Research on human music discovery versus algorithms does not establish a universal winner. Personalised recommendations can help people find music that fits their preferences. Human curation can introduce unfamiliar artists and provide a reason to explore them. Both can also concentrate attention on a limited selection of music.
The result depends on the recommendation method, the listener and the definition of success. Finding another song you immediately enjoy is a different achievement from introducing you to a musical tradition you have never explored. A system can perform well at one and less well at the other.
To understand the evidence, we need to distinguish preference matching, discovery, diversity and social influence—and be careful about what each study actually tested.
This article is published by Kallitone, a platform built around listener comparisons. It examines research that supports and challenges common claims about human and algorithmic discovery. The cited studies do not test Kallitone.
Human recommendations and algorithms take several forms
A friend sharing a song, a radio presenter introducing a local band and a community voting on recordings all involve human judgement. They perform different jobs. A friend can use personal knowledge of you; a presenter can explain a scene; a voting community can gather preferences from many listeners.
Algorithmic recommendation also covers different approaches. Software can use patterns in listening behaviour, relationships between recordings or information about the music to select and order suggestions. The choice of what to optimise matters: immediate listening, variety and repeat engagement are distinct targets.
These methods can overlap. An editor can choose a pool of songs and software can personalise their order. Listeners can make individual judgements while software calculates a ranking. A useful comparison identifies who supplies the judgement, who selects the candidates and how the result reaches a listener.
What counts as successful music discovery?
Before comparing methods, it helps to separate five questions. These are practical distinctions for reading the research, rather than a single standard scoring system.
| Goal | Useful question | What it can miss |
|---|---|---|
| Preference matching | Did the listener enjoy the recommendation? | Enjoyment may come from something very familiar. |
| Novelty | Was the track or artist new to this listener? | A new track can sound much like everything already in their collection. |
| Diversity | How varied was the music encountered or played? | Variety does not establish that the listener liked it. |
| Lasting interest | Did the listener voluntarily return to the music? | Repeated playback is an imperfect measure of emotional attachment. |
| Artist exposure | Which artists received attention, and how was it distributed? | Total listening can grow while attention remains concentrated. |
For an independent artist, a thousand brief encounters and a smaller group of returning listeners mean different things. For a listener, an uninterrupted afternoon of pleasant background music can be valuable even if it produces no new favourite. A discovery claim should say which outcome improved.
What a direct comparison of human and algorithmic recommendations found
A study by François Moreau and colleagues, published online in Applied Economics Letters in April 2026, analysed listening records from 9,778 premium subscribers to a major European streaming platform. The dataset included 4,136 distinct songs discovered through algorithmic recommendations or human-curated playlists.
The researchers used subsequent organic streams—listening initiated outside those recommendation routes—as an indirect measure of how well a discovery matched someone’s preferences. They found evidence that algorithmic recommendations helped listeners find suitable music, while human recommendations performed better when introducing songs by artists unfamiliar to the listener. [1]
This is a useful distinction between discovering a song and discovering an artist. It also has limits: the study used behaviour on one platform, and repeated listening was a proxy for preference matching. Its human recommendations were curated playlists, so the finding cannot automatically be extended to a friend, a critic or an anonymous voting community.
Do recommendation algorithms create musical filter bubbles?
Some findings support concern about narrow listening. In a 2020 Spotify study, Ashton Anderson and colleagues found that algorithmically driven listening was associated with lower consumption diversity. A separate randomised component found that ranking songs by relevance encouraged short-term listening, particularly among listeners whose consumption was already less diverse. [2]
Those results answer different questions. The observational association does not establish that algorithms alone caused narrow tastes. The experiment tested how listeners responded to particular ranking methods; it did not prove that every recommendation system progressively narrows everyone’s musical world.
Quentin Villermet and colleagues reached a similarly qualified picture in Follow the Guides, published in 2021. Analysing a year of listening by roughly 9,000 users, they distinguished organic listening, algorithmic guidance and editorial guidance. Relationships with repetition and artist popularity differed across user groups and recommendation modes. [3]
The practical lesson is to ask whose listening changed, which recommendation route they used and how diversity was measured. “The algorithm creates a bubble” leaves all three unspecified.
A platform can become more diverse while individual listening becomes narrower
Individual and collective diversity can move in opposite directions. David Holtz and colleagues demonstrated this in a randomised Spotify experiment reported in 2020. Personalised recommendations increased podcast streams by 28.9%, reduced average individual-level diversity by 11.5% and increased aggregate diversity by about 6%, compared with recommendations of popular podcasts within demographic groups. [4]
This experiment concerned podcasts, not music. It is relevant here because it illustrates a measurement problem: different people can concentrate on different niches, increasing diversity across the audience while narrowing each person’s consumption. Its numerical results should not be presented as effects on music listening.
When a platform reports that recommendations increased diversity, ask whether that means more varied listening for each person, a wider selection across all listeners or broader exposure for creators.
Recommendations can also be designed to encourage exploration
Personalisation does not have to mean repeatedly selecting the closest match to existing habits. Researchers can explicitly build variety into the material people encounter.
