Blog · Review Essay · Last reviewed June 16, 2026

Dark Matters and the Racial History of Surveillance

Simone Browne's Dark Matters: On the Surveillance of Blackness is a decisive correction to any theory of surveillance that begins with cameras, databases, or the modern security state as if watching were only a technical problem. Browne shows that surveillance is also a racial formation: a way of making some bodies visible, traceable, governable, bordered, and suspect.

The Book

Dark Matters: On the Surveillance of Blackness was published by Duke University Press in 2015. The publisher lists it as a 224-page work in African American studies, Black diaspora studies, cultural studies, surveillance studies, sociology, and social theory. Its archive ranges across the transatlantic slave trade, the slave ship Brooks, Jeremy Bentham's Panopticon, the Book of Negroes, runaway notices, lantern laws, biometrics, cultural production, and post-9/11 airport security.

The book's central move is to refuse a clean origin story for modern surveillance. Browne does not treat racial control as a special case added onto a neutral surveillance apparatus. She argues that practices for identifying, tracking, restricting, counting, displaying, and disciplining Black life are part of the apparatus's formation. Surveillance studies, on this reading, cannot begin with the prison, the camera, the passport, the database, or the platform alone. It has to face the older systems that made racialized bodies available to those later instruments.

That makes Dark Matters especially useful for thinking about AI. Many AI debates still describe harm as bias entering otherwise rational systems. Browne's framework points deeper. The question is not only whether a model has biased data. It is whether the whole pipeline of visibility, classification, suspicion, correction, and control has inherited racial projects while presenting itself as neutral computation.

Racializing Surveillance

The useful definition is precise: racializing surveillance is surveillance that helps produce race as an administrative, visual, spatial, or security category. It does not merely watch people who already belong to a fixed category. It trains institutions to treat some bodies, names, movements, documents, neighborhoods, hairstyles, languages, and associations as signals of suspicion, ownership, mobility risk, fraud risk, or public-order problem.

That definition matters because it keeps the analysis concrete. The problem is not simply that a camera, database, or model has prejudice. The problem is that technical systems can inherit older institutional routines: who is stopped, who is asked for papers, who is made searchable, who is turned into a case file, who must document ordinary life, and who receives the burden of proving they are not the category the system has assigned.

Read this way, Dark Matters is not a metaphor for AI. It is a method for auditing AI's claims of neutrality. Follow the record back to the practices that made it, the institutions that used it, the categories it preserves, and the people who had no equal power to refuse its gaze.

Legibility Before the Database

The book is a study of legibility under force. Branding, passes, notices, manifests, ship diagrams, border documents, and policing practices made people readable to authority in specific ways. That readability was not mere knowledge. It was a condition of capture, sale, restriction, punishment, and administrative control.

This is why Browne's work belongs beside theories of bureaucracy and classification, but also changes them. Legibility is often discussed as the state's attempt to simplify a messy world. Dark Matters makes the racial stakes explicit: some simplifications are produced by making people into categories of danger, property, mobility risk, or bodily evidence. The violence is not only that a system sees badly. It is that the system's way of seeing helps produce the social reality it claims merely to observe.

That point travels directly into model-mediated institutions. Risk scores, identity verification, fraud detection, predictive policing, facial recognition, airport screening, welfare analytics, and hiring tools do not only sort data. They inherit older questions about who must prove innocence, who is made searchable, who is asked for papers, who is treated as an anomaly, and who has the institutional standing to contest the result.

The link to Seeing Like a State is helpful, but Browne makes the frame less abstract. Legibility is not one uniform state project. It is distributed through slave ledgers, border files, police suspicion, commercial identity systems, school records, welfare files, platform logs, and biometric templates. Each record has a politics of who is made visible and who gets to define what visibility means.

Biometrics and the Border

Browne's discussion of biometrics is not a simple claim that all measurement is surveillance. It is a warning about what happens when the body becomes credential, password, risk marker, and administrative object. The promise of biometric systems is that bodies can anchor identity more reliably than documents. The danger is that bodies also become the site where old racial assumptions are reencoded as technical confidence.

