Blog · Review Essay · Last reviewed June 25, 2026

Re-Engineering Humanity and the Programmable Person

Brett Frischmann and Evan Selinger's Re-Engineering Humanity is one of the most useful books for seeing AI as a training environment rather than only a thinking machine. Its central worry is not that smart systems suddenly rebel. It is that smart environments quietly teach people to become easier for machines, firms, and institutions to predict, steer, and administer.

For this review, a programmable person is not a robotized human or a person who has lost all free will. It is a person whose choices, attention, memory, habits, and permissions are repeatedly formatted by systems that make some actions easy, some invisible, and some hard to refuse. The governance question is whether people can see, contest, revoke, and practice alternatives to that formatting.

The key test is not whether technology changes behavior. Every school, workplace, road, ritual, and interface does that. The sharper test is whether a system makes one party's preferred behavior feel like the user's own uncoerced choice while hiding the incentives, records, exits, and alternatives that would let the user judge the arrangement.

The Book

Re-Engineering Humanity was published by Cambridge University Press in 2018. Cambridge Core lists Brett Frischmann of Villanova University and Evan Selinger of Rochester Institute of Technology as the authors, with publication dates in April 2018, ISBNs for digital, hardback, and paperback editions, DOI 10.1017/9781316544846, and subject classifications in philosophy of science, law and economics, law, and philosophy.

The official book site describes the project as a study of big data, predictive analytics, smart environments, fitness trackers, electronic contracts, social media platforms, robotic companions, fake news, autonomous cars, and the wider push to make both worlds and people more programmable. Its table of contents moves from "Engineering Humans" through contracts, extended minds, smart environments, relationship optimization, Turing tests, engineered determinism, and alternative futures.

That range is why the book belongs beside The Technological Society, The Question Concerning Technology, Privacy in Context, The Glass Cage, and Automating Inequality. It is not narrowly a book about AI. It is a book about the social conditions that make AI feel natural: measured conduct, automated defaults, frictionless agreement, optimized environments, and people trained to respond like components in a system.

Current Context

As of June 25, 2026, the book's diagnosis has moved into the language of governance. AI policy now speaks about literacy, human oversight, manipulative systems, deceptive interfaces, data minimization, model memory, recommender choice, and agent permissions. That vocabulary does not prove the book right in every detail. It does show that the problem Frischmann and Selinger named is no longer only a philosophical anxiety about future humans. It is a design, compliance, labor, privacy, and safety problem in deployed systems.

The EU AI Act's Article 4 AI literacy duty entered into application on February 2, 2025. The European Commission says providers and deployers should consider people's technical knowledge, experience, education, training, the context in which AI systems are used, and the people or groups affected. The Commission's Q&A also says the obligation already applies while supervision and enforcement rules apply from August 3, 2026. That matters for this review because literacy is not just knowing what an AI model is. It is knowing when a workflow has been arranged to make a particular response, click, disclosure, or delegation feel normal.

Deceptive-design law and policy now make the consent point concrete. The Digital Services Act official Q&A describes dark patterns as interfaces that trick users and says online platforms must avoid designs that deceive, manipulate, or materially impair free and informed decisions. The FTC's Bringing Dark Patterns to Light report documents consumer-protection risks from disguised ads, hard cancellation paths, buried terms, junk fees, and privacy manipulation. These sources do not use the book's whole vocabulary, but they confirm its central operational claim: interface design can govern behavior without issuing a command.

The AI Act's Article 5 prohibited-practices language is narrower and more severe. It targets specified AI practices that use subliminal, purposefully manipulative, or deceptive techniques, or exploit vulnerabilities tied to age, disability, or social or economic situation, when they materially distort behavior and cause or are reasonably likely to cause significant harm. That is not a general ban on persuasion, recommendation, personalization, or design. It is a legal warning that some forms of machine-mediated influence cross from ordinary interface design into unacceptable harm.

NIST's Privacy Framework adds the privacy-engineering side of the same problem: organizations should manage privacy risk while protecting individuals, not simply collect a notice-and-click record and call it agency. That matters because programmable environments usually run on data flows, not only persuasion. The interface shapes the person; the data system remembers the shaped conduct; later systems treat that record as evidence.

NIST's AI Risk Management Framework supplies the lifecycle vocabulary: govern, map, measure, and manage. Its Generative AI Profile applies that frame to generative systems. NIST's 2026 AI Agent Standards Initiative adds another layer for systems capable of autonomous action, including standards and protocols around security, identity, authorization, and interoperability. Read through Re-Engineering Humanity, those are not only technical controls. They are ways to ask whether a system is training people to become less able to understand, refuse, recover, or act outside the machine-shaped path.

