"Spotify Teardown" by Anna Johansson, Maria Eriksson, Patrick Vonderau, Pelle Snickars, and Rasmus Fleischer (Getty/MIT Press)

What happens when you create a fake music record label and upload bad music to Spotify

A scholarly experiment to figure out how Spotify works (and how to trick it)

Maria Eriksson - Rasmus Fleischer - Anna Johansson - Pelle Snickars - Patrick Vonderau
February 17, 2019 7:00PM (UTC)

Excerpted from "Spotify Teardown: Inside the Black Box of Streaming Music" by Maria Eriksson, Rasmus Fleischer, Anna Johansson, Pelle Snickars, and Patrick Vonderau (MIT Press, 2018). Reprinted with permission from MIT Press.

In October 2014, the album Election Music by Heinz Duthel was uploaded to services such as Spotify, iTunes, and Google Play. The album consisted of one electronic tune that had been modified thirteen times according to the voting results in different municipalities in the 2014 Swedish government election. The tempo had been adjusted on each track according to the number of voters in specific voting districts, so that a high number of registered voters corresponded to a fast beat, and a low number of voters corresponded to a slow beat. Specific instruments in the tunes had also been paired with particular political parties, making it possible to “listen to” local voting results and compare tracks with one another. Hence, album listeners could, for example, detect the slow pace (and thus the low number of voters) on the track “Rinkeby,” a low-income suburb on the outskirts of Stockholm. Listeners could also discover how this track differed from the much faster rhythm on the track “Hörken,” a small rural county in midwestern Sweden with many registered voters. On “Hörken,” the prominent sound of a siren also indicated the local success of the extreme right-wing party Sverigedemokraterna (Swedish Democrats), which hit historic records in the 2014 election.

As readers may know, Heinz Duthel had already made a name for himself in 2008, when he appeared as the author of over two hundred books, many of which are still available for purchase on Amazon or through Apple’s iBooks store. Duthel’s collected works include biographies of Adolf Hitler, Angela Merkel, Pablo Picasso, Muhammed Ali, Joseph Stalin, and Charles Darwin, as well as writings on topics such as anarchism, the internet drug trade, cosmic intelligence, dementia, the philosophy of mathematics, alternative medicine, dialectical materialism, conspiracy theories, the Great Depression, and Thai massage. One of Duthel’s publications in particular caused a debate in 2010, when it turned out that Amazon was still marketing and selling his e-book entitled WikiLeaks Documents Expose US Foreign Policy Conspiracies: All Cables with Tags from 1 to 5000, even though Amazon had just expelled WikiLeaks from its servers in protest of its release of classified government documents. Much like the rest of Duthel’s releases, this book was essentially a reprint of content scraped from the web and thus the product of algorithmically aided bot-authorship. Heinz Duthel is most likely a software agent designed to mass produce literature—albeit a bot writer with a particular taste for classical historical figures, foreign politics, and new age spirituality.


As it turns out, having music launched on Spotify is not always a straightforward process. After recording a series of sounds, our search for suitable aggregators who were willing to distribute them began on Spotify’s website, which lists a series of recommended aggregators. Initially, we explored the deals offered by TuneCore, IndigoBoom, RouteNote, and CD Baby. Like most music aggregators, these companies either charge musicians an annual fee for distributing their sounds or distribute music against a certain percentage of future revenue. Even though we eventually managed to sign a deal with an aggregator for all of our releases, the process of doing so revealed the arbitrary methods by which music is approved for commercial streaming services.

For instance, we found that our album Fru Kost—which contained a recording of breakfast sounds, such as the sound of coffee being poured—was rejected twice on our first attempt to have it released. Once, we were kindly and politely told that this was “not the kind of content” the aggregator was looking to sign up at the moment. Another time, our sounds were rejected with a blunt notification stating that the distributor in question “only distributes music.”

