technologies do not fail because one line of code slipped. They fail because the system inherits an already unequal world, compresses it into data, and then gives the compressed version operational authority.

This is why the book belongs beside Algorithms of Oppression, Weapons of Math Destruction, Race After Technology, and Unmasking AI. All four books attack the same basic alibi: that computation is neutral until a biased user misuses it. Broussard's contribution is to make the alibi unusually practical. She does not let the reader hide inside either math or moral outrage. The question is always: what was built, for whom, with whose categories, against whose interests, under what institutional authority?

The glitch excuse is especially dangerous in AI governance because it preserves deployment momentum. If the problem is framed as a bug, the system can stay in place while engineers patch around the complaint. If the problem is structural, the harder options come into view: stop using the system, change the decision process, narrow the domain, add appeal rights, change data collection, redesign categories, repair accessibility barriers, or admit that a task should not be automated.

The governance move is to treat severe unequal performance as a structural incident until proven otherwise. That means asking for a root-cause record, not only a patch note: which design choice, dataset, category, threshold, vendor claim, interface assumption, or human workflow made the harm repeatable?

That distinction changes the status of a complaint. A person harmed by a discriminatory score, unusable interface, inaccessible verification step, or impossible appeal is not merely reporting a defect for the vendor's queue. They may be describing a civil-rights, consumer-protection, employment, credit, education, healthcare, or public-service failure. Governance that treats every complaint as a support ticket will miss the institutional harm.

Categories Become Machinery

The book's strongest AI-era lesson is that categories are not harmless labels attached after the real technical work is done. They are part of the machinery. A database field for gender, a disability accommodation workflow, a skin-tone distribution in training data, a medical-risk proxy, a credit feature, or a classroom metric can decide what the model can see before the model begins to learn.

That makes bias more durable than a bad prediction. A bad prediction can be corrected. A bad category can keep producing error while appearing orderly. It can tell institutions who counts as normal, who is an edge case, who must appeal, who must produce extra evidence, and who disappears from the measurable world. Once those categories are embedded in procurement, dashboards, APIs, training data, and evaluation metrics, they become harder to contest than ordinary human prejudice because they arrive as infrastructure.

For AI systems, this is a legibility problem. Institutions want people to become machine-readable. But the template for readability is not neutral. It often reflects the people who had the power to define the form, the benchmark, the default body, the default name, the default face, the default speech, the default family, and the default life path. The model then returns that narrow picture as if it had discovered reality.

A high-stakes category therefore needs a warrant, not just a data dictionary. The warrant should explain why the category exists, what legal authority supports it, what alternatives were rejected, which groups may be missing or misread, how the category can be contested, and how it will be retired if it becomes a route for surveillance, exclusion, stigma, or coerced self-description.

The practical question is not only whether a category is statistically useful. It is whether the category is lawful, necessary, contestable, explainable to affected people, and safe in the hands of the institution using it. A category can improve a benchmark and still create a route for surveillance, exclusion, stigma, or coerced self-description.

The Institution in the Model

Broussard's framing also prevents an overly model-centered reading of AI harm. The model is rarely the whole system. A biased face-recognition result matters because a police department, border agency, school, platform, landlord, employer, hospital, or insurer can act on it. A medical algorithm matters because clinicians, billing systems, hospital policy, vendors, and liability fears translate its score into care. A grading system matters because administrators need scalable judgment and students have weak power to contest it.

This is the connection to the site's recurring concern with automated authority. A model becomes powerful when an institution uses it to replace, compress, or pre-structure judgment. The danger is not only that a system is wrong. It is that an organization accepts the system's wrongness as normal paperwork: a score, a flag, a denial, a case note, a risk level, a confidence interval, a ticket closure.

The decision chain should be auditable from model output to institutional act. If a score becomes a denial, flag, investigation, medical note, classroom sanction, hiring screen, or benefit delay, the affected person needs to know what evidence was used, who had authority to disagree with the system, and what remedy can change the record.

That also explains why inclusion alone is not enough. More diverse teams, better datasets, and broader usability work can matter. But a decision system can become more inclusive while still routing people through an unjust process. A fairer classifier attached to an unfair institution may simply distribute harm with better optics. Broussard's book keeps pushing the reader back to purpose: why does this automated decision exist, and who gains authority when it works?

