Blog · Review Essay · Last reviewed July 2, 2026

Computing Taste and the People Inside Recommendations

Nick Seaver's Computing Taste is an anthropology of music recommendation systems and the people who make them. Its strongest lesson for the AI era is that an algorithmic system is never just code. It is also a workplace, a theory of the user, a business model, a metaphor, a metric regime, and a set of people trying to make culture computable without admitting how strange that task is.

The book is a useful corrective to lazy talk about "the algorithm." Recommendation systems do not descend from nowhere. They are made by engineers, product managers, researchers, clients, labels, platforms, listeners, dashboards, and ideas about what listening is supposed to be.

The Book

Computing Taste: Algorithms and the Makers of Music Recommendation was published by the University of Chicago Press in 2022. The Press lists the book at 216 pages, with a paperback ISBN of 9780226822976, cloth ISBN of 9780226702261, and electronic ISBN of 9780226822969. BiblioVault records the DOI as 10.7208/chicago/9780226822969.001.0001 and places the book in the world of recommender systems, music recommendation, ethnography, machine learning, metrics, and algorithmic power.

The subject is specific: the people who build music recommendation systems. That specificity is the reason the book travels. Seaver studies a field where technical workers are not simply optimizing a neutral preference function. They are deciding what counts as music, taste, genre, context, similarity, surprise, discovery, attention, care, and success.

That makes the book a close companion to Filterworld, The Filter Bubble, What Algorithms Want, and Addiction by Design. Those books ask how ranking, personalization, optimization, and interface design reshape attention. Seaver adds the shop-floor anthropology: the system's politics are partly inside the everyday theories held by the people who make it.

The Makers

The book's most useful move is methodological. Instead of treating recommender systems as sealed technical artifacts, Seaver follows the people who design, sell, maintain, and narrate them. The result is not an apology for platforms. It is a better object of criticism.

Product teams are full of theories. They may talk about overload, discovery, personalization, boredom, loyalty, context, mood, listener journeys, genre maps, or musical space. Those theories are not decorative. They become models, labels, dashboards, targets, tests, and business cases. A theory of the listener becomes a way to measure the listener. A theory of music becomes a way to partition a catalog. A theory of success becomes a metric that pulls the system toward more of some behaviors and less of others.

This is why the review should not stop at "recommendations are manipulative" or "recommendations are helpful." Both can be true. A music recommender can expose someone to an artist they would never have found alone. It can also make the route to discovery dependent on a proprietary interface that learns how to capture, classify, and retain attention. The politics are in the arrangement.

The old myth of technical neutrality often imagines a clean pipeline: data enters, model computes, recommendation appears. Seaver shows a messier path. Workers decide what data matters, which examples count as success, which failure modes are tolerable, which clients need reassurance, which metrics survive meetings, which metaphors make the work feel ethical, and which uncertainties can be left unnamed.

Taste as Work

Computing Taste is sharp because it refuses to treat taste as a small private object hidden inside the user. Taste is made socially. It is learned through friends, scenes, radio, critics, playlists, platforms, memory, status, boredom, refusal, place, and repetition. A recommender system enters that formation process and then records part of the response as evidence.

For music platforms, this creates a deep category problem. Listening is not one thing. A person may play a song because they love it, because they are in a car with someone else, because they need background sound, because a child requested it, because a playlist auto-continued, because a song is useful for work, because the mood is temporary, or because they are trying to signal identity. The system still needs to convert the event into a usable signal.

That conversion is where culture becomes computational. Skips, repeats, saves, searches, follows, locations, devices, times of day, playlists, and inferred contexts all become proxies. Each proxy makes some parts of listening visible and leaves others outside the frame. The platform does not need to misunderstand users maliciously. It only needs to act as if the legible part of behavior is a sufficient account of taste.

This is also the bridge to AI companions and assistants. A conversational system may infer preference from prompts, edits, hesitation, rerolls, dwell time, emotional tone, and repeated topics. The same warning holds: behavior is evidence, but it is not the person. When a system treats traces as a stable self, it can become confident in an identity it helped create.

Captivation and Care

Seaver's table of contents names a central tension: recommendation is often framed as care and experienced as capture. The system promises to help the listener navigate abundance. It also has business reasons to keep the listener inside the service. The same feature can be discovery tool, retention mechanism, cultural guide, ad surface, and data collection instrument.

That ambiguity is the book's strongest AI-era contribution. A recommendation system does not have to hate users to govern them. It can be built by people who genuinely value music, diversity, surprise, serendipity, craft, and listener satisfaction. The problem is that those values operate inside metrics, investors, contracts, platform goals, and technical constraints that turn care into captivation.

This matters because many AI products now speak in the language of assistance. The tool helps you write, learn, shop, exercise, date, pray, heal, plan, code, remember, and decide. Help is real. Capture is also real. The governance problem is not solved by proving that one side is fake. It starts by recognizing that care and capture can run through the same interface.

A healthy system should make that tension inspectable. Users should know when a suggestion is driven by similarity, popularity, sponsorship, retention, social proof, inferred mood, catalog availability, contract terms, or safety filtering. Creators should know how ranking changes affect reach. Auditors should be able to see whether a system systematically narrows exposure, pushes low-cost content, amplifies already dominant catalogs, or learns to exploit vulnerable states.

