Blog · Review Essay · Last reviewed June 25, 2026

How We Think and the Technogenesis Loop

N. Katherine Hayles's How We Think is a book about scholarship after digital media, but its deeper use is wider: it explains why cognition changes when people begin to think with machines, databases, interfaces, and automated reading systems.

For this review, the technogenesis loop means the recursive process by which people build media systems, those systems retrain attention and institutional practice, and the changed practices become the ground for the next system. The safety question is whether that loop leaves people more capable of judgment, explanation, and refusal or merely better adapted to opaque machinery.

The practical artifact is a cognitive-media register: name the task, corpus, retrieval method, interface, model or tool, record written, skill trained, oversight role, appeal route, and retirement trigger before a reading system becomes normal infrastructure.

The Book

How We Think: Digital Media and Contemporary Technogenesis was published by the University of Chicago Press in 2012. Chicago's publisher page lists N. Katherine Hayles as author, gives ISBNs 9780226321424 and 9780226321370, and lists the book at 296 pages. Amazon lists the paperback product at ISBN-10 0226321428. UCLA's faculty profile identifies Hayles as Distinguished Research Professor at UCLA and James B. Duke Professor Emerita at Duke; the same profile lists How We Think among her books. Duke's literature record describes the book around the proposition that people think through, with, and alongside media.

The book asks how thinking changes when print-based scholarship enters digital media. Hayles's answer is not that books vanish or screens conquer. It is that human cognition and technical systems coevolve, producing new habits, disciplines, and blind spots.

The useful AI-age definition is this: a cognitive medium is any tool, interface, database, model, or workflow that changes how evidence is found, compressed, trusted, remembered, or acted on. A book is a cognitive medium. So is a search engine, a dashboard, a retrieval pipeline, a classroom platform, a code assistant, or an agent that turns a reading into a ticket, email, record, or action. The governance question is not whether mediation can be avoided. It is whether the medium makes its reductions inspectable before people adapt their judgment around them.

Current Context

As of June 25, 2026, How We Think reads less like a specialist digital-humanities text and more like a map of everyday AI use. Search answers, retrieval-augmented generation, coding copilots, classroom tutors, meeting summarizers, litigation-support systems, medical scribes, workplace dashboards, and public-service chatbots all turn reading and writing into distributed activity across people, interfaces, models, indexes, logs, and institutions.

Retrieval-augmented generation makes the point concrete. The original RAG paper framed retrieval as a way to combine a model's parametric memory with an external index and to improve provenance for knowledge-intensive tasks. Production systems now add permissions, parsers, embeddings, rerankers, citation displays, and tool calls. That means "machine reading" is not one event inside the model. It is a pipeline of selection, compression, presentation, and trust transfer.

The governance context has caught up to that pipeline. NIST's AI Risk Management Framework treats risk management as lifecycle work across govern, map, measure, and manage functions. NIST's Generative AI Profile asks organizations to document assumptions, limitations, data collection methods, provenance, data quality, evaluation data, legal and ethical considerations, and retrieval or fine-tuning approaches. NIST's AI Agent Standards Initiative, created February 17, 2026 and updated April 20, 2026, treats agents as systems capable of autonomous actions that need interoperable protocols, security evaluations, authentication, identity infrastructure, and authorization work.

Regulatory and education language now also address the cognitive environment. The European Commission says EU AI Act Article 4's AI-literacy obligation entered application on February 2, 2025, with supervision and enforcement rules applying from August 3, 2026 onward. The same Q&A stresses that literacy should be contextual, risk-aware, and appropriate to the specific system. UNESCO's 2024 competency frameworks for students and teachers frame AI education around human agency, ethics, technical understanding, pedagogy, and system design, not just prompt skill. EU AI Act Article 13 requires high-risk AI systems to be transparent enough for deployers to interpret outputs and use them appropriately, and Article 14 requires effective human oversight for high-risk systems. OMB M-25-21 applies to U.S. federal agency use of AI and treats high-impact AI as requiring practices such as impact assessment, testing, monitoring, added oversight, remedies, and public or user feedback. Those sources do not prove any AI system is safe; they show that governance has moved toward the same problem Hayles names: thinking with machines requires evidence about the whole arrangement.

