Blog · Review Essay · Last reviewed June 25, 2026

The Social Life of Information and the Context Around the Machine

John Seely Brown and Paul Duguid's The Social Life of Information is an antidote to the fantasy that information becomes more powerful as it becomes more detached. Its AI-era lesson is direct: a system can move symbols, summarize records, and automate workflow while still missing the social practice that makes knowledge usable.

For this review, the social life of information means the roles, institutions, communities of practice, trust relations, maintenance work, and informal judgment that let a record become usable knowledge. The unit of analysis is not the document, prompt, model, or database alone. It is the whole setting that makes a claim meaningful and accountable.

The practical test is a context ledger: before deploying a knowledge base, RAG system, copilot, agent, tutor, triage bot, or dashboard, name the source community, the work practice being represented, the authority of each record, the human repair path, the affected people, and the limits on reuse. If those cannot be named, the system is not yet carrying knowledge; it is carrying decontextualized information.

The Book

The Social Life of Information was written by John Seely Brown and Paul Duguid and first published by Harvard Business School Press in 2000. UC Berkeley's School of Information lists the original publication as March 2000. The current Harvard Business Review Press edition, updated with a new preface, was published on March 14, 2017 and runs 336 pages.

Brown was long associated with Xerox PARC, and Duguid is a historian and social theorist affiliated with Berkeley's School of Information and Xerox PARC. That institutional background matters. The book is not an abstract manifesto against computers. It comes from people who spent time close to the places where networked computing, office automation, and knowledge-work fantasies were being built.

The book's target is information reductionism: the belief that once information can be captured, transmitted, searched, packaged, and personalized, the surrounding social structures become obsolete. Brown and Duguid argue that context, practice, trust, institutions, communities, and informal knowledge do not disappear when information becomes digital. They become easier to overlook.

Current Context

As of June 25, 2026, the book reads less like a dot-com-era correction and more like a handbook for AI deployment. The word "context" now appears in technical phrases such as context windows, retrieval context, memory, vector search, tool context, and prompt context. Brown and Duguid's warning is that those are not the same as social context. A model can receive many tokens and still lack the institutional knowledge, role boundaries, incentives, exceptions, and repair practices that make the tokens reliable.

The current governance vocabulary is catching up to that distinction. NIST's AI RMF Core organizes risk work around govern, map, measure, and manage functions, and NIST's Generative AI Profile asks organizations to treat data provenance, human oversight roles, source verification, citations, anthropomorphization, training data, evaluation data, and retrieval-augmented generation approaches as reviewable system evidence. ISO/IEC 42001 similarly frames AI governance as an organizational management system, not only a model-quality exercise.

Labor and high-risk AI rules point in the same direction. The U.S. Department of Labor's October 2024 AI best-practices roadmap emphasizes meaningful human oversight for significant employment decisions, worker input, rights protection, training, and protection of worker data. EU AI Act Article 10 requires high-risk AI data governance to consider data origin, original collection purpose, assumptions, bias, and the specific context of use; Article 14 requires human oversight that lets people monitor, interpret, override, or stop high-risk systems as appropriate.

The live issue is therefore not whether AI can retrieve more information. It is whether institutions can preserve the social conditions under which that information was produced and should be used: who knows the exception, who owns the record, who can challenge an automated summary, who has authority to act, and who is harmed when the system treats a brittle representation as the whole reality.

Tunnel Design

The book's central warning is against tunnel design. A tunnel designer sees the informational part of an activity and assumes the rest is waste. In that frame, a university is content delivery, an office is document routing, a company is process execution, a library is retrieval, a community is a contact list, and a worker is a bundle of codified tasks.

This diagnosis has aged well because the same mistake keeps returning with better interfaces. A platform can treat journalism as content units, teaching as video plus assessment, medicine as records plus triage, governance as services plus identity, and social life as messages plus graphs. The machine-readable part is real. The error is treating it as the whole activity.