That is also why procurement and vendor governance matter. A tool can arrive with an accuracy claim, a fairness slide, and a security questionnaire while leaving the real questions unanswered: what task is being automated, who can challenge the result, what evidence is logged, what subgroup testing was done, what accessibility testing was done, what uses are forbidden, and who is responsible when a model output becomes an official decision.

Ability Bias and Access

The book is particularly useful because it keeps disability in the foreground rather than treating accessibility as an afterthought. Ability bias is not just a missing compliance checklist. It is a design imagination problem. Systems assume bodies, senses, time, attention, mobility, speech, gesture, and cognition. Those assumptions decide who can pass a verification screen, use a workplace tool, receive a service, appeal a decision, or be recognized as a legitimate user.

AI makes that more important, not less. Voice systems, proctoring tools, hiring assessments, identity checks, workplace analytics, classroom software, companion bots, care robots, and medical triage tools all encode expectations about how a person should look, speak, move, respond, and explain themselves. When the interface becomes the gate, the interface's model of the body becomes policy.

An accessible alternative cannot be a grudging exception hidden after failure. It has to be part of the safety case: documented before deployment, easy to find, staffed by people with authority, logged without stigma, and able to produce the same substantive outcome as the automated path.

This is a more concrete way to think about human-machine cognition. The problem is not only whether a machine understands a person. It is whether the person must reorganize their life so the machine can process them. A humane system gives people more ways to be understood. A high-control system narrows the acceptable human until the user has to perform legibility for the machine.

For safety work, accessibility is not separate from accuracy or fairness. CAPTCHA systems, speech recognition, facial verification, remote proctoring, timed forms, keyboard traps, inaccessible PDFs, and chatbot-only service channels can deny access before any formal decision is made. Testing only the model while ignoring the access path produces a narrow and misleading account of harm.

Recursive Reality

The most important loop is simple. Institutions collect records from an unequal world. Engineers turn those records into models. Institutions act on the models. People adapt to those actions. The adaptations produce new records. The next system treats those records as evidence. At that point, the model is not merely reflecting bias. It is helping manufacture the world that will later justify it.

That is why the book matters for generative AI and agentic systems, even when many examples come from older algorithmic decision tools. Today's answer engines, copilots, multimodal models, and agents do not merely classify people at a distance. They write summaries, draft reports, recommend decisions, fill forms, search records, speak to users, and carry institutional language from one workflow into another. If the underlying categories are biased, the system can spread the bias through fluent prose and routine action.

That makes data provenance and post-market monitoring bias controls, not paperwork. An organization has to know which records taught the system, which outputs changed later records, which complaints signal recurrence, and when a model or workflow should be paused because the feedback loop is amplifying harm.

The loop also changes belief formation. A biased output repeated through dashboards, model summaries, case-management systems, search snippets, and official letters becomes harder to experience as a claim. It feels like the administrative world reporting on itself. That is how a technical artifact becomes common sense: it disappears into procedure.

This is the article's bridge from book review to the site's larger project. Machine-readable reality is recursive when records feed models, models feed decisions, decisions create records, and those records later train or justify the next system. Broussard's contribution is to insist that race, gender, disability, and access are not decorations on that loop. They are often among the first things the loop learns to preserve.

Governance and Safety

The governance implication is that bias review has to begin before model evaluation. A serious review names the decision context, affected groups, data provenance, labels, proxies, thresholds, model version, vendor role, human workflow, accessibility path, appeal path, and feedback loop. It asks not only whether the system is accurate, but who bears the cost of false positives, false negatives, denial, delay, surveillance, stigma, or forced self-disclosure.

Subgroup and intersectional testing matter, but averages are not enough. The same error rate can have different consequences in hiring, school discipline, public benefits, credit, healthcare, fraud detection, identity verification, and content moderation. Governance has to specify the harm, the comparator, the decision authority, and the remedy.

Public rules now make some of this concrete. The EU AI Act points high-risk systems toward data governance and impact assessment. NYC's automated employment decision tool rule requires covered employers and employment agencies to ensure a bias audit, publish a summary, and give notices. The CFPB's adverse-action circular is a useful anti-black-box principle: a complex model does not erase the duty to explain a consequential denial in specific terms.

A systemic-bias case file should preserve the minimum evidence needed to govern the system: purpose, data provenance, category definitions, absent or underrepresented groups, subgroup and intersectional tests where lawful, accessibility tests, proxy-risk review, threshold policy, human workflow, notice language, appeal route, vendor obligations, incident history, monitoring results, accountable owner, and retirement criteria. Without that record, the institution is asking the public to trust a machine it cannot reconstruct.