Training Humans for Machines

The book's strongest move is to reverse the usual AI-risk frame. Public debate often asks whether machines will become more human: conscious, creative, social, deceptive, moral, or agentic. Frischmann and Selinger ask how humans are being made more machine-like. The worry is not metaphorical. It is institutional and behavioral.

Smart systems need legible inputs. Platforms need users who click, rate, swipe, consent, disclose, and continue. Employers need workers whose activity can be counted. Schools need students whose performance fits dashboards. Cities need residents whose movement, payments, reports, and risks can be monitored. The system improves when human behavior becomes predictable enough to model, nudge, price, and optimize.

This is techno-social engineering: not only designing tools, but designing the conditions under which people learn how to behave with tools. It works through defaults, friction, categories, incentives, memory, ranking, feedback, sanctions, and choice architecture. The interface becomes a lesson. The lesson becomes a habit. The habit becomes data. The data becomes the next interface's model of what people are likely to do.

The sharper definition is this: techno-social engineering is the redesign of social environments so that human conduct becomes more predictable, measurable, and administrable for a system's goals. It does not require a single villain or a perfect machine. It only requires enough repeated situations in which the easiest path is the path the system can count.

The design target is not obedience in the old disciplinary sense. It is formatted spontaneity: people still feel like they are choosing, but the range of visible choices, the timing of prompts, the vocabulary of options, the memory of past actions, and the cost of refusal have already been arranged. That is why the book is so useful for AI governance. It turns "alignment" back toward the social world and asks who is being aligned to whom.

One of the book's best examples is electronic contracting. Click-through terms are legally serious, but practically unread. Their genius is not persuasion by argument. It is conditioning by repetition. The user learns that access requires a quick affirmative gesture, that reading is impractical, that refusal is costly, and that the system will treat the click as meaningful agreement.

This matters for AI governance because many current consent designs repeat the same form. Training opt-outs, data-sharing settings, biometric notices, cookie banners, workplace monitoring disclosures, student-proctoring acknowledgments, model-memory toggles, and enterprise connector permissions often ask people to perform agency inside a structure where the real terms have already been made hard to inspect, compare, refuse, or revoke.

The issue is not that every click is fake. The issue is that repeated low-friction consent can produce a citizen, worker, patient, student, or user who has learned not to expect a real negotiation. Once that habit is established, AI systems can inherit an enormous permission surface while appearing to operate through choice.

A serious consent record should therefore include the user journey, not only the final click. Was refusal equally visible? Was revocation durable? Was the data use specific? Did the interface distinguish service operation, personalization, training, human review, retention, and vendor transfer? Did an AI assistant or agent frame the permission request as help, loyalty, safety, urgency, or administrative necessity? A click proves little when the path to that click was engineered to make alternatives disappear.

Some practices should not be laundered through consent at all. A manipulative employment dashboard, a child-facing recommender that exploits attachment, an education tool that infers emotion for discipline, or a welfare portal that makes appeal practically unreachable may need prohibition, procurement rejection, or redesign rather than a better checkbox. Consent is one governance instrument, not a solvent for engineered dependence.

Smart Environments

Frischmann and Selinger are especially good on the political ambiguity of smart environments. A smart device can help. A tracker can support health. A navigation system can reduce cognitive load. An assistant can make a service easier to use. A workplace tool can remove tedious steps. The book does not need these systems to be useless in order to criticize them.

The danger is aggregation. Each tool asks for a small trade: more data for more convenience, more automation for less effort, more personalization for less uncertainty, more prediction for less friction. At the scale of a life, those trades can reorganize judgment. People outsource memory, route choice, relationship maintenance, attention management, purchasing decisions, physical activity goals, and bureaucratic navigation to systems whose defaults they did not design and whose incentives they may not share.

Generative AI intensifies this pattern because it moves smartness into language. The assistant does not only count steps or suggest a route. It drafts replies, summarizes evidence, offers career advice, interprets policy, tutors children, writes code, coaches feelings, and translates institutional procedure into conversational instruction. That makes the interface feel less like a device and more like a social environment.

Agentic systems add action. A tool-using assistant can read a workspace, update a record, send a message, book a service, change a setting, or trigger a workflow. The programmable environment then includes not only what the user sees, but what the delegated system can do through the user's identity. Permission classes, logs, approval gates, rollback, and source boundaries become part of the human environment because they define what kind of practical agency remains available.