The useful phrase for AI is context loss. When a workflow is compressed into prompts, tickets, embeddings, dashboards, summaries, and generated answers, the system may preserve the visible text while losing the background practices that made the text meaningful. A decision can look cleaner precisely because the messy knowledge has been stripped away.

Tunnel design is especially tempting in retrieval-augmented generation. A RAG system can cite a policy, contract, note, ticket, or manual and still miss whether the source is current, locally authoritative, superseded by custom, limited by jurisdiction, contested by workers, or known to fail in practice. The citation proves that a fragment entered the prompt. It does not prove that the system understood the room from which the fragment came.

Agents and Context

One reason the book belongs in this catalog is that it was already thinking about software agents. The Idaho State Archives catalog lists "Agents and angels" among the book's chapters, and Brown and Duguid were writing at a time when digital agents were imagined as delegated helpers that could search, filter, broker, and act on behalf of users.

That early agent discourse now looks familiar. AI assistants promise to read for us, schedule for us, buy for us, negotiate for us, summarize for us, and remember for us. The practical question is not only whether they can perform the task. It is what social knowledge they need to perform it without damaging the relationships, obligations, norms, and tacit commitments around the task.

A calendar entry is not just time. A meeting can carry rank, trust, fatigue, obligation, avoidance, alliance, politics, and care. A procurement request is not just a transaction. A school assignment is not just content. A medical note is not just text. Delegation to agents becomes risky when the agent sees the informational surface and misses the social situation.

That is why agent governance cannot stop at permission scopes. A useful agent record has to name the principal, the affected third parties, the tools exposed, the authority delegated, the corpus retrieved, the memory used, the escalation rule, and the receipt left behind. The system should not be able to convert an ambiguous social request into a completed institutional action without showing which context authorized the move.

Practice Before Process

The strongest chapters are about practice, learning, and organizations. The book argues that knowledge lives in use: in repair habits, apprenticeship, communities of practice, shared judgment, institutional memory, unofficial workarounds, and the background skills people use to make formal processes actually function.

That matters for labor. When managers see only process, they try to optimize the diagram. When they see practice, they have to ask how work is learned, how errors are caught, how novices become competent, how experienced workers notice trouble, and how informal coordination keeps brittle systems from failing.

AI can support practice when it gives people better access to memory, examples, translation, explanation, and coordination. It can also hollow practice out when it converts workers into prompt operators, exception handlers, compliance performers, or monitored endpoints in a system whose real knowledge has moved into proprietary models and managerial dashboards.

The difference is whether the system strengthens a community of practice or extracts from it. A good tool gives workers better examples, preserves corrections, exposes uncertainty, and lets practitioners improve the representation. A bad tool captures their tacit knowledge, hides the model and metrics, and then uses the cleaned-up output to discipline the people whose practice made the system usable.

The AI-Age Reading

Read in 2026, The Social Life of Information is a check on model-centered thinking. The temptation is to ask what the model knows. Brown and Duguid push the better question: what social setting lets any information become knowledge?

Large language models are powerful because they can operate across many symbolic surfaces: code, prose, policy, contracts, transcripts, tickets, textbooks, chats, and search results. But symbolic range can hide situational thinness. A model can produce a plausible answer without knowing which local rule is unofficially decisive, which metric has been gamed, which relationship is strained, which exception carries moral weight, or which institutional memory has never been written down.

This is where recursive reality enters. AI systems do not merely extract information from institutions. They can feed summaries, categories, recommendations, and synthetic explanations back into the institutions that will later become training data, policy evidence, audit records, and ordinary memory. If context is lost on the first pass, the cleaned-up version can return as the new official past.

The governance problem is therefore not limited to hallucination. A system can be source-grounded and still flatten the social setting. It can cite the file and miss the practice. It can summarize the meeting and erase the hesitation. It can classify the case and lose the relationship that made the case intelligible.

