Blog · Review Essay · Last reviewed June 25, 2026

Design Justice and the Politics of Community-Led Systems

Sasha Costanza-Chock's Design Justice is a practical theory of design as power. Its AI-era value is direct: systems that classify, recommend, score, tutor, police, hire, care, or govern people must be judged by who shaped them, who can contest them, and whose reality they quietly erase.

Design justice is not a softer word for usability. It is a governance discipline: the people who bear a system's consequences should help define the problem, the evidence, the acceptable risk, the refusal path, the remedy, and the conditions under which the system should not be built.

The Book

Design Justice: Community-Led Practices to Build the Worlds We Need was published by the MIT Press in 2020. The press lists the paperback at 360 pages, with a March 3, 2020 publication date, and also hosts an open-access edition. MIT Press describes the book as an account of design led by marginalized communities, aimed at challenging structural inequality rather than reproducing it.

Costanza-Chock writes as a scholar, designer, and participant in the design justice community. MIT News describes the book as an examination of how technology can work for more people in society, and the MIT Press author note identifies Costanza-Chock as an associate professor of Civic Media at MIT at the time of publication, a faculty associate at the Berkman Klein Center, and creator of the MIT Codesign Studio.

The book is not only about visual design or product design. It treats design as the shaping of worlds: interfaces, institutions, services, built environments, data systems, participation processes, and the defaults that decide who is expected to adapt to whom.

Current Context

As of June 25, 2026, the design justice question has moved from design critique into AI governance infrastructure. Public agencies, standards bodies, and regulators now ask versions of Costanza-Chock's question in legal or procedural form: who is affected, who was consulted, what harms are foreseeable, what evidence supports the system, and what remedy exists when the system fails?

The EU AI Act makes this visible through Article 27 and the Act's implementation timeline. The Commission's AI Act Service Desk says the majority of rules and Annex III high-risk-system rules start applying on August 2, 2026, while full roll-out is foreseen by August 2, 2027. Article 27 requires covered deployers, including public bodies and some private entities providing public services, to perform a fundamental-rights impact assessment before first use of certain high-risk systems. The assessment must describe the deployer's process, period and frequency of use, affected groups, specific risks of harm, human oversight, internal governance, and complaint mechanisms, and it must be updated when relevant elements change. That is not the whole of design justice, but it turns affected people and fundamental rights into required design evidence rather than after-launch sentiment.

In U.S. federal practice, OMB Memorandum M-25-21 requires agencies to manage high-impact AI through documented risk practices, while M-25-22 treats acquisition as a governance problem: agencies need enough transparency, documentation, testing support, privacy controls, accessibility evidence, and risk-management information to meet their obligations after purchase. Canada's Algorithmic Impact Assessment tool is another practical model; the government describes it as a mandatory questionnaire supporting the Directive on Automated Decision-Making, with risk and mitigation questions that determine the impact level of an automated decision system.

Standards now carry the same pressure. ISO/IEC 42005:2025 provides guidance for AI system impact assessments focused on effects on individuals, groups, and society across the lifecycle. NIST's AI Risk Management Framework organizes risk work around govern, map, measure, and manage. W3C's WCAG 2.2, published as a W3C Recommendation in October 2023 and updated in December 2024, remains a concrete reminder that "user" is never a generic category; visual, auditory, physical, speech, cognitive, language, learning, and neurological access must be designed rather than assumed.

The practical shift is this: participation, accessibility, impact assessment, procurement, and appeal are no longer separate ethics decorations. In high-consequence AI systems, they are part of the system boundary. A model, dataset, interface, vendor contract, community process, human reviewer, access channel, and complaint path all decide what the system really is.

The Universal User Problem

The book's central target is the supposedly universal user. Design often claims to serve everyone while building from the standpoint of people with the most institutional, economic, racial, gendered, bodily, and technical privilege. The resulting system can look neutral because its assumptions have been hidden inside convenience.

This is why the book belongs beside work on classification, surveillance, race and technology, accessibility, and legibility. A form field, risk score, app workflow, biometric gate, classroom platform, identity check, public-benefits portal, or automated support bot is never just a surface. It is a theory of who exists, what they need, how they should move, what proof they owe, and how much friction they are expected to survive.

