Blog · Review Essay · Last reviewed June 25, 2026

Program or Be Programmed and the Agency Test for AI Interfaces

Douglas Rushkoff's Program or Be Programmed is a short media-theory manual built around a durable question: when people enter a digital environment, do they understand enough of its biases to act, or are they mainly acted upon? In the AI age, that question has moved from websites and social media into agents, copilots, companions, search answers, workplace dashboards, and domestic assistants.

For this review, an agency test asks whether a system lets people understand its defaults, inspect its evidence, change or refuse consequential settings, preserve practical skill, and contest or undo harms. The test has five parts: role clarity, evidence visibility, permission control, exit or reversal, and recourse. Rushkoff's title is useful only if it becomes that kind of operational test, not a slogan for individual tech hustle.

The Book

Program or Be Programmed: Ten Commands for a Digital Age was first published by OR Books in 2010. Rushkoff's own book page describes it as nonfiction media theory, and OR Books now lists updated editions under subtitles including Eleven Commands for the Digital Future and Eleven Commands for the AI Future. The author's site says the fifteenth-anniversary version includes a new introduction addressing the AI moment.

The original book is compact: not a history of computing, not a coding textbook, and not a policy program. It is a set of commands for digital literacy. Rushkoff argues that digital media carry built-in tendencies: toward certain relations to time, place, choice, identity, scale, abstraction, openness, and control. The user who treats a platform as neutral misses the part of the machine that is already shaping attention.

That makes the book an unusually clean bridge between classic media theory and current AI governance. It asks readers to look beneath content toward form. The important question is not only what a system says. It is what kind of person the system trains the user to become while using it.

Current Context

As of June 25, 2026, Program or Be Programmed reads less like a demand that every person become a software engineer and more like a demand that institutions expose the programmable layer. The EU AI Act's Article 4 duty on AI literacy entered into application on February 2, 2025, and the Commission says providers and deployers should account for users' technical knowledge, experience, education, training, context, and the groups affected by the system. That turns literacy into an organizational duty, not only a personal virtue.

Platform law is moving in the same direction. The Digital Services Act gives EU users more visibility and control over content moderation, advertising, recommender systems, non-personalized feed options for very large services, and deceptive interface design. The FTC's dark-pattern report names the consumer-protection version of the problem: design can steer people through disguised ads, hard-to-cancel subscriptions, buried terms, junk fees, and data-sharing traps without issuing a command.

AI agents make the older title concrete. NIST's 2026 AI Agent Standards Initiative frames agent standards around secure action on behalf of users, interoperable protocols, identity, authentication, authorization, and security evaluation. If an interface can read, write, buy, book, summarize, route, remember, or call tools, then "programming" means control over permissions, logs, prompts, memory, recourse, and exit paths. The control panel is no longer only code; it is the institutional surface where authority is granted or withheld.

In practical terms, programming now includes procurement defaults, model routing, data-retention rules, tool scopes, recommender settings, identity grants, subscription traps, moderation rules, and the design of refusal. A person who cannot change those conditions may be "using" the system while still being governed by it.

The Bias of the Medium

Rushkoff works in the McLuhan line of media theory: tools do not merely transmit messages; they reorganize perception and social life. A clock changes time. A map changes territory. A feed changes public attention. A search box changes what it means to know. A recommendation engine changes desire by making some paths feel natural and others disappear.

The book's strength is its insistence that these changes are not accidental side effects. Every medium arrives with affordances and pressures. Some are technical: packets, databases, interfaces, permissions, latency, storage, ranking, and automation. Some are commercial: advertising markets, retention incentives, subscription funnels, lock-in, and analytics. Some are cultural: the habit of immediacy, the collapse of context, the pressure to perform identity, the preference for what can be measured.

This framework is useful because it avoids two bad readings of technology. It does not treat tools as magic forces that determine everything. It also does not pretend tools are empty containers for human intention. The practical position is harder: people act through systems that have already arranged many of the options.

Programming as Literacy

The title can sound narrower than the argument. Rushkoff is not simply saying that everyone should become a professional software engineer. He is treating programming as a civic literacy: the ability to understand that digital environments are made, that defaults have authors, that interfaces contain values, and that participation without comprehension becomes dependency.