The book also keeps AI evaluation honest. A benchmark can test whether a model retrieves an answer. It often cannot test whether the system respects the social role of the source, the authority of the author, the needs of the affected person, or the repair channels of the institution. Those are not soft extras. They are the conditions under which information becomes knowledge rather than noise with formatting.

Governance and Safety

The governance implication is to treat social context as system infrastructure. A deployment review should not ask only what model is used, what corpus is indexed, and what accuracy score was achieved. It should ask what work practice the system represents, who maintains that practice, what informal exceptions matter, which records are authoritative, and how affected people can correct the automated version of events.

For institutional systems, the minimum record should connect the AI use to an AI system inventory, data provenance, audit trail, model or system card, and human oversight plan. The inventory names the system; provenance names where information came from; the audit trail shows what happened; the card states intended use and limits; oversight names who can intervene and repair.

For workplace and knowledge-work tools, the safety controls should include worker consultation, role-specific access, retrieval permission checks, source freshness labels, human review for consequential outputs, correction loops, appeal paths, and limits on turning practice traces into surveillance or performance scoring. "AI learned from the team" is not a governance claim unless the team can see, correct, refuse, and contest what was learned.

For agents, the context ledger should travel with action. A completed booking, message, purchase, edit, triage, recommendation, or case update should preserve the initiating user, tool permissions, retrieved sources, relevant policy, model or agent version, human handoff, and revocation or correction route. Without that receipt, the agent becomes a tunnel through which institutional context disappears.

Where the Book Needs Updating

The book was written around the internet and knowledge-management debates of the late 1990s, so it does not anticipate the full platform economy, mobile surveillance, cloud concentration, recommender systems, large language models, or the labor politics of data extraction. Its optimism about complementing institutions needs pressure from later work on platforms, algorithmic management, surveillance capitalism, content moderation, and automated welfare.

The phrase "social life" can also become too comforting if read loosely. Social context is not automatically humane. Communities can exclude. Organizations can conceal abuse. Informal practice can preserve hierarchy, racism, sexism, credential hoarding, and arbitrary discretion. Context should interrupt brittle automation, but it still needs accountability.

That is the necessary update: defend social practice without romanticizing it. A good AI institution has to preserve tacit knowledge, worker voice, appeal, repair, and local context while also exposing power, bias, capture, and hidden labor to public challenge. The answer to tunnel design is not nostalgia for informal authority. It is accountable practice.

What This Changes

The book belongs here because it clarifies a recurring error in machine-mediated life: confusing access to information with understanding of a world. A generated answer can feel like direct contact with knowledge, but knowledge is carried by practices, institutions, maintenance, trust, and people who know when the formal record is insufficient.

That makes the practical lesson concrete. Before deploying an agent, answer engine, knowledge base, chatbot, workplace copilot, or automated decision layer, ask what background social knowledge the system cannot see. Ask who repairs its mistakes. Ask whether workers can correct it. Ask whether affected people can contest it. Ask whether the system strengthens a community of practice or extracts from it until only process remains.

Brown and Duguid's lasting value is that they make the old future look naive in a way that helps diagnose the new one. Information does not become less social when machines handle more of it. It becomes socially consequential at greater speed and scale.

Source Discipline

This review separates book metadata, author background, book interpretation, and current governance context. Harvard, Berkeley, John Seely Brown's author page, catalog records, and journal reviews support publication history and reception. NIST, ISO, DOL, and EU AI Act sources support current governance claims. They do not prove that any particular copilot, RAG system, or agent preserves social context in practice.

For product claims, the evidence has to be system-specific: corpus list, source authority, connector permissions, data provenance, retrieval logs, model or agent version, evaluation set, human review role, worker consultation record, and appeal path. A vendor phrase such as "grounded," "enterprise knowledge," or "human-in-the-loop" is a starting claim, not a finding.

The page also avoids stronger metaphysical claims about AI. The issue is not whether an AI system understands, believes, becomes conscious, or becomes divine. The issue is simpler and more governable: institutions may act as if the system has carried knowledge when it has only carried information across a broken context boundary.

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