The same point applies to recommender and ranking systems. They do not merely present options; they allocate salience, repetition, and social proof. A person who repeatedly meets a ranked world learns what kinds of expression, risk, complaint, purchase, style, or belief are rewarded. A programmable environment can therefore be quiet: no command, no threat, just a long series of adjusted likelihoods.

The Machine-Shaped Person

The phrase "programmable people" can sound too blunt, but it names a real design ambition. Institutions often do not need total control. They need reliable channels of influence. They need a high enough probability that users will accept the recommendation, workers will follow the metric, customers will subscribe, students will comply, drivers will take the route, and citizens will move through the portal instead of the office.

That is why this book sharpens debates about human-machine cognition. Cognition is not sealed inside the head. It is distributed through calendars, maps, phones, feeds, dashboards, forms, routines, social cues, and institutional scripts. When those supports are redesigned for extraction, prediction, and optimization, human thought is redesigned with them.

A person can remain formally free while becoming easier to administer. The loss may appear as convenience: fewer decisions, fewer pauses, fewer awkward negotiations, fewer visible conflicts. But judgment often lives in those pauses. So does dissent, patience, interpretation, privacy, and the ability to ask whether a system's goal should be the goal.

This is the book's most important contribution to AI ethics. The danger is not only that a model misclassifies a person. It is that repeated classification trains people to fit the categories. A welfare applicant learns the portal's language. A worker learns the dashboard's rhythm. A student learns the proctoring system's suspicion cues. A customer learns the subscription funnel. A creator learns the recommender's appetite. The system then treats the resulting behavior as natural evidence of preference, risk, productivity, or need.

For that reason, autonomy cannot be measured only by the presence of options. The audit has to ask what capacities are being practiced. Does the system preserve explanation, patience, bargaining, memory, skepticism, repair, and appeal? Or does it reward speed, disclosure, compliance, dependence, affective attachment, and acceptance of default judgment?

Recursive Reality

The book also clarifies a recurring loop in AI-mediated life. Systems observe people. The observations shape predictions. Predictions shape environments. Environments shape conduct. Conduct becomes new data. The next system then treats the shaped conduct as evidence of what people prefer, need, risk, deserve, or are likely to do.

This is where Re-Engineering Humanity connects to recursive reality. A platform may teach users to communicate in measurable gestures, then use those gestures to define engagement. A workplace dashboard may train workers to perform for metrics, then use metric performance to judge productivity. A school proctoring tool may train students to behave for machine suspicion, then treat nervous compliance as normal exam behavior. An AI assistant may train an organization to ask questions in the format it can answer, then make that format look like the natural shape of institutional memory.

In each case, the model does not merely describe reality. It participates in making the version of reality it later measures. That is why source hygiene, contestability, appeal, provenance, and memory boundaries are not secondary compliance details. They are part of how the world is being made legible.

Governance and Safety

The practical governance lesson is to evaluate smart systems by the kind of person, worker, student, patient, customer, or citizen they train. A system that saves time can still weaken refusal. A system that improves service access can still hide appeal. A system that personalizes care can still create dependence. A system that makes work measurable can still erode the local judgment needed to notice when the metric is wrong.

A useful review starts with a person-shaping audit. What behavior is being normalized? Which choice is the default? Which path carries the most friction? What is the refusal route? What is retained as memory? What evidence survives for appeal? Which skills or relationships may atrophy if the system works as designed? Who benefits from the new predictability, and who bears the cost when the system's categories become reality?

The audit needs two records. The design record preserves prompts, screens, defaults, rankings, scripts, A/B tests, policies, agent permissions, and vendor configurations. The outcome record tracks complaints, reversals, overrides, disparate effects, dependency signals, skill loss, abandonment, appeals, and incidents. Without both records, an institution can prove that a system existed but not what kind of person it trained.

For consent and privacy systems, the controls are symmetric acceptance and refusal, specific data-use choices, revocation that propagates, contextual integrity flow maps, data minimization, retention limits, and versioned evidence of the interface state. For AI assistants, the controls are source visibility, uncertainty display, memory inspection and deletion, mode separation, no hidden conversion target, and clear boundaries between suggestion, advice, and action.

For agentic systems, the controls are stricter. Read, write, send, spend, delete, publish, and permission-change powers should be separate. Tool calls should be logged in terms a reviewer can understand. Irreversible or high-impact actions should require approval. Delegated authority should expire. Connectors should be scoped by role and task. Rollback should be planned before deployment. The user should not discover the authority surface only after the agent has acted.

For public services, schools, workplaces, health care, finance, housing, and legal contexts, the baseline should be non-machine alternatives, meaningful human escalation, notice and appeal, documented human oversight, and monitoring for disparate effects and skill loss. A high-stakes system that trains people to comply with machine-readable categories while removing the route to correction is not merely efficient. It is a governance failure.