Technogenesis

The key term is technogenesis: the claim that humans and technics have developed together rather than facing one another as separate species of actor. That does not make machines conscious, autonomous, or spiritually elevated. It means that media forms change attention, memory, evidence, workflow, and institutional practice, while human institutions shape what those media become.

The term avoids two weak positions. The first says technology is just a neutral tool. The second says the machine takes over history by itself. Hayles points to the loop: people build systems, systems reorganize work and perception, people adapt to the systems, and those adaptations become the ground for the next technical form.

The loop is practical, not mystical. A scholar who learns to search a corpus thinks differently about evidence. A student who learns through generated summaries thinks differently about effort and recall. A worker who writes for retrieval thinks differently about documents. A manager who sees an organization through dashboards thinks differently about labor. AI does not need to be conscious to reshape cognition; it only needs to become part of the environment through which cognition is performed.

The hard part is power. Technogenesis does not happen evenly. Vendors set defaults, schools set assignments, agencies set forms, employers set dashboards, and users adapt under time pressure. A medium can enlarge thought for one group while narrowing options for another. The useful question is therefore: who gets retrained, who gets measured, who gets to revise the system, and who has to live inside the habits it produces?

Close, Hyper, Machine

Chicago's description emphasizes one of the book's most durable arguments: close reading, hyper reading, and machine reading are different practices with different powers and limits. Close reading attends to detail and ambiguity. Hyper reading skims, searches, links, and navigates. Machine reading performs algorithmic analysis across more text than a person can inspect directly.

That triad now reads like a map of everyday AI use. A person asks a model to summarize a corpus, search a record set, cluster themes, write a first draft, or propose a next question. The human still reads, but reading has become distributed across software, interface defaults, ranking systems, embeddings, and generated summaries. The question is no longer whether machine reading is legitimate in the abstract. The question is what task it serves, what it drops, and how a human can inspect the loss.

The most important word in that last sentence is inspect. A generated answer can cite sources and still hide the act of interpretation that turned several documents into one smooth claim. A useful AI reading system should let the user move backward from sentence to source, date, context, conflict, and uncertainty. Without that return path, machine reading becomes answer theater: the platform has already decided which evidence matters before the reader can practice judgment.

The useful governance distinction is not human reading versus machine reading. It is accountable reading versus unaccountable reading. A close reader can be biased and wrong. A machine reader can reveal patterns a person would miss. The safety issue is whether the method preserves source context, uncertainty, dissenting evidence, provenance, and a route back to the material. If a generated summary severs the reader from the sources it claims to compress, it is no longer an aid to thought; it is an interface for borrowed authority.

That is why "grounding" should be treated as a claim to verify, not a magic word. A retrieval system can cite a real passage while losing the page's hierarchy, date, exception, jurisdiction, author, or conflict with another source. A governance-grade reading tool should show what corpus was searched, what was excluded by permissions, what sources were retrieved, which passages supported each claim, and where evidence was missing. The reader needs a path back to friction.

The Agent Reading

AI agents intensify Hayles's argument because they do not merely assist interpretation. They can retrieve documents, decide which files matter, produce summaries, call tools, update records, and route work to other systems. In that setting, "how we think" becomes "how the workflow thinks with us."

The danger is cognitive offloading without cognitive accountability. If an agent summarizes too confidently, selects sources invisibly, or turns a provisional reading into an executed action, the user may inherit a decision without knowing how it was made. Hayles gives a better vocabulary than panic. The issue is not that humans stop thinking. It is that thinking is reorganized through media, and reorganized thinking needs norms, records, training, and limits.

