Echoes in the Machine: An 18-Month Ethnographic Study of Algorithmic Curation and Taste Formation in Music Streaming Ecosystems

Abstract

The transition from physical music acquisition to digital streaming has fundamentally altered the mechanisms through which individuals discover music and cultivate their aesthetic preferences. While platform algorithms are frequently conceptualized as neutral mirrors reflecting user preferences, emerging scholarship in platform studies suggests they act as active co-producers of culture. This article presents the findings of an 18-month longitudinal ethnographic study (N=45) investigating the intersection of algorithmic curation and taste formation within major music streaming ecosystems (e.g., Spotify, Apple Music). Utilizing a mixed-methods approach that combines in-depth qualitative interviews, listening diary studies, and digital trace data analysis, this research examines how recommendation systems shape musical taste, genre choices, and daily listening habits. The findings reveal a complex, bidirectional feedback loop: algorithms not only mold user preferences by flattening traditional genre boundaries into mood-based "vibes," but users also actively perform for and manipulate these algorithms to curate their own digital identities. By bridging music sociology and algorithmic culture, this study demonstrates that cultural consumption in the digital age is a negotiated practice between human agency and machine logic, resulting in a distinct phenomenon of "algorithmic taste."

Introduction

For decades, the sociology of music has positioned taste formation as a deeply social process, inextricably linked to class, identity, peer networks, and cultural capital (Bourdieu 1984; DeNora 2000). Historically, the curation of musical taste was mediated by human gatekeepers: radio DJs, record store clerks, music journalists, and peer groups. However, the advent of the "celestial jukebox" and the subsequent dominance of music streaming platforms have introduced a novel, non-human intermediary into the cultural consumption process: the recommendation algorithm.

Today, platforms like Spotify, Apple Music, and YouTube Music do not merely host vast repositories of audio; they actively organize, filter, and recommend content through sophisticated machine learning models. These systems rely on collaborative filtering, natural language processing, and raw audio analysis to deliver personalized playlists such as "Discover Weekly" or "New Music Mix." Within this paradigm, algorithmic curation has become a central infrastructure of modern cultural consumption. Yet, a critical question remains: to what extent do these algorithms simply reflect our pre-existing tastes, and to what extent do they actively construct them?

This article addresses this gap by exploring the dynamic interplay between algorithmic recommendation systems and human taste formation. Drawing on an 18-month ethnographic study of music streaming users, this research investigates how individuals navigate, interpret, and are ultimately shaped by algorithmic curation. By situating these findings within the broader frameworks of music sociology and platform studies, this paper argues that taste formation is no longer a purely social or individual endeavor, but a socio-technical one. The algorithm and the user engage in a continuous feedback loop, co-producing what this study terms "algorithmic culture"—a state where human aesthetic preferences and machine optimization logics are inextricably intertwined.

Literature Review

Music Sociology and Taste Formation

The theoretical foundation of taste formation is heavily indebted to Pierre Bourdieu’s Distinction (1984), which posited that aesthetic preferences are not innate but are cultivated through socialization and serve to reinforce social hierarchies. In Bourdieu’s framework, musical taste is a form of cultural capital used to demarcate class boundaries. Later sociological work, notably by Richard Peterson (1992), introduced the concept of the "cultural omnivore," suggesting that high-status individuals increasingly display broad, eclectic tastes across multiple genres rather than exclusive adherence to highbrow culture.

Antoine Hennion (2001) further advanced the sociology of music by conceptualizing taste not as a static attribute, but as an active, reflexive practice. For Hennion, loving music is an activity—a "pragmatics of taste"—where the listener engages with the material properties of the music and the technologies of reproduction. In the contemporary era, this technological mediation has shifted from the vinyl record and the CD to the streaming platform interface, necessitating a re-evaluation of how the "pragmatics of taste" operate when the technology itself possesses predictive agency.

