Updating to Remain the Same and the Habit Loop of New Media

The Book

Updating to Remain the Same: Habitual New Media was published by the MIT Press in hardcover and ebook on May 27, 2016, and in paperback on August 11, 2017. MIT Press lists the hardcover ISBN as 9780262034494, the ebook ISBN as 9780262333788, and the paperback ISBN as 9780262534727. Its publisher description locates the book's subject not at the moment a medium appears new, but after repeated use has embedded it in bodies and everyday life.

Chun develops that point through the compact relation “habit + crisis = update.” The phrase is useful because the update is not simple novelty. A perceived crisis—obsolescence, insecurity, disconnection, falling behind—makes revision feel necessary; revision then preserves the user's attachment while changing the conditions of participation.

Simon Fraser University's current faculty profile identifies Chun as Canada 150 Research Chair in New Media, Professor in the School of Communication, and Director of the Digital Democracies Institute. Her work connects cultural formations to technological artifacts and protocols, which is why this book moves between bodies, interfaces, networks, privacy, publicity, and political economy rather than offering a psychology of screen time.

The book belongs beside Programmed Visions, The Culture of Connectivity, The Twittering Machine, and Filterworld. It shifts attention from spectacular technological change to the repetitive routines through which technology becomes the background grammar of everyday life.

Current Context

As of August 12, 2026, Chun's argument applies to AI less as a claim about one dramatic invention than as an account of gradual installation. Summaries, autocomplete, saved memory, recommendation slots, answer boxes, recurring tasks, and agent permissions become the default route one small acceptance at a time.

The mechanisms must be kept separate. A transcript can be retained as history; a saved or inferred memory can condition later responses; product analytics can change an interface; a recommender can rank what appears next; and an interaction may or may not be reused for model improvement. Calling all five “learning” hides which data moved, which system changed, and which consent or control applies. A disciplined habit audit records the actual path.

The EU Digital Services Act regulates several parts of that path with different scopes. Article 25 bars online platforms from deceptive or manipulative interface design that materially impairs free and informed decisions. Article 27 requires plain-language information about recommender parameters and available controls. Article 28 requires platforms accessible to minors to provide a high level of privacy, safety, and security. Articles 34 and 35 require designated very large platforms and search engines to assess and mitigate systemic risks, while Article 38 requires those largest services—not every platform—to offer at least one recommender option not based on profiling.

On July 10, 2026, the European Commission issued a preliminary finding that the design of Instagram and Facebook breached the DSA. The investigation focuses on infinite scroll, autoplay, push notifications, and highly personalized recommenders, and the Commission's stated concern includes physical and mental wellbeing, minors, and vulnerable adults. This is an enforcement allegation in an ongoing process, not a final adjudication. Separately, the Commission's non-binding 2025 guidance for in-scope platforms accessible to minors recommends disabling by default features such as streaks, autoplay, read receipts, ephemeral content, and push notifications, and calls for safeguards around integrated AI chatbots; following the guidance is voluntary and does not itself guarantee compliance.

The AI Act now adds a different layer. Article 50's transparency duties for specified AI systems and generated or manipulated content began applying on August 2, 2026, with a limited transition until December 2, 2026 for the Article 50(2) marking-and-detection duty of systems placed on the market earlier. Regulation (EU) 2026/1744 also revised Article 4: providers and deployers must take context-sensitive measures that support staff AI literacy, but they need not guarantee a particular level for every person. NIST's Generative AI Profile remains voluntary guidance; its “Human-AI Configuration” risk includes automation bias and over-reliance. NIST's 2026 Agent Standards Initiative is an initiative for research and voluntary standards work, not a completed agent-safety standard.

U.S. consumer-protection enforcement supplies a narrower interface example. A 2025 federal court order resolving the FTC's Amazon Prime case requires clear enrollment disclosures, affirmative consent, a clear way to decline, and cancellation through the same medium used to sign up. Those requirements bind that order; they are not a universal AI law. Their transferable design principle is symmetry: a repeated path should not make authorization easy and withdrawal obscure.

Habit as Infrastructure

Chun's central move is to treat habit as infrastructure rather than private weakness. New media announce novelty, speed, participation, and constant transformation; their durable power comes from repeated bodily practice: checking, scrolling, posting, saving, updating, searching, and carrying a networked device as an ordinary condition of availability.

A cue-action-reward diagram can describe one episode, but it cannot carry the book's institutional argument. Habitual media have three coupled layers. At the bodily layer, repetition reduces deliberation. At the technical layer, interfaces store traces, rank responses, and revise the next encounter. At the institutional layer, employers, schools, markets, peer groups, and public agencies can make one route costly to refuse. The system becomes infrastructure when those layers reinforce one another.

