Blog · Review Essay · Last reviewed June 25, 2026

Sorting Things Out and the Politics of Classification

Geoffrey C. Bowker and Susan Leigh Star's Sorting Things Out is one of the best books for seeing what disappears behind a clean category. For AI readers, its lesson is direct: every classifier, dataset, dashboard, risk score, and institutional workflow inherits decisions about what counts, who fits, and whose life becomes difficult when the system needs sharper boxes than reality provides.

For this review, classification means the maintained system of categories, codes, thresholds, labels, and rules that lets an institution recognize something as a case it knows how to process. A classification is not only a word. It is a route through records, labor, authority, money, care, suspicion, and appeal.

The practical question is not whether categories should exist. Shared categories make public health, benefits, science, disability access, safety reporting, and institutional memory possible. The question is whether the category has a source, owner, update process, appeal path, and evidence record before it becomes the hidden grammar of a machine decision.

The Book

Sorting Things Out: Classification and Its Consequences was published by The MIT Press in hardcover on September 29, 1999 and in paperback on August 25, 2000. MIT Press lists the paperback ISBN as 9780262522953, the hardcover ISBN as 9780262024617, and the book at 392 pages. MIT Press Direct lists DOI 10.7551/mitpress/6352.001.0001 and electronic ISBN 9780262269070.

Bowker and Star were central figures in science and technology studies, infrastructure studies, and the sociology of information systems. MIT Press frames the book as an investigation of how categories and standards shape worldviews and social interaction. That publisher summary is accurate, but the book is more concrete than the phrase suggests: it follows categories into archives, hospitals, death records, work classifications, race systems, and standards bodies.

The cases range across mortality records, the International Classification of Diseases, the Nursing Interventions Classification, apartheid race classification in South Africa, virus taxonomy, and tuberculosis. That range is the point. Classification is not a narrow library problem. It is how modern institutions make work, illness, identity, death, risk, and responsibility travel through records.

The book belongs beside work on legibility, surveillance, software, and cybernetics because it studies the quiet layer underneath them: the categories that let a system recognize a person or event as something it knows how to process.

Current Context

As of June 25, 2026, classification is no longer only an information-science topic. It is a live governance requirement. The World Health Organization says all Member States are committed to using the most recent version of the International Classification of Diseases, and that 2022 was the first year ICD-11 was officially in effect. The book's ICD case therefore remains current: medical classification is not static background metadata. It is maintained global infrastructure for mortality statistics, care records, reimbursement, research, and public health comparison.

The EU AI Act makes the category question explicit for high-risk AI systems. Article 10 requires training, validation, and testing datasets to be governed according to intended purpose, including design choices, collection processes, data origin, preparation operations, assumptions, bias examination, and gap identification. Article 13 requires high-risk systems to be sufficiently transparent for deployers to interpret outputs and use the system appropriately. Article 27 requires specified deployers to assess affected groups, risks, human oversight, governance, and complaint mechanisms before first use. These are classification controls in legal form: the law asks where the data categories came from, what they assume, who they affect, and whether the institution can explain and contest their use.

U.S. federal policy points in the same direction. OMB M-25-21 requires agencies to complete AI impact assessments before deploying high-impact AI use cases, including review of the intended purpose, expected benefits, data and model quality, potential impacts, safeguards, and residual risk acceptance. NIST's AI Risk Management Framework is voluntary, but it names risks to individuals, organizations, and society and treats bias, data representativeness, explainability, privacy, and accountability as parts of trustworthy AI. A classifier is not governed merely because it is accurate on a benchmark; it is governed when its categories, evidence, limits, and remedies are visible enough to challenge.

W3C's PROV work gives the technical vocabulary behind that demand. PROV defines provenance as information about entities, activities, and people involved in producing a data item or thing so that quality, reliability, and trustworthiness can be assessed. In Bowker and Star's terms, provenance is how the category stops pretending to be natural. It shows who made the box, what work it did, and what history it carries.

Classification as Infrastructure

The core move of Sorting Things Out is to treat classification systems as infrastructure. A category is not only a label. It is a routing device. It tells a hospital what happened, a database where to place a record, a government what benefits or restrictions apply, an employer what counts as work, and a model what features can be learned.

