Blog · Review Essay · Last reviewed June 19, 2026

Race After Technology and the New Jim Code

Discriminatory technology is rarely just a matter of bad data, biased coders, or insufficient representation, and Ruha Benjamin's Race After Technology is one of the clearest short books for seeing why. It traces social hierarchy becoming infrastructure: older racial arrangements rewritten as apps, scores, filters, databases, defaults, and cheerful products that claim to be neutral.

The New Jim Code, in this review, means more than biased software. It names an institutional pattern: racial hierarchy translated into technical systems through data collection, objectives, thresholds, interfaces, procurement, and feedback loops.

The practical test is exposure and recourse. Who is made machine-readable? Who is acted on? Who can see the evidence? Who has authority to correct, refuse, or appeal the classification?

The Book

Race After Technology: Abolitionist Tools for the New Jim Code was published by Polity in 2019. Benjamin's author page lists the subtitle, publisher, year, and several honors, including the 2020 Oliver Cromwell Cox Book Award, a 2020 CITAMS honorable mention, and the 2020 Brooklyn Public Library Literary Prize for Nonfiction. Academic listings in Social Forces, Nature Machine Intelligence, Cultural Sociology, and New Media & Society identify the book as a 2019 Polity Press title.

The book's importance is not that it was first to notice algorithmic bias. Its strength is that it refuses the thin version of the problem. Benjamin is not mainly asking whether a model contains prejudice as an accidental contaminant. She is asking how technology can inherit racial order while appearing modern, efficient, objective, innovative, and even humane.

That shift matters. If bias is only a bug, the cure looks like better training data, more diverse teams, technical audits, and fairness metrics. Those can matter. But Benjamin's argument is larger: the institution deciding what to build, whom to monitor, what to predict, whose pain counts as evidence, and what counts as progress may already be carrying the old arrangement into the new tool.

That is also why the book still reads as current in 2026. The central risk is not that one old classifier was unfair. It is that organizations keep turning contested social judgments into scalable decision infrastructure, then asking audit tools to clean up a deployment choice that should have been questioned earlier.

The New Jim Code

Benjamin's central phrase, the New Jim Code, names technologies that reproduce or deepen racial hierarchy while presenting themselves as race-neutral or benevolent. The phrase is a deliberate echo of Michelle Alexander's The New Jim Crow, her 2010 study of how mass incarceration rebuilt a racial caste system through ostensibly colorblind law. Benjamin's claim is that the same continuity now runs through code: each era keeps the hierarchy while upgrading the machinery that enforces it, and the latest machinery presents itself as the least biased of all. The Ohio State University research guide summarizes the point directly: automation can hide, accelerate, and intensify discrimination while appearing neutral compared with older forms of racism.

The phrase works because it holds two ideas together. "Code" means software and technical rule-making, but it also means the social codes that teach institutions whom to suspect, serve, exclude, protect, classify, or ignore. The system does not need an explicitly racist instruction if its inputs, categories, objectives, procurement logic, and deployment setting already encode the hierarchy.

This is why the book belongs near work on legibility and classification. A person becomes a record; the record becomes a category; the category becomes a risk score, ranking, alert, denial, or automated suspicion. The social fact is then returned to the world as if it were a technical finding.

The phrase is useful because it rejects a false choice: either explicit racist intent or neutral technology. Many harmful systems sit between those poles. They inherit proxies, missing data, uneven surveillance, historical labels, and institutional incentives, then produce outputs that look cleaner than the world that made them.

Classification and Exposure

Princeton's coverage of the book lists the kind of examples Benjamin uses: gang databases with overwhelmingly Black and Latinx entries, automated beauty judging that selected almost entirely white winners, recidivism risk scoring that misclassified Black defendants at higher rates, and seemingly small but revealing language glitches such as a map system misreading Malcolm X Boulevard.

The recidivism example is the most studied of these. It refers to COMPAS, the risk-scoring tool examined in ProPublica's 2016 "Machine Bias" investigation, which compared scores for more than ten thousand defendants in a Florida county against who actually reoffended. The headline finding was not that the tool was less accurate for Black defendants; overall accuracy was roughly equal. It was that the errors fell along the color line. Among people who were not rearrested, the algorithm had wrongly flagged Black defendants as future criminals at nearly twice the rate of white ones, about 45 percent against 24 percent. A system can be evenhanded in the aggregate and still distribute its mistakes by race.