The “habit + crisis = update” relation explains why change can deepen continuity. A security warning, threat of obsolescence, broken compatibility, missing feature, or fear of social exclusion creates urgency. The update offers restoration, yet may also alter defaults, data use, ranking, price, identity requirements, or permissions. The user keeps performing the old routine through a system with newly enlarged authority.

For governance, a material update is therefore a new risk boundary, not merely a release note. If a writing tool gains persistent memory, a search box becomes an answer engine, or an assistant gains the ability to send, buy, publish, or delete, continuity of branding does not mean continuity of risk. Review should compare the system before and after the change.

Habit is not inherently capture. Repetition supports craft, care, accessibility, maintenance, community, and memory. The sharper test is whether a routine remains revisable: can a person understand it, interrupt it, choose another route, and leave without losing disproportionate access to work, relationships, education, or public life?

Networks and the Personalized You

Chun links habitual media to networks as a dominant social metaphor. Network language promises connection, flexibility, personalization, and individual empowerment. It also makes society appear as clusters of addressable individuals, each treated as a target, profile, node, or personalized feed endpoint.

The book is especially good on the operational intimacy of personalization. A system says “you” while acting on a statistical subject assembled from traces, correlations, categories, and expected behavior. Tailoring may be genuinely useful, but a relevant output is not proof that the system knows the person in the human sense. The address is also a routing decision: what to show, withhold, prioritize, price, recommend, or ask next.

That is why personalization is not only a privacy problem. It is a feedback problem. A system repeatedly presents what its profile predicts; the person adapts to what is visible, quick, rewarded, or recognized; the adaptation then becomes further evidence. As Filterworld helps show, the profile can become a curriculum rather than a neutral description.

Source discipline matters here. Product personalization, advertising profiles, recommender ranking, persistent assistant memory, and foundation-model training can share data or infrastructure, but they are not the same process. The first audit question is not “does the AI learn me?” It is: what observation was made, where was it stored, which decision used it, for how long, and can the person inspect or correct it?

Generative interfaces intensify the effect because tailored language can feel like recognition. That makes the argument in Discriminating Data a useful sequel: correlation can grant or deny recognition without recovering the person behind the proxy. A fluent address should therefore increase the demand for provenance and contestability, not lower it.

Privacy, Publicity, and Exposure

One of the book's strongest contributions is its refusal to treat privacy as a simple individual setting. Chun argues that networked media invert and scramble privacy and publicity: supposedly personal devices are intensely communicative, and supposedly public participation is often governed by private platforms, unequal exposure, harassment, and data extraction.

That point changes the design goal. Privacy cannot mean disappearance from common life. A defensible public system must let people appear, speak, experiment, organize, and sometimes use pseudonyms without turning every exposure into a permanent profile or an invitation to attack. The relevant right is protected publicness: exposure with context, boundaries, and recourse.

The safety problem is relational. Friends, coworkers, household members, devices, locations, and stylistic similarities can provide evidence about someone who never made the original disclosure. A settings page organized around one account therefore cannot by itself govern collective inference. Nor can a single acceptance govern every later use of a post, search, voice sample, support chat, or uploaded document.

Persistent AI memory makes the distinction operational. A governable record should say whether an item was stated by the user, inferred by the system, imported from a connected source, supplied by another person, or generated by an agent; it should also state who can read it and whether it can affect action. The model-memory attack surface begins when those sources are flattened into one trusted past.

Public safety also requires contextual boundaries. Material shared for support, creativity, education, or civic participation should not silently become a credential, risk signal, advertising category, or durable personal premise. The problem is not publicity itself. It is exposure whose future audience, purpose, and authority cannot be contested.

The AI-Age Reading

Read in 2026, Updating to Remain the Same offers a way to analyze the platform routines beneath AI. This translation is the review's argument, not a claim that Chun wrote about today's models.

A common loop has five decisions: the interface proposes a route; the person acts or delegates; the product retains some trace; that trace may alter memory, ranking, defaults, or follow-up; and the person adapts to the revised environment. The last adaptation can become new input. This is recursive reality at the level of routine, but it should be demonstrated product by product rather than assumed from the presence of a chat box.

The safety risk is not only a wrong answer. It is an accumulating dependency that becomes hard to see once one interface is the first route to writing, searching, choosing, remembering, purchasing, or contacting other people. Frequency alone does not prove harm. More probative signals include rising exit cost, declining correction or override, skill displacement, permission expansion, failed withdrawal, and worse outcomes for people who cannot take the default route.