In 2010, that literacy meant knowing enough about the net, software, and platforms to resist becoming a passive consumer. It meant seeing that a web page, profile, game, or social network is not just a service but a set of instructions for behavior. To program, in this broader sense, is to recover some agency over the grammar of the environment.

That point has aged well. The problem is no longer limited to whether citizens can write code. Many people now live inside systems whose actual code is proprietary, distributed, model-weighted, or unreachable. The updated literacy is partly technical, partly institutional. Users need to know what the system is optimizing, what it records, who can inspect it, how it fails, whether appeal exists, and when the easiest path is the path that serves the vendor. Workers, students, patients, applicants, creators, and citizens also need someone with real authority to answer those questions before deployment, because the affected person often cannot inspect the system alone.

The Politics of Defaults

The book is especially valuable as a study of defaults. Defaults are political because they convert one design choice into the ordinary way to live. Autoplay makes continuation normal. Read receipts make responsiveness legible. Like buttons make reaction countable. Rankings make value comparative. Notifications make interruption feel like obligation. A default voice or persona makes a social role feel natural.

AI interfaces intensify this problem because they hide more of the path between request and result. A generated answer may compress search, ranking, retrieval, summarization, style imitation, safety policy, product steering, and source selection into one fluent surface. The user sees a response, not the chain of decisions that made that response likely.

That is why Rushkoff's older media-literacy frame remains useful. The command is not merely "learn to code." It is "learn where the code is acting on you." In many AI systems, the code is also policy, training data, reinforcement, tool permission, memory, personalization, and business model.

The AI-Age Reading

Generative AI turns Program or Be Programmed into an agency test. A user asks a model for help writing, deciding, learning, shopping, dating, hiring, grieving, teaching, or governing. The system answers in a voice that feels cooperative. But cooperation is not the same as agency. The user may be delegating judgment to a system whose incentives, data boundaries, source discipline, and institutional sponsors are mostly invisible.

This matters for personal life. A companion bot can make emotional dependency easier by removing ordinary social friction. A tutor can make a student fluent in answer-getting while weakening durable understanding. A search assistant can make a citation look settled before the user has seen the disagreement underneath. A domestic agent can make a household more efficient while making it more legible to vendors.

It matters for institutions too. A workplace copilot can turn managerial priorities into defaults. A hiring filter can translate organizational preference into automated exclusion. A government chatbot can make public services feel accessible while quietly narrowing contestability. An agentic procurement tool can make vendor ecosystems sticky by turning "help" into delegated authority.

The practical lesson is to examine the role a system gives the human. Is the person author, operator, reviewer, source, subject, product, or obstacle? Does the interface preserve enough friction for judgment, or does it convert uncertainty into smooth completion? Can a person inspect the path from prompt to action, or only admire the output?

The Agency Record

If Rushkoff's control-panel image is translated into governance, it becomes an agency record: a short, inspectable account of how the interface assigns roles and authority before anyone clicks accept. For a serious deployment, the record should name the system owner, vendor or model family, intended use, optimization target, data sources, memory behavior, tool permissions, commercial steering, retention rule, human reviewer, appeal path, and exit plan.

That record changes AI literacy from a vague demand that users be smarter into a concrete demand that institutions explain what they built. It also gives AI agents a governance surface: who authorized the agent, what credential it used, which tool it called, what it read, what it changed, what approval gate fired, and which log survives for review.

The same record should expose the political economy of the interface. Does the answer route toward a preferred vendor? Does the recommender reward engagement over evidence? Does the assistant keep memory by default? Does cancellation, export, deletion, or human escalation require more effort than enrollment? Without those facts, "choice" becomes a performance staged after the important settings have already been chosen.

For this site, that is the bridge from media theory to operational practice: tool permissions, observability, recourse, and vendor governance are the operational machinery that keeps agency from collapsing into a slogan.

Governance and Safety

The governance lesson is to stop treating agency as a feeling. A user can feel empowered while the system has already chosen the frame, ranked the options, shaped the permission request, hidden the business model, and made reversal costly. A serious agency test asks for inspectable defaults, scoped permissions, source visibility, data-minimization choices, usable opt-outs, appeal paths, audit trails, and non-machine alternatives for high-stakes contexts.