These examples are not all the same technically. That is part of the point. Discrimination can appear through training data, label choices, feedback loops, institutional use, user-interface defaults, surveillance intensity, error tolerance, and unequal exposure to experimental systems.

Exposure is the crucial word. Technological harm is not evenly distributed. Some people meet automation first as convenience: recommendation, speed, personalization, novelty. Others meet it first as suspicion: fraud detection, policing, welfare eligibility, border screening, school discipline, workplace monitoring, tenant scoring, or biometric misrecognition.

A system can therefore be accurate enough to sell and still unjust enough to govern badly. The test is not only benchmark performance. It is who is placed inside the machine's jurisdiction, who can contest its output, and who is forced to live with its mistakes.

Biometrics make the point sharply. NIST's face-recognition evaluations describe demographic differentials for contemporary algorithms, and NIST's 2019 study warned that performance depends on the algorithm, application, and data. That matters because a benchmark is not a deployment permission slip. Identification, verification, watchlist search, access control, and police investigation place different burdens on the people who are misclassified.

Benevolent Interfaces

One of Benjamin's most useful moves is her attention to technological benevolence. Harmful systems do not always arrive with hostile branding. They arrive as safety, efficiency, health, personalization, modernization, fraud prevention, child protection, beauty, convenience, or inclusion.

The Institute for Advanced Study excerpt from the book uses the Beauty AI contest to show the pattern. A project framed as technical novelty and health-oriented assessment produced racially skewed results, then made those results look like machine judgment rather than social preference hardened into a model.

This is a general interface problem. A clean design can launder a dirty social process. A dashboard can make coercion look administrative. A score can make a contested judgment look measured. An AI answer can make a ranking decision look like knowledge. A companion interface can make asymmetrical data extraction feel like care.

Benjamin helps readers see that the moral danger often lies in the friendliness of the system. The more benevolent the interface feels, the less likely users are to ask who was classified, excluded, watched, or made available for intervention in order for that benevolence to appear.

The AI-Age Reading

The AI-era reading of Race After Technology is straightforward: foundation models, agents, AI search, workplace copilots, hiring filters, police analytics, health triage tools, and automated public services do not escape social history because their outputs are generated by statistical systems.

Shakir Mohamed's review in Nature Machine Intelligence makes that connection explicit, placing the book inside machine-intelligence work and noting examples across healthcare, policing, welfare, dating, and hiring. The point is not that every model is equally harmful. The point is that AI systems are deployed into already unequal worlds and can make those worlds easier to administer without making them more just.

For AI governance, Benjamin's book is an antidote to a narrow safety culture. It asks questions that capability discourse often skips. Who is the default user? Who is the default suspect? Which communities become test beds? Whose data is treated as public raw material? Who has appeal rights? Who is represented in the design meeting, and who is represented only as a data point?

Those questions do not replace technical evaluation. They make it honest. A model card that describes performance but not institutional purpose is incomplete. An audit that reports error rates but not exposure patterns is incomplete. A fairness metric that ignores power over deployment is incomplete.

Generative AI adds a new surface without changing the basic problem. A system that drafts, ranks, summarizes, screens, or searches can turn racialized records into fluent defaults: whose dialect is "corrected," whose name is treated as anomalous, whose neighborhood is over-risked, whose history is summarized as a problem, or whose complaint is down-ranked because it does not resemble the training examples privileged by the institution.

Current Context

As of June 19, 2026, the governance environment has moved closer to Benjamin's diagnosis. NIST Special Publication 1270 treats AI bias as a lifecycle and sociotechnical problem, not merely a statistical defect, and the NIST AI Risk Management Framework gives organizations a govern, map, measure, and manage structure for risks to individuals, organizations, and society.

The EU AI Act makes related duties concrete for high-risk systems. Article 10 requires data governance practices for training, validation, and testing data, including examination and mitigation of biases likely to affect safety, fundamental rights, or anti-discrimination law. Article 27 requires fundamental-rights impact assessments for certain deployers of high-risk AI systems, including descriptions of use, affected groups, risks, human oversight, and mitigation. The Act applies on a phased timeline, so claims about it should stay jurisdiction-specific and date-specific.

U.S. agencies have also rejected the idea that automation creates a civil-rights vacuum. The 2023 joint statement from the FTC, DOJ Civil Rights Division, CFPB, and EEOC said existing agency authorities apply to automated systems. The EEOC's iTutorGroup settlement showed the point in a simple form: software that allegedly rejected older applicants was treated as an employment discrimination matter, not as a technical accident. In credit, CFPB Circular 2022-03 says creditors using complex algorithms still must give specific and accurate adverse-action reasons.