Whether a human or an automatic software system was behind this editorial decision remains unknown, yet the initial rejections that we received indicate that music aggregators are gatekeepers who perform the job of cleaning, sorting, and selecting which types of sounds end up on platforms such as Spotify. Thereby, they also play a crucial role in deciding what counts, or does not count, as music in today’s digital landscape.


Music aggregators such as Record Union—a service that ended up accepting our breakfast sounds against a $20 yearly fee—claim to “liberate music” and “democratize the access to global digital distribution.” However, their very existence indicates that the Spotification of media has raised new and fuzzy barriers around digital music, as compared to earlier “open” Web 2.0 sites, such as Myspace. We also learned that what binds these aggregator companies together is not only their financial setup but also the strategies by which they prepare music and artists for entering digital marketplaces. One has to go through several standardized vetting processes before any songs or albums can be submitted. Our careful documentation of these procedures has been a key source of insight into how music streaming infrastructures operate today.

For one, we learned that signing up with an aggregator involves taking several steps to name and categorize sound content, steps that are central to the process of making digital music searchable and commodifiable. Before we could submit our sounds, the aggregator RouteNote asked us to choose between thirty-nine different genre specifications, which were organized in two steps: “first genre” and “second genre.” These options included traditional genres, such as “Electronic” and “Instrumental,” but also more curious genres, such as “Anime,” “Indian,” “Fitness & Workout,” “Enka,” and “Kayokyoku”—not to mention “German Pop” and “German Folk.”

Such genre attributions speak to the micropolitics of (self-)representation that artists are encouraged to participate in. (Not attributing a genre to our sounds wasn’t an option.) This highlights that genres continue to matter, despite repeated modernist claims that genres have lost their importance in relation to the contemporary music scene. Music metadata, such as genre specifications, fundamentally influence how streamed music moves and is displayed, not least because metadata underlie music recommendation algorithms and other automatized ways of bundling musical pieces together. Recorded music simply needs metadata and paratextual materials, or else it is extremely difficult to find or sell. Within the ecosystem of music streaming, aggregators perform the crucial role of prompting and collecting such music metadata, thus making music and artists “algorithm ready.”


Yet aggregators do not just collect metadata from individual artists when they sign a contract. They are also responsible for assembling pieces of metadata from other more obscure and centralized sources. Once our different sounds had been uploaded, for example, we were struck by the lack of descriptive and contextual data surrounding our music, such as who we were and where our music came from. Given the extensive amount of information we had provided to our aggregators, this seemed somewhat paradoxical. Even though our sounds were available on Spotify and other services, we were not in charge of our own artist pages. In fact, there was no possibility at all for us to directly intervene in how our music was being presented by textual and visual means. The process of adding artist biographies also proved to be tedious, since control over artist metadata is frequently outsourced in several steps. Indeed, music metadata now forms an industry in itself, as is shown by the multitude of companies that deal with such data.

For instance, we found out that Spotify gathers its artist biographies from AllMusic, a website that provides reviews, search functions, and recommendations concerning music. In turn, AllMusic extracts its artist biographies from TiVo (formerly Rovi and Macrovision). TiVo was in full ownership of AllMusic until 2013, when it sold off the consumer access side to All Media Network, while still controlling its content licenses. TiVo now specializes in metadata provision, intellectual property management, and analytics provision within the culture industries. Since its inception in 1983, the company has been a major player in antipiracy debates over copyright protection. TiVo can be described as an “infomediary,” that is, a company that monitors, collects, and controls information about cultural products in ways that affect how users find and experience them. To adjust how our sound materials were presented on Spotify—and other streaming services—we would therefore need to get into direct contact with a global metadata corporation that is in control of large-scale music metadata archives and databases.