The safety rule is practical: no consequential classification should operate without notice, an evidence record, accessibility testing, a human with authority to change the outcome, vendor obligations that survive procurement, incident logging, post-deployment monitoring, and a non-automated or alternative path where automation is not fit for the person or the decision. In some settings the right answer is not a fairer model, but a narrower use, a changed institution, or a decision that should not be scored at all.

Bias Audit Theater

Broussard's anti-glitch argument is also a warning about audit theater. A weak audit asks whether a vendor can produce a fairness table. A useful audit asks whether the table can change the deployment. If the audited population is too narrow, the tool version is unclear, disabled users are absent, subgroup results are averaged away, complaints are not connected to remediation, or the contract lets the vendor withhold logs, then the audit has become a credential rather than a control.

The problem is familiar from The Audit Society and Trust in Numbers: verification can become a ritual that protects the institution from accountability while appearing to satisfy accountability. In bias work, that ritual is especially dangerous because a public audit summary may persuade affected people that the system has already been checked, even when the check did not test their route through the system, preserve their evidence, or give anyone authority to change the outcome.

The corrective is an audit with leverage. The record should name the system version, decision context, data provenance, category warrant, affected groups, lawful subgroup and intersectional testing, accessibility findings, unresolved harms, vendor limits, appeal outcomes, incident history, remediation deadlines, owner, and suspend-or-retire criteria. An audit that cannot delay, narrow, alter, or end a deployment is not governance. It is documentation after the institution has already decided to proceed.

Where the Book Needs Friction

The strength of More than a Glitch is its range and clarity. The weakness is the same. Readers looking for a deep technical audit manual, a detailed procurement playbook, or a full legal theory of algorithmic discrimination will need other books and policy materials beside it. Broussard gives a broad diagnostic and reform argument; she does not exhaust every domain she enters.

There is also a risk that the anti-glitch frame can flatten different failure modes. Some harms come from explicit exclusion, some from weak measurement, some from bad proxies, some from inaccessible design, some from historical data, some from vendor secrecy, some from institutional incentives, and some from using an otherwise accurate system for a task that should not be automated. The practical work is to preserve those distinctions after accepting the larger point that bias is not accidental noise.

The book is best read as an inspection habit. When a system harms people unevenly, do not start by asking how to patch the output. Start by asking what social world the system assumes, what institution it serves, what categories it hardens, and whether automation has made the wrong thing easier.

What This Changes

More than a Glitch changes the burden of proof. A developer, vendor, or agency should not be able to say that a system is neutral until critics prove bias. The stronger presumption is that every consequential system carries assumptions about race, gender, disability, class, geography, language, documentation, and bodily presentation. Governance begins by making those assumptions inspectable before deployment, not after harm becomes public.

The review standard follows from that. Ask who defined the task, who is represented in the data, who is missing, who can appeal, who audits the system, what remedies exist, what uses are prohibited, whether the interface works for disabled users, whether a true output can still be unjust, and whether a false output can become institutional fact before anyone can contest it. Then put the answers in an AI system inventory, impact assessment, audit trail, and incident reporting process that can be inspected later.

The book's durable lesson is that the machine is not separate from the world it measures. It is a social instrument that can turn old exclusions into new infrastructure. Calling the result a glitch is a way of keeping the machine innocent. Broussard's answer is cleaner: inspect the system, inspect the institution, and be willing to turn it off.

Source Discipline

Use sources for the level of claim they can support. MIT Press and Broussard's author site establish book metadata, author affiliation, and publisher framing. Academic and public reviews help situate the book's reception and recurring cases. NIST, ISO, EU, CFPB, EEOC, FTC, NYC, DOJ, and W3C sources support governance context, but they do not prove that a particular deployed system is lawful, fair, accessible, or safe.

Claims about bias should be scoped to the system, domain, population, jurisdiction, date, and decision. A finding about a facial recognition benchmark is not proof about every identity system. A hiring-tool audit requirement is not a general guarantee of fair employment automation. A model's subgroup metric does not settle whether the institution should use the tool at all.