Post-9/11 airport and border systems matter here because they combine movement, suspicion, identity, database matching, and discretionary power. A border interface is rarely just an interface. It is a scene where a person becomes a record, a score, a query, a match, a refusal, or an exception. When that process is automated, the human subject may never learn which category mattered, which database answered, or what path exists for repair.

AI intensifies the same structure. A face-recognition model, language classifier, fraud filter, or behavioral detector may treat its output as probabilistic, but institutions often translate probability into action. Once the action is taken, the affected person meets a bureaucracy, not a calibration curve.

Current official sources keep Browne's warning live. NIST's 2019 Face Recognition Vendor Test on demographic effects remains a central government reference for measuring how recognition algorithms can perform differently across demographic groups. The FTC's biometric policy statement likewise treats biometric systems as privacy, security, bias, and discrimination risks, noting that face templates and embeddings are biometric information and that some systems can perform differently across demographic groups. The legal and technical issue is not only accuracy in a lab. It is whether a biometric error or category travels into exclusion, suspicion, arrest, denial, or watchlisting.

Dark Sousveillance

The book is not only an account of domination. Browne also develops dark sousveillance: practices of watching back, refusing capture, rerouting visibility, using performance, art, opacity, counter-records, and collective knowledge to resist racializing surveillance. This is one of the book's most important contributions because it avoids treating watched people as passive objects of systems.

Resistance can mean making surveillance visible. It can mean withholding what the system wants to know. It can mean producing records that contradict official records. It can mean creating art that reveals the politics of vision. It can mean everyday tactics for moving through hostile infrastructures without granting them moral authority.

For AI governance, this matters because many institutional proposals still imagine accountability as something granted from above: audits, dashboards, model cards, explainability reports, compliance forms. Those matter, but Browne's frame asks whether affected communities can watch the watchers, shape the questions, preserve counter-evidence, refuse extractive data demands, and define harms in their own terms.

The AI-Age Reading

The current AI stack is a legibility machine. It turns speech, images, faces, work, movement, preference, social connection, creativity, and administrative history into model-usable signals. It then returns predictions, rankings, recommendations, summaries, flags, denials, and permissions that institutions can treat as knowledge.

Dark Matters sharpens the ethics of that stack. The basic question is not "Does the system see?" It is "What history of seeing does the system continue?" A model trained on institutional records may inherit the institution's prior surveillance. A safety tool may learn from complaint and policing data already shaped by unequal scrutiny. A border system may describe itself as identity management while expanding the number of moments where identity must be proven. A workplace system may call itself productivity software while teaching management to treat workers as streams of measurable behavior.

The book also warns against technological innocence. A system can be new and still participate in old arrangements. It can use modern sensors, cloud databases, embeddings, and neural networks while repeating older logics of suspicion, differential visibility, and racialized administrative control. The interface looks contemporary. The social role may not be.

That is why Browne is useful for AI policy, not only surveillance studies. Bias mitigation is too narrow if it only asks whether outputs are statistically fair. The stronger test asks who was made visible to build the system, who is made accountable to its categories, who can disappear from its view, who can appeal its decisions, and who benefits from the asymmetry between seeing and being seen.

NIST's AI bias guidance supports this broader reading because it treats bias as a sociotechnical problem that can arise across data, design, deployment, and use, with harms possible even without malicious intent. That aligns with Browne's central lesson: racializing surveillance can persist through institutional routine, not only through explicit racist instruction.

Governance and Safety

As of June 16, 2026, official governance has begun to name parts of the problem, though unevenly. The EU AI Act prohibits AI systems that categorize individuals from biometric data to infer race, political opinions, religion, sexuality, and related sensitive traits, while leaving room for specified law-enforcement dataset handling. It also restricts real-time remote biometric identification in publicly accessible spaces for law enforcement to narrow listed objectives, with fundamental-rights impact assessment, database registration, prior authorization in ordinary cases, and limits on adverse decisions based solely on the system output.

