<serendipidadealgorítmica/>

A study of how digital platforms turn personalization into a synonym for discovery.

And of what is left of the listener's sensibility, the capacity to be moved before you can name it, when even chance becomes something calculated.

A manageable accident is no accident.
<a passage at random>
<the question>

What happens to music discovery when recommendations are personalized?

A song on the bus radio. A track someone saved on a USB stick. A tune heard at a party, without knowing who sings it. For a long time, discovering music happened through encounters like these, sometimes by pure chance.

Today much of our listening runs through platforms that log every play, pause and skip, and use that data to decide what comes next. The recommendation arrives ready-made, made for you.

How does Spotify build, technically and discursively, the equivalence between personalization and music discovery, and what conditions does that construction set for listening?

To think through this question, the research turns to an old word, coined to name the encounter with what you were not looking for.

<serendipity>

What is serendipity?

Chance, sagacity and the encounter with the unexpected.

Horace Walpole coined the word in 1754, after the three princes of Serendip, who kept making discoveries “by accidents and sagacity” of things they were not in quest of. Robert K. Merton and Elinor Barber stress the second half. Chance alone does not produce discovery. Someone has to recognize what turned up.

Music works the same way. A track heard by chance only becomes a discovery when it meets a repertoire and an openness in the listener. Two figures from Yoruba cosmology help the research think of this chance as an opening to encounter.

Èṣù

The orisha of the crossroads. In a narrative collected by Reginaldo Prandi, Orunmila's nature was destiny and Exu's was accident. An accident that is not a failure but the condition for movement.

Ọ̀ṣun

The orisha of fresh waters, “that which runs, seeps, flows”. Water finds its way as it runs, like a sagacity attuned to movement.

The system

It starts from what has already been recorded. In the research's reading, where there would be a crossroads, the direction tends to arrive in advance.

Circuit diagram of an AM and FM radio, in light blue
// Merton and Barber
serendipity = accident + sagacity
// simplified sketch of recommendation, not how Spotify actually works
listening_traces →︎ content_selection →︎ recommendation →︎ possible_discovery?
<sources for this section>
  • Robert K. Merton and Elinor Barber, The travels and adventures of serendipity, Princeton University Press, 2004. Walpole's phrase is on p. 2.
  • Reginaldo Prandi, Mitologia dos Orixás, Companhia das Letras, 2001, p. 76, cited in Raimundo and Hinkel, 2022.
  • Joseph M. Murphy and Mei-Mei Sanford (eds.), Ọ̀ṣun across the waters, Indiana University Press, 2001, p. 5.
full references in the thesis
<on the platforms>

algorithmic serendipity on streaming platforms

The phrase brings together two ideas that seem to pull against each other, unexpected discovery and data-driven selection.

Spotify works neither by pure chance nor by absolute predictability.

Geoffray Bonnin and Dietmar Jannach describe the challenge of music recommendation. Close enough to your taste to be accepted, new enough to surprise. The idea of organized chance was already there in 1968, in the Cybernetic Serendipity exhibition in London.

familiarity ←︎   →︎ novelty

That gap has a history. Before the platforms, other mediators already decided what reached our ears.

<sources for this section>
  • Jasia Reichardt (ed.), Cybernetic Serendipity: the computer and the arts, Studio International, 1968.
  • Geoffray Bonnin and Dietmar Jannach, “Automated generation of music playlists: survey and experiments”, ACM Computing Surveys, 2014.
full references in the thesis
<mediation>

Before the algorithm, mediation

From radio and record labels to streaming.

Listening to music was never a direct practice. Radio, record labels, television and informal circuits have always chosen what reached our ears, and they still coexist. In Ana Rieper's documentary Vou Rifar Meu Coração, brega, a Brazilian popular romantic style, circulates through parties, bars, raffles and copied discs.

  1. Radio

    Editorial, commercial or the DJ's own taste.

  2. Record Labels

    Decided what deserved to be pressed.

  3. TV and MTV

    Image and sound building hierarchies of visibility.

  4. Bootleg CDs

    Street markets and vendors, on the margins of the industry.

  5. Streaming

    Curation loses its face.

What streaming changes is not whether there is mediation, but how it is built.

The radio host had a voice and could be challenged. The recommendation arrives without stating its criteria. How does Spotify describe its own selection?

<sources for this section>
  • Ana Rieper (dir.), Vou Rifar Meu Coração, documentary, Aruac Filmes, 2011.
full references in the thesis
<the research>

What the research looks at

Spotify's discourse, the corpus and the method.

read the full text
<how the research was done>
  1. The object

    What Spotify says about its recommendation systems. Not the code, nor what listeners experience.

  2. The corpus

    28 posts from 2015 to 2025, 15 from the Spotify for Artists Blog and 13 from the Spotify Engineering Blog.

  3. The method

    Discourse Analysis, drawing on Eni Orlandi and Michel Pêcheux, along three axes, personalization, discovery and control.

  4. The software

    Correspondence Factor Analysis in IRaMuTeQ, to map the vocabulary of the corpus.

The research reads the texts in which Spotify speaks in its own name. The words, the metaphors and the silences.

Glenn McDonald, a former data engineer at the company, points out that there is no single “Spotify algorithm”, there are many. The corporate discourse speaks in the singular, and that singular is what the research questions.

The study describes a discourse and the conditions it creates for listening. It does not measure effects on individual listeners.