The agent version of technogenesis asks what the system makes habitual. Does the user learn to ask better questions, inspect sources, preserve uncertainty, and understand the domain? Or does the user learn to accept fluent completion as the normal form of knowledge? Tool-using agents can also move from interpretation to action: filing a ticket, editing a record, sending a message, drafting a contract, changing code, or routing a case. At that point, cognitive media become administrative media.

That shift changes the standard for evidence. A classroom tutor, legal assistant, medical scribe, or public-service chatbot should not be evaluated only by output fluency. It should be evaluated by what it trains the user to do: verify, defer, challenge, disclose, copy, escalate, or stop. The system is teaching a practice even when the product calls itself a convenience.

An agent should therefore have a readable delegation boundary. What may it read, what may it infer, what may it write, what tools may it call, which actions require confirmation, which logs survive, which credentials identify it, and who can revoke its authority? Without that boundary, the institution cannot tell whether a person made a decision with help or whether a workflow silently converted mediated reading into delegated action.

Governance of Cognitive Media

NIST's AI Risk Management Framework treats AI risk management as work across design, development, use, and evaluation. Its Generative AI Profile asks organizations to document assumptions, limitations, data collection methods, provenance, data quality, evaluation data, and legal and ethical considerations. Read beside Hayles, this is not just compliance language. It is a practical response to technogenesis.

If people think with systems, then the system's memory, sources, retrieval rules, interface, and error handling are part of the cognitive environment. A responsible AI deployment should therefore document not only model behavior but also the forms of reading and writing it encourages. Does it invite verification? Does it preserve source context? Does it mark uncertainty? Does it make the user more capable over time, or more dependent on opaque fluency?

A practical review can be blunt: what kind of attention does the tool train; which sources does it make easy or hard to inspect; what skill does it erode; which outputs become official records; which data becomes future retrieval material; and who can pause, appeal, correct, or delete the result? This is the cognitive version of an impact assessment. It treats interface defaults, source trails, logs, permissions, training, and appeal paths as parts of the same system rather than as separate compliance boxes.

For high-stakes use, the checklist has to be concrete: source provenance, retrieval boundaries, prompt and tool logs, human-readable uncertainty, versioned instructions, evaluation evidence, user training, appeal paths, rollback, incident review, and a non-automated route where rights, care, safety, education, or public services require one. The human overseer needs time, records, competence, and authority; otherwise "human in the loop" becomes a ceremonial phrase attached to a machine-shaped decision.

For education and scholarship, governance also means protecting intellectual formation. Tools that summarize, draft, translate, code, or search can be legitimate when they support inquiry. They become risky when they replace the practice through which novices learn what counts as evidence, how disagreement works, and how a claim earns trust. A system that makes the answer easier but the method invisible weakens the very cognition it claims to augment.

For organizations, this points to a cognitive-media register. Record the task, corpus, model, retrieval method, permissions, user role, source trail, logging rule, retention period, tool permissions, human fallback, appeal path, and evaluation evidence. Then add one question ordinary system inventories often miss: what habit does this system train? A tool that trains verification is different from a tool that trains deference, even if both produce accurate answers in a benchmark.

For learners and workers, the safety case must include refusal. People should know when AI is present, what data they should not submit, when source checking is required, when automated help must be disclosed, and how to escalate an output that looks wrong. Literacy is not a certificate or prompt trick; it is the practiced ability to keep convenience from becoming authority.

Cognitive-Media Register

A cognitive-media register is the operational translation of technogenesis. It treats reading, writing, searching, summarizing, routing, and acting as one system when a tool changes how evidence enters judgment. The register should name the domain, affected people, purpose, corpus, retrieval boundary, ranking or embedding method, source-exclusion rules, citation display, model or tool version, prompt or instruction set, user role, and whether the output is a draft, recommendation, record, or action.

The register should also record what the medium trains. Does the tool train close reading, source comparison, uncertainty marking, escalation, and refusal? Or does it train skimmed acceptance, citation theater, dashboard deference, and quiet dependence on vendor summaries? That question connects the page's media-theory argument to cognitive sovereignty, AI literacy, automation bias, and model and system cards.