Those rules matter because they turn Browne's historical argument into design constraints: do not build systems that infer racial categories from bodies; do not normalize ambient biometric search; do not let law-enforcement exceptions become ordinary infrastructure; do not let a biometric output become punishment without a human, legal, and evidentiary process around it.

The UN Special Rapporteur's 2020 report on racial discrimination and emerging digital technologies gives the human-rights frame: AI and networked technologies can reproduce racial discrimination through design and use, and the responsibility belongs to states and corporations, not only to individual users. That frame is useful for public agencies, vendors, schools, employers, airports, platforms, and police departments because it shifts the burden from "prove the model is racist" to "prove the institution can justify, limit, monitor, and repair the system it deploys."

A practical safety test follows.

Purpose. Is the system necessary for a legitimate purpose, or is it expanding suspicion because automation made watching cheaper?

Data lineage. Which historical records, policing patterns, complaint data, border data, platform data, or identity datasets shaped the system?

Group effects. Has performance been measured across relevant groups and contexts, including intersectional conditions, not only averaged across a benchmark?

Contestability. Can affected people receive notice, challenge the match or category, correct records, obtain human review, and get evidence preserved for appeal?

Community power. Can the communities most exposed to the system shape procurement, deployment limits, audits, data retention, and shutdown triggers?

Non-use. Are there contexts where the system should not be deployed at all because the harm cannot be repaired after the fact?

Where the Book Needs Care

Dark Matters is theoretically dense and intentionally interdisciplinary. Readers looking for a narrow policy checklist will not find one. Browne moves through archives, visual culture, Black feminist theory, sociology, surveillance studies, and cultural analysis. That breadth is the point, but it requires slower reading than a standard technology-policy book.

The book should also not be reduced to a simple analogy engine for every new AI tool. Its historical claims are specific. The point is not that all surveillance is the same, or that every database is equivalent to the archive of slavery. The point is more rigorous: contemporary technical systems are never outside history, and racialized surveillance has been one of the histories through which modern visibility, mobility, identity, and suspicion were built.

Used carefully, that discipline prevents two errors. It prevents technologists from treating racism as a dataset flaw that can be patched after deployment. It also prevents critics from using history as decoration. Browne's work demands that technical critique stay attached to concrete practices: documents, borders, bodies, categories, institutions, and the people made to live under them.

The governance limit is equally important. Anti-surveillance language can become too broad if it ignores genuine needs for evidence, access, safety, and rights enforcement. Some records protect people from disappearance, wage theft, abuse, denial, or state violence. Browne's lesson is not that visibility is always bad. It is that visibility must be accountable to the people placed under it, especially when the institution doing the seeing has coercive power.

What This Changes

The recurring problem across AI interfaces, bureaucratic dashboards, ranking systems, and synthetic publics is not merely automation. It is the conversion of social life into machine-readable form, followed by institutional action that can feel objective because it has passed through a system.

Dark Matters makes that conversion harder to romanticize. It reminds the reader that legibility is not always liberation. Visibility can mean recognition, evidence, and protection. It can also mean exposure, capture, classification, and control. A humane technical politics has to distinguish those conditions instead of treating more data as more justice.

The book's strongest AI-era lesson is that accountability must include counter-visibility. People and communities subject to automated systems need more than explanations delivered after harm. They need the power to inspect, refuse, contest, document, organize, and sometimes remain opaque. Without that power, the next intelligent interface can become an old surveillance relation with a smoother surface.

Source Discipline

This review separates Browne's book claims, current-law claims, technical measurement claims, and interpretation. Duke University Press and the DOI record support book metadata, subject areas, chapter scope, and publisher description. NIST supports claims about face-recognition demographic testing and AI bias as a sociotechnical lifecycle problem. EUR-Lex supports claims about the EU AI Act's biometric categorization and remote biometric identification provisions. FTC supports claims about biometric information, embeddings, commercial risks, bias, and discrimination. The UN digital-library record supports the human-rights source for racial discrimination and emerging digital technologies.

This page does not claim that every biometric or AI system is identical to slavery's archive, and it does not claim that any AI system is conscious, divine, or AGI. It argues that the social function of a system must be audited historically and institutionally, not only technically.

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