The safety threshold should rise when a system shapes vulnerable users, captive workers, students, patients, children, people seeking public benefits, or people in crisis. In those settings, "the user can leave" is often fictional. The minimum standard is not only disclosure; it is demonstrable preservation of refusal, escalation, correction, human relationship, and recovery after the system fails.

Where the Book Needs Friction

The book's large frame is powerful, but it can make technology look more unified than it is. Fitness trackers, clickwrap contracts, autonomous vehicles, social media, smart homes, AI assistants, and predictive analytics are not the same system. They operate through different technical architectures, legal regimes, markets, and user relationships. A strong application of the book's argument needs deployment-level specificity.

The LSE Review of Books review by Ignas Kalpokas makes a useful point along these lines. It credits the book for making technology creep visible and for showing how data loops can condition behavior, while also pressing on the underdeveloped shape of the authors' "new humanism." That criticism is fair. The book is better at diagnosing the drift toward machinic behavior than at specifying a fully worked institutional alternative.

There is also a risk of nostalgia. Human beings have always been shaped by tools, institutions, media, and built environments. The right question is not whether technology changes us. It is who gets to design the change, what forms of agency are preserved, what forms of dependence are created, and whether people retain enough friction to notice when a helpful system has become a governing one.

The book also needs class, disability, and labor pressure. Some smart systems reduce barriers and make participation possible. Some friction is exclusion, not agency. Some people were already subject to rigid scripts long before digital platforms arrived. The argument becomes stronger when it distinguishes humane friction from bureaucratic obstruction, assistive automation from extraction, and genuine accessibility from a smoother path into surveillance.

That distinction should be explicit. The answer to programming people is not to add random difficulty. It is to add agency-preserving friction: pauses before irreversible action, visible alternatives, source trails, reviewable reasons, human escalation, and routes out. Bad friction blocks help. Good friction keeps help from becoming capture.

What This Changes

For a consent interface, ask whether refusal is usable, revocation is durable, and the choice is understandable before the system treats it as permission. For an AI assistant, ask whether it preserves uncertainty, source context, and user judgment, or whether it rewards fast acceptance. For an enterprise agent, ask whether it strengthens institutional memory or silently turns old permissions into new authority. For a school, workplace, welfare agency, or clinic, ask whether automation expands care and appeal, or simply makes people conform to machine-readable categories.

The book's value is not that it rejects all automation. It makes automation morally concrete. A system can save time while shrinking discretion. It can make services smoother while hiding power. It can personalize experience while narrowing imagination. It can make people feel served while training them to stop asking for terms they can understand and institutions they can contest.

Re-Engineering Humanity is therefore a useful bridge between media theory, surveillance studies, AI governance, and philosophy of technology. It asks a hard question that every smart environment should face: after people live with this system long enough, what capacities will they have practiced, and what capacities will they have surrendered?

The answer should be written before deployment, not inferred after dependence forms. If a product will train workers to obey dashboards, children to trust chatbots, patients to narrate symptoms into proprietary triage, or citizens to navigate services through automated categories, the institution owes the public a theory of preserved agency and evidence that the theory survives contact with use.

Source Discipline

This review separates book facts, author framing, critical reception, regulatory vocabulary, standards guidance, and product claims. Cambridge Core and the official book site establish publication metadata, scope, authorship, and chapter structure. The LSE and Prometheus reviews help locate reception and criticism. European Commission, FTC, EDPB, and NIST materials establish current governance language; they do not prove that any particular AI product is safe, fair, compliant, or autonomy-preserving.

Claims about programmable people require deployment evidence. A product demo shows a workflow, not social effect. A consent screen shows an interface state, not meaningful consent. A benchmark shows measured performance, not retained human agency. A regulatory duty shows a legal expectation, not operational compliance. A serious source trail should identify the system, date, jurisdiction, affected population, task, data flow, interface version, human role, permission record, appeal path, and incident history.

Regulatory language should also stay precise. Article 4 AI literacy is not the same as user education in general. DSA dark-pattern restrictions apply to online platforms in the DSA frame, not to every interface everywhere. AI Act Article 5 prohibitions are specific and harm-conditioned, not a blanket ban on all nudges. NIST frameworks are voluntary guidance unless incorporated by contract, procurement, policy, or another binding obligation.

The argument is also bounded. This page does not claim that AI systems are conscious, divine, or AGI. It treats AI systems as engineered interfaces, models, workflows, and institutional processes that can shape people through defaults, permissions, incentives, memory, and action channels.

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