Next, what the analysis made it possible to see.

<sources for this section>
  • Glenn McDonald, You have not yet heard your favourite song: how streaming changes music, Canbury Press, 2024, p. 43.
  • Eni Puccinelli Orlandi, Análise de discurso: princípios e procedimentos, Pontes, 2007.
  • Michel Pêcheux, Semântica e discurso: uma crítica à afirmação do óbvio, Editora da Unicamp, 1988.
  • Full list of the 28 posts in Appendix A of the thesis.
full references in the thesis
<findings>

What the discourses reveal

Personalization, discovery and the limits of the algorithmic profile.

Each finding starts from a post in the corpus and separates what Spotify says from the reading the research makes of it.

  • ENGNov 2015

    An origin that looks like friendship

    The post says Discover Weekly is “the playlist a friend would make for you”. Reading: technical mediation is recast as friendship.

    origin effect, Eni Orlandi
  • ENGDec 2021

    381 million versions

    The post speaks of “381 million individual versions of Spotify”. Reading: the number works as an argument. The calculation comes from similarity between profiles.

    subjectivist illusion, Michel Pêcheux
  • ENGOct 2023

    Taste that gets contaminated

    The post presents the feature for removing plays from your taste profile. Reading: taste becomes something that can be soiled and purified.

    dataism, Rafael Evangelista
  • ENGApr 2023

    Human and machine in one word

    The post coins the term algotorial. Reading: the word fuses human curation and calculation, and erases the tension between them.

    technodiscourse, Marie-Anne Paveau
  • ARTFeb 2021

    The cost in the corner of the page

    The post announces Discovery Mode through one artist's success story. Reading: the lower royalties stay at the margins of the text.

    silence, Eni Orlandi
  • ARTJan 2022

    What counts as discovery

    The guide splits discovery into three paths, made by editors, made for you and made by you. Reading: a friend's tip or a song at a party fall outside the grid.

    discursive device, Marie-Anne Paveau

In the factor analysis, the vocabulary splits into two poles, accounting for 37.55% and 27.37% of the variance. Calculation on one side, the discovery narrative on the other.

<engineering>llm, semantic, model, embedding, recommendation
<curation and listening>playlist, editor, fan, artist, music

The system records traces of behavior before they add up to a profile and hands them back to the listener as identity. Antoinette Rouvroy and Thomas Berns call this algorithmic governmentality, a steering by statistical projections, under the appearance of free choice.

Spotify turns personalization into a synonym for discovery and presents its mediation as neutral infrastructure.

For Gilbert Simondon, every individual carries an excess that no form exhausts. The profile captures fragments and hands back statistics. What is left out is the taste that does not yet know it is a taste.

Discovery is still possible. What the research discusses are the conditions under which it happens and the way the platform names it.

<posts analyzed>
  • Spotify Engineering Blog, “What made Discover Weekly one of our most successful feature launches?”, Nov. 2015.
  • Spotify Engineering Blog, “How Spotify uses ML to create the future of personalization”, Dec. 2021.
  • Spotify Engineering Blog, “Exclude from your taste profile”, Oct. 2023.
  • Spotify Engineering Blog, “Humans + Machines: algotorial technology”, Apr. 2023.
  • Spotify for Artists Blog, “The 8 biggest highlights for artists from Spotify Stream On 2021”, Feb. 2021.
  • Spotify for Artists Blog, “Learn how fans discover your music: Made to Be Found”, Jan. 2022.
<authors cited>
  • Eni Puccinelli Orlandi, Análise de discurso: princípios e procedimentos, 2007.
  • Michel Pêcheux, Semântica e discurso, 1988.
  • Rafael Evangelista, “Por uma etnografia do poder na inteligência artificial, no capitalismo de vigilância e no colonialismo digital”, 2025.
  • Marie-Anne Paveau, Análise do discurso digital: dicionário das formas e das práticas, 2021.
  • Antoinette Rouvroy and Thomas Berns, “Governamentalidade algorítmica e perspectivas de emancipação”, Revista Eco-Pós, 2015.
  • Gilbert Simondon, “A individuação à luz das noções de forma e de informação: introdução”, [1958] 2005.
full references in the thesis
read the full thesis
<about the researcher>
Black and white portrait of Vinicius Miranda

Vinicius Miranda David Paulino

@viniciusmirandavid

Professor of Communication and Advertising. Master's in Social Sciences from the Federal University of São Paulo (UNIFESP), postgraduate specialization in Media, Information and Culture from the University of São Paulo (USP/CELACC), and a bachelor's degree in Advertising from Faculdade Reinaldo Ramos / Centro de Educação Superior Reinaldo Ramos (FARR/CESREI), in Campina Grande, Paraíba, Brazil. Researcher at PimentaLab/UNIFESP, where he studies communication, the arts, artificial intelligence and digital culture.

thesis
Subjetivações tecno-mediadas e a serendipidade: o algoritmo curador do Spotify (Techno-mediated subjectivations and serendipity: Spotify's curator algorithm)
advisor
Prof. Dr. Henrique Zoqui Martins Parra
funding
CAPES scholarship, Finance Code 001
program
Graduate Program in Social Sciences (PPGCS), School of Philosophy, Letters and Human Sciences, Federal University of São Paulo (UNIFESP), Guarulhos, Brazil
read the full text