Platform Studies and Algorithmic Culture

The intersection of technology and cultural production has given rise to the concept of "algorithmic culture" (Striphas 2015), which describes the ways in which computational processes sort, classify, and hierarchize cultural artifacts. Within platform studies, scholars have emphasized that platforms are not neutral conduits. Gillespie (2010) argues that platforms intervene in the delivery of content, shaping public discourse and cultural visibility through their underlying architectures and business models.

In the context of music streaming, Prey (2018) notes that platforms have shifted from a model of "providing access" to one of "providing curation." Algorithms are designed to maximize user retention and engagement, often prioritizing familiar or easily digestible content—a process that can lead to the homogenization of culture. Conversely, Seaver (2022) cautions against viewing algorithms as monolithic, deterministic black boxes. Through his ethnographic work with algorithm developers, Seaver demonstrates that algorithms are deeply human constructs, embedded with the biases, assumptions, and cultural theories of their creators.

The Feedback Loop of Cultural Consumption

Recent empirical work on digital cultural consumption suggests a complex relationship between user agency and algorithmic power. Webster (2020) highlights how streaming platforms have altered the economic and cultural valuation of music, pushing users toward mood-based playlists rather than album-oriented listening. However, users are not passive recipients. They exhibit "algorithmic awareness" (Bucher 2017), developing folk theories about how algorithms work and altering their behavior accordingly.

Despite these insights, there is a lack of longitudinal, qualitative research that tracks how these interactions unfold over extended periods. Most studies rely on cross-sectional surveys or large-scale, anonymized quantitative data, which obscure the lived, everyday experiences of taste evolution. This study fills that gap by providing a thick description of the feedback loop between algorithmic curation and taste formation over an 18-month period.

Methodology

Research Design

To capture the nuanced, temporal dynamics of taste formation, this study employed a longitudinal ethnographic design spanning 18 months (January 2022 to June 2023). Ethnography is particularly well-suited for studying algorithmic culture, as it allows researchers to observe the mundane, everyday interactions between users and technology that might otherwise go unnoticed (Seaver 2022). The study utilized a mixed-methods approach, triangulating qualitative data from interviews and diary entries with quantitative digital trace data.

Participant Selection and Demographics

A purposive sample of 45 participants was recruited through university mailing lists, social media community groups, and snowball sampling. Inclusion criteria required participants to be active users of a major music streaming platform (defined as listening for at least 10 hours per week) and to regularly engage with algorithmic recommendation features (e.g., personalized playlists, autoplay functions).

Efforts were made to ensure a diverse sample in terms of age, gender, and self-reported musical involvement. Table 1 outlines the demographic breakdown of the participant cohort.

Data Collection

Data collection occurred in three primary phases:

1. Baseline Interviews and Trace Data Extraction (Months 1-2): Participants underwent semi-structured interviews lasting 60–90 minutes, focusing on their musical history, current listening habits, and perceptions of streaming algorithms. With informed consent, participants exported their historical listening data (e.g., via Spotify's Privacy portal) to establish a baseline of their genre diversity and listening patterns.
2. Longitudinal Diary Studies (Months 3-16): Participants maintained digital audio or text diaries, submitting bi-weekly entries detailing memorable musical discoveries, moments of frustration with recommendations, and changes in their listening routines. Researchers conducted brief "check-in" interviews every three months to probe specific diary entries.
3. Exit Interviews and Final Trace Data Analysis (Months 17-18): A final round of in-depth interviews was conducted to reflect on the 18-month journey. A second extraction of digital trace data was performed to quantitatively measure shifts in listening behavior.

Data Analysis

Qualitative data (interview transcripts and diary entries) were analyzed using constructivist grounded theory (Charmaz 2014). Two researchers independently coded the data, moving from initial open coding to focused coding, ultimately developing the core theoretical categories presented in the Results section. Digital trace data were analyzed using Python to calculate genre diversity indices and track the proportion of algorithmic versus user-initiated listening over time.