This review does not treat AI systems as conscious, divine, AGI, or inevitable. It treats them as institutional machinery: built by people, trained on records, bought through procurement, deployed under incentives, and contestable through law, design, audit, accessibility work, and public pressure. Current claims were rechecked on June 25, 2026 and should be rechecked as rules, standards, and enforcement positions change.

Related Pages

• Algorithmic Bias, AI Audits and Assurance, Algorithmic Impact Assessments, AI System Inventory, AI Red Teaming, Notice and Appeal, Right to Explanation, and Human Oversight of AI Systems turn the book's diagnosis into governance checks.
• AI in Employment, AI in Government and Public Services, AI in Healthcare, Biometric Categorization, AI Data Provenance, AI Post-Market Monitoring, and AI Incident Reporting cover domains and records where exposure, evidence, access, repair, and appeal are central.
• Race After Technology and the New Jim Code, Unmasking AI and the Coded Gaze, Artificial Unintelligence and technochauvinism, Algorithms of Oppression and search authority, Weapons of Math Destruction and scored society, Data Feminism and power-aware evidence, Invisible Women and data gaps, and Sorting Things Out and classification infrastructure supply adjacent book-length frameworks.
• The Audit Society, Trust in Numbers, Claim Hygiene Protocol, Privacy and Data Stewardship, and Vendor and Platform Governance are the operational follow-through: source discipline, data boundaries, procurement accountability, and audits that can actually change a deployment.

Sources

• MIT Press, More than a Glitch, publisher record for title, subtitle, formats, ISBNs, publication dates, page count, award notes, summary, and author biography, reviewed June 25, 2026.
• Meredith Broussard, official author site, author role, affiliations, and book listing, reviewed June 25, 2026.
• Fabian Lütz, review of More than a Glitch, LSE Review of Books, December 28, 2023, reviewed June 25, 2026.
• Melike Asli Sim, review of More Than a Glitch: Confronting Race, Gender, and Ability Bias in Tech, International Journal of Communication 18, 2024, reviewed June 25, 2026.
• New Books Network, Meredith Broussard, More than a Glitch, episode page and summary, May 18, 2023, reviewed June 25, 2026.
• Association of American Publishers, 2024 PROSE Awards finalists and category winners, Popular Science and Mathematics finalist listing, reviewed June 25, 2026.
• NIST, Towards a Standard for Identifying and Managing Bias in Artificial Intelligence, Special Publication 1270, 2022, reviewed June 25, 2026.
• NIST, AI Risk Management Framework, released January 26, 2023 and current resources reviewed June 25, 2026.
• NIST AI Resource Center, AI RMF Core, official govern, map, measure, and manage functions, reviewed June 25, 2026.
• ISO, ISO/IEC 42005:2025, AI system impact assessment, official standard page, reviewed June 25, 2026.
• European Commission AI Act Service Desk, Article 10: Data and data governance, Article 27: Fundamental rights impact assessment for high-risk AI systems, and Article 113: Entry into force and application, reviewed June 25, 2026.
• Office of Management and Budget, M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, official memorandum for U.S. federal AI use, high-impact AI, impact assessment, monitoring, and remedies or appeals, reviewed June 25, 2026.
• FTC, DOJ Civil Rights Division, CFPB, and EEOC, Joint Statement on Enforcement Efforts Against Discrimination and Bias in Automated Systems, April 25, 2023, reviewed June 25, 2026.
• EEOC, iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit, September 11, 2023, reviewed June 25, 2026.
• CFPB, Consumer Financial Protection Circular 2022-03, adverse-action notification requirements for complex algorithms, reviewed June 25, 2026.
• New York City Department of Consumer and Worker Protection, Automated Employment Decision Tools, reviewed June 25, 2026.
• U.S. Department of Justice, ADA.gov, New Rule on the Accessibility of Web Content and Mobile Apps Provided by State and Local Governments, April 8, 2024, reviewed June 25, 2026.
• Federal Register, Extension of Compliance Dates for Title II web and mobile accessibility requirements, April 20, 2026, reviewed June 25, 2026.
• W3C, Web Content Accessibility Guidelines (WCAG) 2.2, W3C Recommendation, reviewed June 25, 2026.
• NIST, Face Recognition Vendor Test Part 3: Demographic Effects, NISTIR 8280, December 19, 2019, reviewed June 25, 2026.

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

• Amazon, More than a Glitch by Meredith Broussard, affiliate listing, reviewed June 25, 2026.

Return to Blog · Return to Books