Infrastructure is powerful because it fades into the background. People notice roads when they break, forms when they exclude them, medical codes when insurance denies a claim, and identity categories when the system cannot hold their life without distortion. Bowker and Star ask readers to study those background systems before they become naturalized as common sense.

This is a useful correction to shallow technology talk. Software often presents itself as a neutral layer over reality. Classification shows that the layer is already political before the first line of automation runs. A database schema, risk band, eligibility code, demographic field, incident taxonomy, content label, or safety severity scale has already decided what the later system can notice.

The stronger formulation is this: classification turns the continuous world into institutional states. Once something has a state, it can be counted, routed, priced, ranked, denied, escalated, audited, or ignored. That is why classification belongs with AI data provenance, algorithmic transparency, opaque scoring systems, and algorithmic impact assessments.

The Work of Invisibility

Bowker and Star are especially good on invisibility. Categories feel obvious when they work for the people who designed them or the institutions that depend on them. They become visible to people who do not fit, to workers who maintain the scheme, and to anyone forced to translate messy life into official terms.

That matters because invisibility is not the same as neutrality. A standard can hide labor. A code can hide discretion. A database field can hide a moral decision. A required choice can hide the person whose answer is "none of the above." The more smoothly the infrastructure operates, the easier it is for outsiders to miss the damage it imposes at the edges.

The authors' method is patient rather than theatrical. They look at archives, standards, work practices, and institutional histories. That patience is part of the book's strength: it teaches readers to distrust the glamour of the finished interface and look for the negotiations buried in the form.

In AI systems, invisibility often moves one layer deeper. The visible label may be "risk," "fraud," "priority," "unsafe," "high confidence," "match," "eligible," or "low quality." The hidden work sits in the training labels, annotation guide, default taxonomy, data-cleaning rule, threshold, feedback loop, and business process that makes the label actionable. A serious review asks for those buried materials, not only for the final prediction.

Classification and Suffering

The book's moral force comes from showing that classification produces consequences. The same scheme that helps institutions coordinate can also produce administrative suffering when a person's reality is misread, flattened, delayed, or made illegible.

The apartheid case makes this explicit, but the quieter medical and work-practice cases matter too. A disease category affects recognition, funding, stigma, treatment, and research. A nursing classification can make care labor visible to administrators, but it can also reshape the work by privileging what can be coded. Classification can rescue knowledge from invisibility and impose a new invisibility at the same time.

That double nature is why the book is more useful than a simple anti-bureaucratic complaint. Bowker and Star do not ask readers to abandon categories. They ask readers to notice who pays for them, who can change them, and who is treated as an exception to be managed rather than a signal that the system is incomplete.

That is the governance lesson. A category should be treated as a claim about the world, not as the world itself. When people repeatedly fail to fit, the institution should not only create a workaround. It should ask whether the classification is harming the people it claims to organize.

The AI-Age Reading

Artificial intelligence makes classification faster, cheaper, more ambient, and easier to detach from accountability. A model can classify applicants, patients, students, workers, neighborhoods, images, emotions, texts, threats, fraud signals, support tickets, and user intent. Once the classification exists, another system can act on it.

The hard problem is not only accuracy. Accuracy asks whether the model placed a case in the expected box. Bowker and Star push the deeper question: should this box exist, who made it, what history does it carry, and what happens to the person who cannot live cleanly inside it?

This is where classification becomes recursive reality. A label changes treatment. Changed treatment changes behavior. Changed behavior becomes new data. The new data validates the label. In credit, policing, welfare, workplace management, education, and platform moderation, a weak category can harden into social fact when institutions keep feeding its consequences back into the system.

AI governance therefore needs category review, not just model review. Dataset labels, target variables, risk bands, incident taxonomies, safety thresholds, user segments, and escalation categories are all sites of power. They need appeal paths, maintenance owners, affected-community review, public documentation, and a way to retire categories that have become harmful.

This matters for generative and agentic systems too. A chatbot may appear open-ended, but it still classifies requests, sources, users, safety risks, tool calls, policy violations, and support intents. An agent can act only after some schema has made the world machine-addressable: task state, account role, permission, merchant, file type, priority, confidence, or exception. The interface sounds fluid because the classification work has moved into prompts, tool schemas, retrieval indexes, and hidden policy layers.