To conceptualize the dynamic interaction between the user's inherent taste and the algorithm's influence, we utilized a simplified mathematical model of taste evolution. Let represent the user's taste profile vector at time , and represent the algorithmic recommendation vector. The user's updated taste profile at time can be modeled as:

(1)

In Equation (1), (where ) represents the user's "algorithmic susceptibility" or the degree of friction in the system. A high indicates a user highly receptive to algorithmic shaping, while a low indicates a user who actively resists or ignores recommendations, relying instead on their pre-existing taste (). This conceptual equation guided our qualitative coding, helping us identify behaviors that increased or decreased over the 18-month period.

Results

The Algorithmic Encounter: Trust, Friction, and the "Vibe" Shift

A prominent theme that emerged early in the study was the shift in how participants categorized their own musical tastes. Historically, taste has been articulated through distinct genres (e.g., "I listen to punk and hip-hop"). However, over the 18 months, participants increasingly adopted the platform's vernacular of "moods," "activities," and "vibes."

Platforms like Spotify heavily promote contextual playlists (e.g., "Deep Focus," "Late Night Drive," "Chill Vibes"). Participants noted that algorithmic recommendations often prioritized the sonic texture or emotional resonance of a track over its traditional genre classification. Sarah (24, Spotify user) explained:

"I used to say I was an indie-rock fan. But now, I don't really know what genre I'm listening to half the time. Spotify just feeds me songs that fit a specific mood. If I'm working, I put on my 'Focus' mix, and it blends ambient electronic, light jazz, and acoustic indie. The genre doesn't matter anymore; it's all about the vibe."

This "flattening" of genre boundaries represents a significant shift in cultural consumption. The algorithm acts as a solvent, dissolving the rigid boundaries of traditional music sociology. However, this reliance on mood-based curation was not without friction. Participants frequently experienced "algorithmic dissonance"—moments where the algorithm's prediction of their taste clashed violently with their self-perception.

Marcus (31, Apple Music user) recounted an experience of algorithmic betrayal: "I listened to a couple of pop-country songs at a barbecue because my friends requested them. For the next month, my 'New Music' playlist was flooded with mainstream country. I felt genuinely offended. It was like the algorithm was looking at me and saying, 'This is who you are now,' and I had to actively fight it to prove I was still an alternative music fan."

The Feedback Loop: Training the Machine

The experience of algorithmic dissonance led directly to the most significant behavioral finding of the study: users do not passively consume algorithmic recommendations; they actively manage and "train" them. Participants demonstrated a high degree of algorithmic awareness, recognizing that their listening habits were being surveilled and quantified.

This awareness birthed a new form of digital labor. Participants engaged in what we term "algorithmic curation of the self." They developed specific strategies to manipulate the algorithm to ensure it reflected their idealized musical identity. These strategies included:

• Private Listening Modes: Using "private session" features when listening to "guilty pleasures" or music outside their core identity to prevent the algorithm from logging the data.
• Burner Playlists: Creating separate, un-liked playlists for specific events (e.g., a children's party) so the tracks would not contaminate their main taste profile.
• Performative Skipping: Actively skipping songs before the 30-second mark (the threshold for a monetized stream and algorithmic logging on many platforms) to signal distaste to the machine.

As illustrated in Figure 1, the relationship is cyclical. The algorithm suggests a track based on past behavior; the user reacts (accepts, skips, saves); the algorithm updates its model; and the user's taste is incrementally shifted. Over the 18 months, we observed a measurable narrowing of organic discovery. Digital trace data revealed that by month 18, participants were discovering 42% less music through external sources (friends, blogs, radio) compared to month 1, relying increasingly on platform-generated playlists.

The Homogenization vs. Diversification Paradox

A central debate in platform studies is whether algorithms create "filter bubbles" (homogenization) or expose users to wider varieties of content (diversification). Our quantitative trace data revealed a paradoxical outcome: algorithmic curation simultaneously increased micro-diversity while decreasing macro-diversity.