The AI-Age Reading

Read in 2026, Computing Taste is no longer only about songs. It is about the cultural layer of AI systems. Recommendation has moved from "what should play next?" toward "what should the user encounter, believe, ask, write, buy, feel, and become next?" The route from music recommendation to AI assistance is not direct, but the governing pattern is similar: observe behavior, infer preference, arrange an encounter, record the response, and update the next encounter.

Generative AI changes the loop by adding production. A recommender selects from a catalog. A generative system can also create a summary, voice, image, lesson, explanation, playlist, synthetic companion, or personalized argument. That means the system may both read taste and manufacture the object through which taste is expressed.

The risk is closed-loop culture. A platform infers what travels, creators adapt to what travels, generated material imitates what traveled, users respond to the generated material, and the response becomes evidence for the next system. In that loop, culture does not simply get recommended. It gets formatted for recommendability.

Seaver's focus on makers keeps the critique grounded. The problem is not a ghostly algorithmic will. It is an institutional field in which people make choices under pressure and then describe those choices as if the system merely followed the data. For AI governance, that distinction is essential. If people make the system, people can document, contest, constrain, and redesign it.

Governance and Safety

The governance lesson is concrete: recommender systems should be treated as sociotechnical systems with accountable owners, not as invisible conveniences. A basic review should name the product goal, ranking objective, protected user groups, creator impact, ad and sponsorship influence, content supply constraints, user controls, logging practices, experiment policy, and appeal paths for people harmed by ranking.

The EU Digital Services Act gives one live policy vocabulary. Article 27 requires covered online platforms that use recommender systems to explain the main parameters of those systems and any user options to modify or influence them. Article 38 adds that very large online platforms and very large online search engines must provide at least one recommender option that is not based on profiling as defined by the GDPR.

That framework is not universal. It is EU law for covered services, and music recommendation is only one part of a larger platform environment. But it supplies a useful governance test for AI products: can users understand the main parameters, change meaningful settings, find non-profiling routes where required, separate sponsored influence from relevance, and get a record when a recommendation becomes consequential?

NIST's AI Risk Management Framework adds a lifecycle discipline. Its core functions - govern, map, measure, and manage - are a practical reminder that recommendation risk is not only an accuracy problem. Teams need to map contexts, measure effects, manage harms, and govern the organization that keeps changing the system after launch.

For AI-era curation, the minimum standard should include source transparency, ranking-change records, A/B test governance, user reset controls, creator notice for major distribution changes, privacy limits on inferred taste, independent researcher access where appropriate, and incident review for recommendation-driven harm. The point is not to abolish mediation. The point is to make mediation answerable.

Where the Book Needs Friction

The book is strongest as ethnography and theory. Readers looking for a simple consumer-protection checklist or a technical explainer of modern recommender architectures will need companion sources. Seaver is less interested in providing a policy recipe than in showing how algorithmic culture is made in practice.

There is also a scope limit. Music recommendation is unusually rich because music is affective, social, contextual, repetitive, and identity-laden. Lessons from music should not be flattened into every recommender domain. A medical triage system, hiring screen, welfare fraud model, shopping recommender, search engine, and short-video feed all have different stakes, records, remedies, and legal duties.

The most important limit is institutional access. Ethnography can make technical workers visible, but it can also depend on what companies, interviewees, and field sites allow to be seen. That is not a failure of the book. It is part of the argument. Algorithmic systems are difficult to study because the knowledge needed to govern them is held inside firms, contracts, dashboards, experiments, and logs that publics often cannot inspect.

For this site, the right reading is not "the makers are innocent" or "the makers are villains." The better reading is that makers are part of the evidence trail. Their metaphors, incentives, dashboards, and informal theories matter because they become machinery.

What This Changes

Computing Taste gives the archive a way to talk about the human interior of algorithmic systems without falling back into the comforting idea that human involvement automatically makes a system humane.

The book asks a hard question: what happens when people who care about culture build machines that must count culture in order to operate at platform scale? The answer is not that culture dies. The answer is that culture is translated into signals, spaces, clusters, metrics, journeys, and experiments. Each translation helps the machine act, and each translation leaves something behind.

For AI agents, assistants, tutors, companions, and answer engines, the same discipline applies. Do not ask only what the model outputs. Ask what theories of the user, task, world, and good outcome were built into the product. Ask which traces become evidence. Ask which forms of refusal, ambiguity, mood, taste, and context the system cannot see. Ask who gets to revise the theory when the user changes.

Seaver's book matters because it rehumanizes recommendation without sentimentalizing it. The people inside the system are real. So are the power relations, metrics, and loops they build. A serious AI culture has to study both.

Source Discipline

This review uses the University of Chicago Press and BiblioVault for bibliographic metadata, chapter framing, DOI, and publisher description; Amazon for the affiliate retail link; EUR-Lex for Digital Services Act claims; NIST for risk-management vocabulary; and ACM RecSys for the field context around recommender systems.

Claims about law are jurisdiction-specific. The Digital Services Act examples here describe EU obligations for covered online platforms, with additional requirements for very large online platforms and very large online search engines. NIST material is voluntary risk-management guidance, not a statute.

The review treats "recommendation" as a sociotechnical system rather than a single algorithm. When the page says the system acts, it means people, interfaces, data practices, ranking models, metrics, contracts, and product goals acting together.

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