Local rules and audits add another layer. New York City's automated employment decision tool rule requires bias audits and candidate notices for covered employment tools, but it also illustrates Benjamin's warning: audit law is only as strong as its definitions, enforcement, documentation, and paths for affected people to challenge the result.

Governance and Safety

The governance implication is that a discrimination review must start before model evaluation. A serious review names the decision context, affected groups, data provenance, labels, proxies, thresholds, model version, vendor role, human workflow, appeal path, and feedback loop. It asks not only whether the system is accurate, but who bears the cost of false positives, false negatives, low-quality service, exclusion, surveillance, or representation as risk.

Subgroup and intersectional testing matter, but average accuracy is not enough. The same error rate can have different consequences in bail, hiring, school discipline, benefits eligibility, clinical triage, fraud detection, and content moderation. Governance has to specify the harm, the comparator, the decision authority, and the remedy.

The safety rule is practical: no consequential classification should operate without notice, an evidence record, a human with authority to change the outcome, vendor obligations that survive procurement, and post-deployment monitoring for drift and disparate impact. In some settings the right answer is not a fairer model, but a narrower use, a non-automated process, or a decision that should not be scored at all.

Where the Book Needs Friction

The book is short, accessible, and deliberately broad. That makes it powerful as an entry point, but it also means some technical mechanisms receive less detail than an engineering reader may want. Predictive policing, computer vision, automated scoring, data infrastructure, user-interface design, and platform ranking each have different failure modes.

That distinction matters because good governance needs both social diagnosis and implementation detail. A database error, a biased training distribution, an invalid proxy, a badly scoped objective, a procurement incentive, and an abusive deployment context may require different interventions.

Still, this is not a serious weakness of the book's argument. It is a boundary. Benjamin is giving readers a political and ethical grammar, not a compliance checklist. The responsible next move is to bring that grammar into audits, procurement rules, participatory design, appeal systems, civil-rights enforcement, and technical practice.

A second boundary is evidentiary. A vivid example can reveal a pattern, but it should not be stretched into a universal claim about every tool in a category. The stronger move is to use the example to ask what evidence would be needed in the specific deployment: data lineage, validation study, subgroup performance, decision logs, notice language, complaint records, and proof that someone can actually change a bad outcome.

What This Changes

Race After Technology is a guide to how recursive reality becomes racialized.

A classification system does not only describe people. It changes what institutions can see, what they can ignore, what they can automate, and what they can justify later. Once those outputs circulate, they shape the next round of records, decisions, incentives, and beliefs. The loop can make an imposed category look natural because the world has been reorganized to confirm it.

That is why civil-rights literacy belongs inside AI literacy. The question is not only whether machines think, reason, align, or hallucinate. It is whether machine-mediated institutions are building a world in which some people are more searchable, more punishable, more predictable, more exposed, and less able to contest the reality being assigned to them.

Benjamin's practical demand is not despair. It is design discipline joined to political imagination: refuse neutral theater, inspect defaults, follow exposure, build appeal, include affected communities before deployment, and treat technical systems as arrangements of power. The machine is never outside the society that trains, funds, buys, and believes it.

That makes the book a companion to this site's recurring concern with machine-readable reality. When records feed models, models feed decisions, decisions create records, and those records train or justify the next system, social order becomes recursive. Benjamin's contribution is to insist that racial power is not a side issue in that loop. It is often one of the oldest patterns the loop is asked to reproduce.

Source Discipline

Use sources for the level of claim they can support. Benjamin's author page and academic reviews establish book metadata and reception. ProPublica's COMPAS investigation supports a specific 2016 finding about one tool, data set, jurisdiction, and error pattern; it should not be cited as proof that all risk assessment tools behave identically. NIST's biometric tests support claims about tested algorithms and test conditions, not blanket claims about every real-world deployment.

Regulatory sources are also bounded. The EU AI Act, EEOC, CFPB, FTC, and New York City materials show enforceable or official governance expectations in particular jurisdictions and domains. They do not prove that a given vendor is compliant, fair, or safe. They are useful because they identify the questions a responsible deployment has to answer in public language: data governance, discrimination, adverse-action reasons, impact assessment, notice, oversight, and remedy.

This review does not treat AI systems as conscious, divine, 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, and public pressure.

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