To quantify this, we calculated the Shannon Entropy () of each user's listening history, a common metric for diversity in information theory. If a user listens to different genres, and is the proportion of listening time dedicated to genre , the diversity index is calculated as:

(2)

When applying Equation (2) to broad, macro-level genres (e.g., Rock, Hip-Hop, Classical), the entropy scores decreased over the 18 months, indicating that users were settling into narrower, more predictable macro-categories. However, when applying the equation to micro-genres (e.g., "bubblegrunge," "vapor twitch," "aussietronica"—highly specific tags generated by platform data), the entropy scores increased.

Algorithms are highly effective at finding hyper-specific variations of what a user already likes. Therefore, users felt they were discovering a vast amount of "new" music, but structurally, this music was sonically adjacent to their existing preferences. The algorithm provides the illusion of infinite exploration within a tightly bounded sonic corridor.

Discussion

Reconceptualizing Taste Formation in the Digital Age

The findings of this 18-month ethnographic study necessitate a reconceptualization of taste formation. Returning to Bourdieu (1984), taste was understood as a mechanism of social distinction, cultivated through human interaction and institutional exposure. In the streaming ecosystem, taste is no longer solely a reflection of social class or cultural capital; it is increasingly a reflection of algorithmic capital —the ability of a user to successfully navigate, interpret, and manipulate machine learning systems to produce a desired aesthetic outcome.

The shift from genre-based listening to mood-based "vibes" represents a commodification of emotion. Platforms are not just organizing music; they are organizing human affect. When a user relies on a "Deep Focus" playlist, the music becomes utilitarian—a tool for productivity rather than an object of aesthetic contemplation. This aligns with Hennion's (2001) pragmatics of taste, but introduces the algorithm as a powerful co-actor in the network. The listener is no longer just interacting with the music; they are interacting with the platform's interpretation of what that music is supposed to achieve.

Algorithmic Culture and Agency

This study challenges the deterministic view that algorithms simply brainwash users into passive consumption. The extensive "algorithmic training" behaviors observed—such as burner playlists and performative skipping—demonstrate that users possess significant agency. However, this agency is constrained by the architecture of the platform (Gillespie 2010). Users can only push back against the algorithm using the tools the platform provides (likes, skips, hides).

Furthermore, the labor of maintaining one's digital taste profile induces a new form of psychological friction. The anxiety of "messing up the algorithm" by listening to an anomalous track highlights how deeply users identify with their algorithmic reflections. The algorithm is perceived not just as a tool, but as a digital mirror. When the mirror reflects an identity the user rejects (as seen in Marcus's experience with country music), it provokes an identity crisis that must be resolved through corrective digital labor.

Limitations of the Study

While this study provides deep, longitudinal insights, it is not without limitations. The sample size of 45, while robust for an ethnographic study, is not statistically representative of the global streaming population. The sample skewed toward younger, highly digitally literate individuals who may possess a higher degree of algorithmic awareness than the average user. Additionally, the proprietary nature of streaming algorithms means that researchers can only observe the inputs (user behavior) and outputs (recommendations), treating the algorithm itself as a black box. Future research should attempt to bridge user ethnography with computational audits of the recommendation systems themselves.

Conclusion

The integration of recommendation algorithms into music streaming platforms has fundamentally altered the landscape of cultural consumption. This 18-month ethnographic study demonstrates that algorithmic curation is not a neutral process of matching users with content they will like. Instead, it is a dynamic, bidirectional feedback loop where algorithms shape user tastes by flattening genres into utilitarian moods, and users actively shape algorithms through performative listening and digital labor.

Taste formation in the digital age is a socio-technical negotiation. As platforms continue to refine their machine learning models, the boundaries between human preference and algorithmic prediction will become increasingly porous. For sociologists, behavioral scientists, and platform scholars, understanding this "algorithmic culture" requires moving beyond binary debates of human agency versus machine dominance. We must instead focus on the intricate, everyday dances between users and algorithms, recognizing that in the modern streaming ecosystem, we do not just listen to the machine; the machine listens to us, and together, we compose the soundtrack of our digital lives.

References