Costanza-Chock's critique of universalism is especially useful because it does not stop at representation. It asks who has actual power in the process. Adding more diverse test users late in the cycle is not the same as letting affected communities define the problem, reject the frame, choose the tradeoffs, or decide that a tool should not be built.

The AI-era version is the universal dataset. A benchmark can hide whose language, body, history, work pattern, disability, paperwork, or neighborhood was treated as normal. A model card can document aggregate performance while leaving the affected person to discover that the workflow cannot handle their case. Design justice pushes the question upstream: whose reality became the schema?

The practical warning is the "edge case" label. Institutions often use it to mean statistically rare, but people experience it as a denial of fit: a name the form cannot hold, a disability the workflow treats as fraud, an address the identity system rejects, a language path that disappears after launch, or a family structure the benefits portal cannot represent. A design justice reading asks when an edge case is actually evidence that the system's model of the public is too narrow.

Process Is Power

The Design Justice Network's principles clarify the book's practical politics. The principles center directly impacted people, prioritize community impact over designer intention, treat change as accountable and collaborative rather than a prize delivered at the end, and describe the designer as a facilitator rather than an expert. They also ask designers to share knowledge and work toward community-led and community-controlled outcomes.

That process language can sound modest until it is applied to high-stakes systems. If a platform, agency, school, hospital, employer, or AI vendor controls the question, the data, the procurement process, the interface, the appeal route, and the public story, then participation may become theater. People are asked to comment on a world that has already been mostly designed around them.

Design justice treats process as part of the system. A tool built through extractive consultation carries extraction into its final form. A tool built through accountable participation is more likely to preserve refusal, local knowledge, repair, and shared ownership.

That means participation has to attach to decision rights. A listening session that cannot change the scope is research, not governance. A community review board without access to procurement records, data provenance, testing results, incident logs, and pause authority is a decorative interface. Process is power only when it can alter the system.

A useful test is whether participation leaves a decision trace. The record should show what affected people changed, what the institution refused to change, why the refusal was made, who is accountable for that choice, and how the issue can be reopened after deployment. Without that trace, co-design becomes a story told by the institution rather than a governed constraint on the system.

The AI-Age Reading

AI systems make the design justice argument harder to ignore. A model-mediated interface can do more than exclude. It can infer, rank, personalize, diagnose, summarize, persuade, and remember. It can transform a design assumption into an automated decision that arrives with the polish of intelligence.

The old universal user becomes the universal dataset, the universal benchmark, the universal risk category, the universal productivity metric, or the universal assistant. The danger is not only bias in model outputs. The danger is a design process that compresses social worlds into machine-readable proxies, then treats those proxies as the natural shape of the task.

Consider an AI hiring system. Design justice asks more than whether the model is accurate. Who defined merit? Who was harmed by previous hiring patterns? What data was used as evidence of success? Can applicants understand, challenge, or refuse the system? What accommodations are available? Which workers inside the institution can slow or stop deployment? Which communities helped decide whether automation belongs in that gatekeeping role at all?

The same test applies to AI tutors, welfare triage, synthetic companions, recommender systems, moderation tools, medical chatbots, policing analytics, and workplace dashboards. The interface may feel helpful, but helpfulness is not accountability. A system can be friendly while concentrating power.

This is also a record problem. Once a model summarizes, scores, routes, or remembers someone, the design assumption can become an institutional record. If the person cannot see the source, correct the label, appeal the route, or force a human review, the system has converted design into durable power while calling it service.

The agentic version widens the circle of affected people. An assistant may draft a case note about a patient, rank a job applicant, flag a tenant, summarize a classroom interaction, or schedule a care task for someone who never touched the interface. Design justice therefore cannot stop at "the user." It must identify the account holder, the operator, the subject of the record, the person affected by the downstream action, and the community whose categories are being normalized.

Governance and Safety

A design justice governance file should start before procurement. It should name the affected groups, who represented them, what authority they had, what disagreements were unresolved, what harms were considered unacceptable, what non-technical alternatives were considered, and what would trigger a decision not to deploy. The file should travel with the system into procurement, launch, monitoring, incident review, and retirement.