Nowadays, it is possible to submit artist biographies to TiVo via email. But in the fall of 2014, when we first tried to have our sounds released, TiVo only accepted physical copies of artist information that were sent to its headquarters in Santa Clara, California. According to the instructions then available, every submission was supposed to contain a press kit, including an artist biography, photographs of the musician(s), and physical copies of the music concerned. TiVo (or Rovi, as the company was called at the time) then promised to save and distribute all the information it had been provided, including the recorded music, for future purposes. On one occasion, we made an attempt to send such a press kit—including a recorded CD—to the office in Santa Clara, although the letter must have been rejected or lost along the way, since the biographical details were never updated on any streaming services. The difficulties in taking control of our artist biographies reveal how a small number of actors have secured a vital role in regulating narratives around streamed music. While Spotify may be the end point where such metadata is presented to the public, the same data is collected, stored, and managed by a number of other services.

From a slightly different angle, we have also used distributed sounds as a basis for investigating the financial logics and premises of streaming services. For example, the release of breakfast sounds by Fru Kost was used to critically assess the issue of limited payouts to artists. This was done by examining if it was possible to analyze Spotify’s business model by automating fake Spotify listeners. The experiment resembled a Turing test, in which we asked ourselves what happens when—not if—streaming bots approximate human listener behavior in such a way that it becomes impossible to distinguish between a human and a machine. Does Spotify care if humans or machines are listening to its music? And in what ways are such differentiations nestled within streamed financial logics?

At Humlab, we therefore set up a so-called SpotiBot experiment, in which we programmed multiple bots to repeatedly play one of our self-distributed sounds: the song “Kaffe” (coffee) by Fru Kost. Our record label releases thus became a testing ground for exploring Spotify’s infrastructure for royalty payments. In detail, the bots that did the “listening” were scripted algorithms that exhibited human-like behavior. Using Web UI testing frameworks, the SpotiBot engine roughly consisted of three hundred bots that were put into action in approximately fifty parallel Spotify sessions. These bots were simultaneously put on a fixed repetition scheme that involved listening to a selected song and ran from one hundred to n times using different Spotify Free bot accounts. At the time, the registration of these accounts was easily automated since Spotify did not require any CAPTCHAs to be solved. As is well known, the original Turing test questioned whether a human could distinguish between a person and a machine. In our case, Spotify’s algorithms were instead implicitly challenged to make this distinction. Our bots interacted with the Spotify system, which tried to decide (via various unknown fraud detection tools) if communication was human or machine based.


One of the major controversies regarding the transition to music streaming services has centered on payouts to artists. With revenue (in USD) steadily increasing into the billions in 2017, music streaming is now the beating heart of the music business, yet many artists struggle with reduced royalties. A heated discussion has focused on the issue of whether or not music streaming will be able to generate a sustainable income for musicians. Statistics on streamed royalty payouts vary and are difficult to find, since agreements and contracts are exclusive and proprietary. In addition, playlists have become increasingly important as a gateway regulating royalty; i.e. an artist who is not featured on a certain playlist gets zero revenue. Estimates, however, usually state that revenue per played track runs as low as $0.005 at Spotify. Within the music industry, this has led to considerable debate. Most famously, Taylor Swift decided to remove her entire back catalog from Spotify in 2014, arguing principally that the ad-supported Spotify Free devalued her music (even if she later, during the summer of 2017 put all her music back again). As we discussed in chapter 1, confidential agreements and record label contracts with Spotify differ, but royalties are usually disbursed to artists (or more precisely, aggregators or record labels) once a song is registered as a play, which generally happens after thirty seconds—hence, the “thirty second royalty rule.” Therefore, our hypothesis was that playing Fru Kost tracks repeatedly for twenty-five seconds would not be a problem, since such plays would not involve any financial transactions and would hence be of little interest to fraud detectors. If our bots would play a song for more than thirty seconds though, we suspected that Spotify would be more cautious about evaluating and securing the “humanness” of the registered users.