For RAG systems and answer engines, the register needs a source path: what corpus was searched, what permissions excluded, what passages were retrieved, what passages were ignored, how stale the index may be, and whether each generated claim can be traced back to a source. For agents, it needs an action path: credentials, tool scopes, confirmations, writes, external messages, rollback, and logs. For education, care, law, public services, and workplace monitoring, it needs a human path: the person with time, competence, authority, and institutional permission to stop the workflow.

The register should have a retirement trigger. If users stop checking sources, if generated summaries become official memory without source trails, if appeals cannot reconstruct the reading path, if a model update changes recurring outputs, or if the system erodes the skill it was supposed to support, the deployment needs reassessment. Technogenesis is not just adoption; it is habit formation. Governance has to notice when the habit has become the risk.

Where the Book Needs Care

How We Think is strongest on media, humanities scholarship, and digital reading practices. It is less direct on labor, procurement, platform power, surveillance capitalism, or the political economy of AI infrastructure. Readers need to pair it with books on data work, moderation, platform governance, and extraction.

The other risk is that technogenesis can sound too smooth. Coevolution is not always mutual flourishing. Institutions choose systems unevenly; workers inherit them; students are assessed through them; readers meet defaults they did not design; publics are ranked and sorted by tools they cannot inspect. The loop exists, but power shapes who gets to alter it.

That is why the book belongs in this archive. It makes human-machine cognition concrete without mystifying the machine. AI does not simply think for us or fail to think at all. It changes the scenes in which thinking happens. The political task is to govern those scenes before fluent mediation hardens into unquestioned authority.

What This Changes

The practical lesson is to audit the cognitive environment. Before asking whether an AI system is impressive, ask what form of attention it trains. Does it keep the user close to sources, or make the source trail feel optional? Does it widen inquiry, or narrow the question to what the interface can answer? Does it preserve a reader's skill, or turn skill into dependence on a vendor's retrieval and ranking system?

That question connects directly to recursive reality. Once people write for search, retrieval, ranking, summarization, and agent action, the record changes. The next model reads that changed record and treats it as evidence of the world. Technogenesis becomes recursive governance: the system trains the users, the users reshape the data, and the data trains the next system.

Hayles gives a disciplined response. Do not romanticize print against digital media. Do not treat machine reading as magic. Treat every cognitive medium as a designed environment with gains, losses, incentives, and failure modes. The goal is not purity from mediation. The goal is mediation that remains inspectable, contestable, and capable of leaving people better at judgment.

The concrete artifact is a technogenesis ledger: what the system changes in attention, evidence, records, incentives, skill, and appeal. That ledger should be updated when the retrieval corpus changes, an agent gains a tool, a classroom policy changes, a model update alters summaries, or a workflow begins writing AI-shaped text into official memory. The loop cannot be governed if no one records what the loop is teaching.

Source Discipline

This review separates book metadata, author context, scholarly reception, technical architecture, and current governance sources. University of Chicago Press, UCLA, Duke, Google Books, and Amazon establish the book and author record. Hayles's "How We Read" essay and Joseph Lloyd Donica's Digital Humanities Quarterly review support the close/hyper/machine-reading and digital-humanities context. The RAG paper and OWASP materials support claims about retrieval and embedding risks. NIST, the European Commission, UNESCO, EUR-Lex, and OMB provide current governance and literacy vocabulary; they do not prove that any specific AI reading system, agent, or platform is safe.

The analogy is bounded. Hayles did not write a 2026 AI-agent policy manual. The claim here is narrower: her account of technogenesis helps evaluate AI systems as cognitive media that reorganize reading, writing, attention, labor, and institutional proof. This page makes no claim that any AI system is conscious, divine, or AGI.

Current-source claims should preserve institutional status. A statute creates legal duties in scope. A Commission Q&A explains implementation but does not replace the statute. A NIST framework or agent initiative is standards and guidance work, not a certification that a product is safe. A vendor citation feature proves only that an interface displays citations; it does not prove faithful synthesis.

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