NIST's Generative AI Profile helps name one part of the problem: its Human-AI Configuration category includes automation bias, over-reliance, anthropomorphism, and possible emotional entanglement. The practical response is not to claim that a system feels or intends. It is to test whether interface language, memory, and repeated success encourage people to grant outputs more authority than their evidence warrants.

Agents make habituation consequential because a routine can become an external act. An assistant that moves from drafting an email to sending it, from comparing products to buying one, or from locating a file to deleting it crosses an authority boundary even if the conversational surface barely changes. Read access, memory access, tool use, outbound action, and data reuse require separate scopes; yesterday's convenience should not become tomorrow's standing consent.

Chun's account also clarifies the value of humane friction. Friction can be waste, but it can also create a pause for source inspection, consequence preview, peer consultation, refusal, or recovery. The design task is not maximum friction. It is placing reversible, legible checks at the points where a habit changes data, money, reputation, rights, or another person's world.

Governance and Safety

The governance lesson is to audit repetition and change together. A feature can be acceptable once yet harmful as a pattern, and a familiar routine can become riskier after an update. The evidence should connect interface design to retained data, altered authority, and measured outcomes.

A practical habit-change record can stay compact:

• Baseline and delta: record the old and new default, model or provider, ranking objective, memory behavior, retention rule, permission scope, affected population, and exit path. Treat a new tool, new data flow, or new external action as a material change.
• Behavior and outcome: measure more than clicks or time spent. Test error correction, override and undo rates, consent withdrawal, permission creep, task success without personalization, accessibility, skill retention where relevant, and outcomes for minors or other exposed groups.
• Control symmetry: make refusal, cancellation, reset, deletion, export, appeal, and a non-profiled route available where relevant and legally required. Do not bundle permission to perform a task with permission to retain its context or reuse its data.
• Authority and evidence: for consequential agent actions, use least-privilege credentials, previews or approval thresholds, expiry and renewal, and logs linking instruction, source, memory, tool call, output, and human decision.

Legal duties remain jurisdiction- and service-specific. DSA Article 38's non-profiled option applies to designated very large platforms and search engines; the Commission's minors guidance is non-binding; the Amazon cancellation requirements arise from a particular U.S. order; NIST guidance is voluntary; and AI Act Article 50 covers specified transparency situations rather than every habit-forming design. These limits are reasons to state product policy clearly, not excuses to collapse everything into “compliance.”

Protective habits deserve the same design attention: backups, accessible defaults, source checks, consent reminders, cooldowns before publication, limits on sensitive inference, and recurring opportunities to revise identity data. The goal is not to abolish repetition. It is to keep repetition from silently converting convenience into capture.

Where the Book Needs Friction

The book is theoretically dense and sometimes compressed. Its strength is conceptual diagnosis, not product telemetry, causal inference, or regulatory architecture. A reader cannot move directly from its account of habit to a claim that a particular notification, recommender, companion, or agent causes harm; that requires evidence about the product, population, exposure, counterfactual, and outcome.

It also predates today's mainstream generative chat, persistent assistant memory, synthetic-media pipelines, and tool-using agents. The AI application on this page is therefore an extrapolation. It is strongest when it identifies a question—what repetition and update make normal—and weakest when “habit” is used as a complete technical explanation.

Habit language can also individualize structural constraint. If a school, workplace, welfare office, landlord, insurer, or public service requires a system, repeated use may be compliance rather than preference. Network effects, procurement, labor discipline, accessibility needs, and lack of alternatives can make an interface unavoidable without making it addictive.

Conversely, frequent use is not itself evidence of injury. Backups, assistive settings, community check-ins, creative practice, medication reminders, moderation queues, and deliberate learning can extend agency. The necessary distinctions are use from dependence, convenience from coercion, personalization from recognition, and repetition from capture.

What This Changes

The practical lesson is to audit the habit and its updates, not only the advertised feature. Before a material change, preserve a baseline; after release, record what changed in data, ranking, memory, permissions, exit cost, and measured outcomes. Without that before-and-after record, “continuous improvement” can become institutional amnesia.

This connects several recurring arguments through one mechanism. Software memory explains how continuity is actively regenerated; platform grammar explains how available actions become common sense; persistent model memory gives retained traces future authority; and the Agent Tool Permission Protocol governs the point where a repeated suggestion becomes an external act.

The decisive test is interruptibility under real conditions. Can a person pause, inspect, correct, reset, refuse, export, delete, appeal, or choose another route without losing core access? Can a worker or student decline data reuse while completing the required task? Can someone distinguish what they stated from what the system inferred? Does a new version renew consent when authority materially expands?