To our surprise, however, our SpotiBot engine was able to play Fru Kost tracks repeatedly for both twenty-five and thirty-five seconds. The bot “selenium57,” for example, played Fru Kost’s “Kaffe” 229 times; the “selenium_bot,” as many as 1,141 times. That is, after thirty-five seconds of “Kaffe,” the bots started playing the song again, and again, and again. It is important to note that the experiment did not run completely smoothly, and many of our bots did not “listen” to the exact amount of repeated plays they had been programmed to perform.16 Yet the intervention did demonstrate that it was possible to automatically and repeatedly play tracks on Spotify via bots. The experiment further showed that no major difference existed between our bots playing an artist such as Fru Kost for twenty-five or thirty-five seconds. In Spotify’s system, it seemed as though both listening sessions were simply treated as plays, since we could closely monitor our own royalty payments. While the experiment only generated a meager $6.28, each play was registered and archived properly via the aggregator Record Union. In other words, Spotify did not seem to care whether a computer or a human was listening to its music archive. Or at least, it appeared as though the company had not taken any serious steps to make sure that such a distinction could be made. Our Fru Kost experiment raises several ethical issues; we were, after all, contravening some of Spotify’s user agreements. However, we never cashed the royalty check generated by our experiment. After the intervention, during the summer of 2016, Spotify also started using CAPTCHAs as well as reCAPTCHAs with image identification, allegedly to better protect its system. Repeating the same experiment after that change would have been considerably more difficult.

As a last comment, our record label experiments and interventions have also given us an awareness of the migratory lives of streamed music files—migrations into digital spheres that go beyond the reach of aggregator analytics tools. As it turns out, our sounds have rarely stayed in place and, instead, have taken on a kind of second life on the web. About a year after our first releases, some of our music had completely disappeared from the services where we first put them, simply because we forgot to renew our aggregator contracts. This is a problem that we suspect many artists deal with, and it illustrates how subscription logic for streaming platforms affects not only music consumption but also artistry.


Our album The Silence of Scholarly Life by the Ethnologist, for example, never really succeeded in attracting listeners, yet it had an interesting afterlife. During the year it was available on Spotify (2015), songs from the album were played a meager sixty times in total. Most of the streams occurred at the time of the album launch (and most likely because we played it ourselves). When the album was removed from Spotify due to an unrenewed subscription, however, it did not disappear completely. Instead, tracks from the album are still available for download on unauthorized websites. The site mp3red.me, for instance, at one point rated the track “The Dishes” as the best song on the album, claiming that “19 people think this track is stunning.” And on the site Musixmatch, Portuguese lyrics were added to the nonvocal track “Lunch”—lyrics that upon closer inspection appear to originate from the song “Ela é tipo” by the Brazilian band Banda Swell. These types of unexpected appropriations reveal how the social lives of streamed music files stretch way beyond commercial platforms and are open to all kinds of reuses, rewirings, and recontextualizations. Through their entanglements with algorithms, global music metadata arrangements, and automatized web-scrapers, streamed music files get wrapped up in the whirlpools of data traffic that surround digital streams. The movements and appropriations of our sounds thereby testify to the fuzzy and porous boundaries of contemporary music streaming services.






Maria Eriksson

Maria Eriksson is a co-author of "Spotify Teardown: Inside the Black Box of Streaming Music" (MIT Press, 2019)

MORE FROM Maria Eriksson

Rasmus Fleischer

Rasmus Fleischer is a co-author of "Spotify Teardown: Inside the Black Box of Streaming Music" (MIT Press, 2019)

MORE FROM Rasmus Fleischer

Anna Johansson

Anna Johansson is a co-author of "Spotify Teardown: Inside the Black Box of Streaming Music" (MIT Press, 2019)

MORE FROM Anna Johansson

Pelle Snickars

Pelle Snickars is a co-author of "Spotify Teardown: Inside the Black Box of Streaming Music" (MIT Press, 2019)

MORE FROM Pelle Snickars

Patrick Vonderau

Patrick Vonderau is a co-author of "Spotify Teardown: Inside the Black Box of Streaming Music" (MIT Press, 2019)

MORE FROM Patrick Vonderau

Fearless journalism
in your inbox every day

Sign up for our free newsletter

• • •