Governance and Safety

The practical safety issue is classification without recourse. A category can deny a benefit, trigger a fraud review, lower a worker rating, route a patient, prioritize a police lead, downrank a post, or mark a child as at risk. If the affected person cannot know the category, see the source data, challenge the fit, or trigger correction, the institution has converted a maintained classification into private fate.

Good governance begins before procurement or model training. Buyers and builders should ask what taxonomy the system uses, who authored it, which domain experts and affected groups reviewed it, whether the categories are legally or clinically meaningful, how uncertain or mixed cases are handled, and whether local knowledge can override central fields. Those questions belong in AI procurement, AI system inventories, audit trails, human oversight, and algorithmic recourse.

Safety controls should distinguish classification error from category harm. A model can assign the intended label and still harm people if the label is illegitimate, too coarse, stigmatizing, historically loaded, or used outside its original purpose. A dataset can be statistically representative and still encode a category that should not determine access to housing, work, healthcare, education, or public services.

The system should also preserve disagreement. Some records need uncertainty, multiple codes, patient or worker testimony, local notes, exception handling, and escalation paths. Forcing every case into one clean field can make the database easier to query while making the institution worse at reality.

The Category Record

The concrete artifact is a category record: a maintained file for each consequential category, label set, target variable, risk band, or taxonomy used in a high-impact system. It should be short enough to maintain and precise enough to audit.

This record should connect to data provenance, model and system cards, AI bills of materials, data retention, change management, and public registers. Otherwise classification remains the most powerful part of the system and the least documented.

Where the Book Needs Friction

Sorting Things Out is rich, but it can be demanding. It is less a clean manifesto than a dense study of infrastructure, standards, and practice. Readers looking for a simple theory of classification will find something more complicated: cases that keep disturbing any one-sentence rule.

That is also why the book should not be turned into a reflexive suspicion of all standards. Shared categories can support public health, disability access, labor recognition, scientific coordination, rights claims, safety reporting, and democratic accountability. The question is not whether classification should exist. The question is how classifications are built, contested, revised, and made accountable to the people they sort.

Academic reception has treated the book as a major contribution to information infrastructure and science studies. Reviews in Public Understanding of Science, Issues in Science and Technology Librarianship, College & Research Libraries, and Technology and Culture placed it in conversation with information science, user studies, knowledge architecture, and the social consequences of standards. Stefan Helmreich's later essay on the race-classification chapter also shows how seriously scholars continued to work with its claims.

The book also predates contemporary machine-learning pipelines, data brokers, platform moderation at global scale, and current AI law. Its vocabulary needs operational updating: classification politics now appears in annotation contracts, prompt taxonomies, system-card limitations, high-risk AI databases, impact assessments, and agent tool schemas. That does not weaken the book. It makes the book's slow infrastructure method more necessary.

What This Changes

Sorting Things Out is a book about the hidden grammar of institutional reality.

Before a model predicts, a database has already named. Before an agent acts, a workflow has already decided what counts as a valid state. Before a dashboard turns red, someone has defined the threshold. The machine appears intelligent because the world has been prepared for machine recognition.

The practical lesson is category hygiene. Name the categories. Date them. Source them. Ask who maintains them. Track who is harmed by them. Build appeals for people who do not fit. Preserve local knowledge when central systems demand clean fields. Treat classification errors as institutional evidence, not merely user friction.

Bowker and Star's lasting warning is that sorting is never only sorting. It is world-making at administrative speed. In the AI era, that speed increases the obligation to inspect the boxes before they become destiny.

Source Discipline

This review separates book metadata, publisher description, secondary academic reception, current classification infrastructure, current AI governance claims, and this site's interpretation. MIT Press and MIT Press Direct support the bibliographic claims. WHO supports the ICD current-context claim. The EU AI Act Service Desk, OMB, NIST, W3C, and ISO support the governance context checked on June 25, 2026.

Claims about law and standards are scope-bound. EU AI Act Articles 10, 13, and 27 apply in specific ways to high-risk AI systems and specified deployers; OMB M-25-21 applies to covered U.S. federal agencies; NIST AI RMF and ISO/IEC 42005 are standards or guidance, not universal law; WHO's ICD is a health classification infrastructure, not a general theory of social categories.

This page does not claim that any AI system is conscious, divine, or AGI. It treats AI systems as institutional systems that classify, route, and act through records. The governance issue is whether the categories that make action possible can be inspected, corrected, revised, and retired.

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