Chun's enduring value is the reminder that technological politics often arrives as maintenance. The screen asks for an update, the feed refreshes, the assistant remembers, and the familiar route acquires new powers. Governing that continuity is less dramatic than debating a distant machine future, and more useful.

Source Discipline

This review separates four kinds of support. MIT Press and SFU establish publication and author facts; scholarly reviews provide reception, not legal or technical evidence; EUR-Lex and official Commission pages establish EU rules, guidance, and procedural status; FTC and NIST materials establish a U.S. enforcement example and voluntary risk-management context. None of them proves that a particular AI product is harmful merely because it is used often.

Scope words matter. The Commission's July 2026 Meta finding is preliminary. Its minors guidance is voluntary and non-binding. DSA duties vary by service type and size. AI Act Article 50 addresses specified transparency cases. The revised Article 4 requires measures supporting literacy, not a guaranteed individual proficiency level. NIST publications are voluntary. The Amazon example is a court order governing identified Prime flows, not a general rule for every subscription or AI agent.

The AI analysis is an explicit extrapolation from a 2016 book. It distinguishes product history, memory, analytics, recommendation, and model improvement because those processes can have different inputs, purposes, controls, and legal bases. Claims about dependence or harm still require product-specific evidence. This page does not attribute consciousness, divinity, or artificial general intelligence to any system.

Related Pages

• Programmed Visions, The Culture of Connectivity, and Filterworld on regenerated memory, platform grammar, and cultural feedback loops.
• Addiction by Design, Lurking, and No Sense of Place distinguish compulsive design, user formation, and collapsed social context.
• The Model Memory Becomes an Attack Surface and The Platform Risk Assessment Becomes the Feed's Confession apply the argument to retained context and measurable platform effects.
• Recommender Systems, Digital Services Act, Deceptive Design Patterns, AI Memory and Personalization, and AI Literacy supply definitions and legal context.
• Humane Friction Standard, Agent Tool Permission Protocol, AI Audit Trails, Notice and Appeal, and Privacy and Data turn interruptibility into operational controls.

Sources

• MIT Press, Updating to Remain the Same: Habitual New Media, publisher record, thesis, author note, edition dates, pagination, and ISBNs, reviewed August 12, 2026.
• Simon Fraser University School of Communication, Wendy Hui Kyong Chun, current faculty role and research profile, reviewed August 12, 2026.
• Simone Natale, review of Updating to Remain the Same: Habitual New Media, New Media & Society, March 2017, reviewed August 12, 2026.
• Zara Dinnen, “habit + crisis = update, somehow”, Computational Culture, November 28, 2017, reviewed August 12, 2026.
• European Union, Regulation (EU) 2022/2065, Digital Services Act, official text, especially Articles 25, 27, 28, 34, 35, and 38, reviewed August 12, 2026.
• European Commission, Guidelines on the protection of minors under the DSA, non-binding recommendations including defaults, persuasive design, and AI-chatbot safeguards, published July 14, 2025 and last updated April 22, 2026; reviewed August 12, 2026.
• European Commission, preliminary findings on the design of Instagram and Facebook, July 10, 2026 DSA press release; reviewed August 12, 2026.
• European Union, Regulation (EU) 2024/1689, Artificial Intelligence Act, official text containing Article 50 transparency duties, reviewed August 12, 2026.
• European Union, Regulation (EU) 2026/1744, July 8, 2026 amendment revising AI Act Article 4 and adding the limited Article 50(2) transition, reviewed August 12, 2026.
• European Commission, Transparency obligations under Article 50 of the AI Act, official FAQ on scope, application from August 2, 2026, and the limited transition, reviewed August 12, 2026.
• Federal Trade Commission, Bringing Dark Patterns to Light, September 2022 staff report on manipulative interface design, reviewed August 12, 2026.
• U.S. District Court for the Western District of Washington, stipulated final order in FTC v. Amazon.com, Inc., filed September 25, 2025, enrollment, consent, cancellation, and redress terms; reviewed August 12, 2026.
• Federal Trade Commission, FTC Secures Historic $2.5 Billion Settlement Against Amazon, September 25, 2025 official summary of the order and allegations, reviewed August 12, 2026.
• National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 26, 2024, voluntary guidance including Human-AI Configuration risk; reviewed August 12, 2026.
• National Institute of Standards and Technology, AI Agent Standards Initiative, current initiative page on voluntary standards, protocols, security, identity, authorization, and interoperability, reviewed August 12, 2026.

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

• Amazon, Updating to Remain the Same by Wendy Hui Kyong Chun, affiliate link, reviewed August 12, 2026.

Return to Blog · Return to Books