For AI systems, the minimum evidence is concrete: problem definition, affected-population map, data provenance, accessibility review, disability accommodations, language-access plan, model or vendor documentation, subgroup testing, human-oversight design, notice text, appeal path, complaint route, incident response, retention rules, and reassessment triggers. This connects directly to algorithmic impact assessments, AI procurement, AI system inventories, and audit trails.

The evidence should also distinguish advisory, consent, veto, and operating authority. Community input that can only advise is different from a board that can delay launch, require a non-automated path, narrow data collection, demand accessibility remediation, publish dissent, or trigger retirement. Procurement should name that authority before a vendor contract makes change expensive.

Safety requires refusal rights, not only feedback. A community should be able to say that a system should not be built, should be narrowed, should preserve a human channel, should remove a data source, should add a language or access path, should delay launch, or should be retired after harm. Without that power, participation becomes a pressure valve for decisions already made.

Accessibility belongs in the same safety file. A public chatbot that cannot be used with assistive technology, a benefits form that fails on mobile, a scoring workflow that ignores disability accommodations, or a medical assistant that cannot handle plain-language or multilingual interaction can produce harm without a dramatic model failure. The exclusion is the failure.

The hard balance is that communities are not unitary. Affected people may disagree. Some may want speed, others privacy; some may want automation, others a human channel; some may want public reporting, others protection from visibility. Design justice does not erase conflict. It gives conflict a governed place before the interface hardens.

Where the Book Needs Care

Design Justice is strongest as a framework and a set of grounded practices. It should not be reduced to a slogan that every technical disagreement can be solved by invoking community. Communities are not simple, unified entities. They contain disagreement, hierarchy, expertise, fatigue, and unequal capacity to participate.

The book is also demanding. Community-led design takes time, money, facilitation, conflict work, institutional humility, and willingness to give up control. Organizations can adopt the vocabulary while keeping the old procurement, management, and liability structures intact. That is a failure of adoption, not a reason to discard the framework.

For AI governance, the point is practical: participation must have teeth. A community review board with no veto, an impact assessment with no remedy, or a feedback form with no obligation to respond can become another interface for absorbing dissent.

There is also a participation burden. The people most affected by a system may already be managing scarcity, disability, surveillance, insecure work, immigration risk, or institutional distrust. A serious process funds participation, protects privacy, offers language and access support, compensates time, and avoids making vulnerable people repeatedly narrate harm for an institution that still keeps final power.

What This Changes

The recurring lesson is that intelligent systems inherit the politics of their design process. A model can only appear neutral when the surrounding institution has already decided which lives, harms, categories, and appeals will count.

Design Justice gives a counter-discipline for the age of fluent machines. Start with the people most affected. Treat lived experience as knowledge, not anecdote. Ask whether the system should exist. Keep power visible. Make refusal and repair real. Judge the design by its consequences, not its intentions or elegance.

That makes the book a useful companion to AI safety and AI governance work that begins too late in the pipeline. The alignment problem is not only inside the model. It is also in the room where the problem was framed, the population was abstracted, the metric was blessed, the vendor was selected, and the affected people were asked to adapt.

The site returns often to legibility because systems that make people readable also decide what kind of person can be read. Design justice adds the missing verb: communities should help decide what forms of legibility are acceptable, what must remain uncollected, what records must be correctable, and when opacity protects people better than measurement. That is the bridge to contextual integrity: a data flow can be technically efficient and still violate the social context that gave the information meaning.

Source Discipline

This review separates the book's argument from current legal and standards claims. MIT Press, MIT Press Direct, MIT News, and the Design Justice Network support claims about the book, the author context, the open-access edition, and the principles. Current claims about impact assessments, procurement, accessibility, and AI governance rely on official public sources: EU AI Act materials and timeline pages, OMB memoranda, Canada's AIA tool, ISO, NIST, and W3C.

Design justice should not be cited as if it automatically proves a system unsafe, unlawful, or illegitimate. It is a framework for asking who held power in the design process and what evidence, remedies, and refusal rights followed. A vendor claim of "co-design," a workshop photo, a user-research summary, or a public comment period is evidence only of that activity, not proof that affected communities controlled the outcome.

This page makes no claim that any AI system is conscious, divine, or AGI. The power at issue is institutional: systems can classify, route, remember, and exclude people through designs that become hard to contest after deployment.

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