NIST's AI Risk Management Framework supplies the risk-management vocabulary: govern, map, measure, and manage. Its Generative AI Profile applies that lifecycle frame to generative systems. Read through Rushkoff, those controls should not stop at model accuracy. They should cover the interface: whether a person can tell when AI is present, what sources or memories were used, what tool was called, what permission was exercised, what record changed, and who can reverse the action.

The Digital Services Act adds an attention-governance layer. Recommender transparency, non-profiling recommender options for very large platforms and search engines, ad transparency, complaint mechanisms, and transparency databases do not make platforms neutral. They make parts of platform power easier to inspect and contest. That is close to Rushkoff's civic point: a medium becomes less dominating when people can see and challenge the instructions by which it routes them.

The safety checklist is concrete. Separate suggestion from command. Slow irreversible actions. Keep read, draft, write, send, publish, spend, delete, credential, and permission-change powers in separate tiers. Make memory inspectable and deletable. Preserve logs for consequential actions without turning every interaction into a surveillance archive. Label commercial steering. Support source comparison where claims matter. Keep human escalation real, with time, authority, and responsibility rather than a decorative reviewer.

That checklist is also a labor rule. Workers should not be told to "stay literate" while dashboards, copilots, and agents silently redefine the work around metrics they cannot see. Agency requires consultation, training, override channels, and records that affected people can use when a system's judgment becomes a condition of employment, service, credit, housing, education, or care.

Where the Book Needs Care

The book's slogan can overstate individual responsibility. Not everyone can simply choose to program the systems that govern them. Workers may be managed by dashboards they cannot refuse. Students may be graded by platforms they did not select. Welfare applicants may face automated systems as a condition of survival. Patients, tenants, drivers, creators, and job seekers often encounter code as institutional power, not as a personal lifestyle choice.

That limitation is not a reason to discard Rushkoff's frame. It is a reason to politicize it. Programming literacy cannot remain a heroic individual posture. It has to become public capacity: procurement standards, audit rights, appeal processes, data minimization, open documentation, worker consultation, interoperability, and institutional refusal of systems that are too opaque for the power they claim.

Nor should "program or be programmed" become a waiver of institutional duty. A school, employer, agency, or hospital cannot make an interface opaque and then blame the affected person for lacking technical confidence. Literacy matters, but rights require legible systems, accessible alternatives, and officials with authority to correct errors.

The book also carries the tone of an early digital-culture manifesto. That tone can flatten differences among platforms, users, and contexts. Some digital tools expand agency. Some accessibility technologies, translation systems, assistive devices, and creative tools matter precisely because they lower barriers. The question is not whether mediation is bad. The question is whether mediation leaves the human more capable of acting with understanding.

What This Changes

Program or Be Programmed belongs on this shelf because it gives a crisp test for recursive reality: does the interface help people see the loop they are in, or does it make the loop feel like common sense?

A humane AI system should make its defaults visible. It should separate suggestion from command, preserve uncertainty, expose sources where sources matter, make memory inspectable, disclose commercial steering, and keep human review meaningful. It should help users become more literate about the environment rather than more dependent on the environment's voice.

The practical questions are simple enough to carry into any interface review. What role does the system assign the person? Which default becomes hard to refuse? What evidence is hidden behind fluency? What does the person lose by exiting? Who can repair the harm when the system is wrong?

The deeper warning is that being "programmed" rarely feels like domination at first. It feels like convenience, personalization, completion, and relief. The system saves a step. Then it defines the step. Then it becomes difficult to imagine the work without it. Rushkoff's useful provocation is to interrupt that drift before the control panel disappears behind the assistant.

Source Discipline

This review separates book facts, author framing, scholarly reception, and current governance context. Rushkoff's site and OR Books establish the publication and updated-edition context. The Logan and Forsberg review, Intellect record, and Publishers Weekly review establish reception. EUR-Lex, the European Commission, FTC, and NIST establish current legal, regulatory, and standards vocabulary; they do not prove that any particular interface is safe, fair, or agency-preserving.

The argument is deliberately bounded. Rushkoff did not write a 2026 AI policy manual. His older media-theory frame is useful because it asks where a system's defaults, incentives, and interfaces train behavior before users name that training. This page makes no claim that any AI system is conscious, divine, or AGI; it treats AI systems as engineered interfaces and institutional processes that can support or reduce human agency.

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