Lorenzo Porcaro, Emilia Gómez and Carlos Castillo studied recommendations of electronic music over 12 weeks. Their study began with 110 participants from Southern Europe; the reported analysis retained 94 after participation exclusions. People receiving more diverse recommendations showed greater willingness to explore the offered playlists. [5]
The distinction between findings matters here too. Openness to electronic music increased during the study, but the difference in openness attributable to recommendation diversity was not statistically significant in the author version examined. Participants also listened under structured study conditions.
The study provides evidence that deliberate diversification can encourage exploration in a particular setting. It does not show that every automated playlist expands musical taste, or that a more varied list always produces a lasting change.
Popularity bias can disadvantage listeners with less mainstream tastes
In a reproducibility study presented at ECIR 2020, Dominik Kowald, Markus Schedl and Elisabeth Lex evaluated recommendation algorithms using Last.fm data. The tested models favoured popular artists, and the group with the least mainstream listening preferences received the least accurate recommendations. [6]
This shows why an overall accuracy score can conceal unequal performance. A system that works well for common preferences may serve a smaller musical community less effectively.
It was an evaluation of particular algorithms and a particular dataset, rather than an audit of every commercial streaming service. Popularity bias is a risk to measure in a system; its size and consequences should not be assumed to be identical everywhere.
There is a separate question for artists: whether a discovery method gives less established recordings meaningful opportunities to be heard. Listener satisfaction and the distribution of exposure deserve their own measurements.
Human choices are also influenced by other people’s choices
Replacing automated recommendations with people does not remove social influence. In a 2006 Science experiment, Matthew Salganik, Peter Dodds and Duncan Watts created an artificial music market involving 14,341 participants. Listeners encountered previously unknown songs with or without information about previous participants’ choices.
Stronger social influence made success more unequal and less predictable across the experimental groups. Musical appeal still mattered, but visible prior choices helped shape what became successful. [7]
This was a study of social information in a music market, not a direct test of professional editors against recommendation software. It demonstrates a problem that can exist wherever people see what others favour: popularity can influence the next decision as well as record earlier decisions.
A curator’s explanation may give you a valuable reason to listen. A visible popularity score supplies a different kind of information. Treating both as “human discovery” hides that distinction.
What does the research say about blind music evaluation?
Blind presentation changes the information available during a judgement. Withholding an artist’s name or public profile can keep those cues out of the interface. It leaves the listener’s taste, familiarity with the recording and response to the sound in place.
Research on visual information shows why strong claims need care. In a 2013 PNAS paper, Chia-Jung Tsay found that participants could identify music competition winners from short silent videos more successfully than from audio in the tested conditions. [8]
Samuel Mehr, Daniel Scannell and Ellen Winner replicated the original effect in 2018 but found important limits. Changing the performance clips changed the result. When performance-quality differences were clearer, hearing the music produced more accurate judgements. [9]
Together, these studies support examining the role of presentation while resisting the claim that appearance always outweighs sound. Predicting the winner of an elite performance competition is also different from choosing a favourite independent song. Neither study validates blind voting as an objective measurement of artistic worth.
Where Kallitone fits into this evidence
Kallitone uses human preferences as inputs to a discovery and ranking process. In Explore, listeners compare two recordings without starting from the artists’ public profiles. The standard requirement is at least 15 seconds from each track, with passages selected randomly from within the songs. Votes contribute to Elo-style ratings. [10]
This is a specific design choice: give a recording an initial hearing through a comparison before its public identity becomes the focus. Software still handles matching and rating calculations. “Human-powered” describes the source of the musical judgements.
The research above helps explain why that design is worth testing. It does not establish that Kallitone produces more diverse listening, better recommendations or fairer exposure than another service. Those outcomes require platform-specific evidence.
A useful evaluation would compare otherwise similar listening conditions with and without identity or popularity cues, then measure what people voluntarily return to later. It would also examine which artists receive exposure and how results vary across musical communities. Those are proposed measures, not results this article claims Kallitone has already achieved.
What listeners and independent artists can take from the research
For listeners, there is a practical reason to keep several routes into music open. Personalised recommendations can help when you want something close to an existing preference. A trusted presenter, editor or friend can offer an unfamiliar starting point and explain why it interests them. A blind comparison lets you notice your response before looking at the artist’s profile.
Try judging a discovery source by what happens after the first encounter. Which artists did you return to? Did you hear something outside your usual choices? Did an initially unfamiliar recording become a favourite? This is an informal listening habit, rather than a controlled experiment, but it gives you a more useful personal measure than how many tracks passed through a playlist.
For artists, discovery involves several steps: being presented, receiving attention and earning a return listen. A placement or favourable vote can create an opportunity; sustained interest is a further outcome. Look for evidence that a route reaches people who continue to engage with your work.
The strongest question to ask of any discovery service is concrete: what does it help people discover, how is that measured, and who benefits from the result?
Research sources
- Moreau, F., et al. (2026). Consumer satisfaction with new content discovery through algorithmic and human recommendations. Applied Economics Letters, published online 13 April 2026. DOI: 10.1080/13504851.2026.2655917.
- Anderson, A., Maystre, L., Mehrotra, R., Anderson, I. and Lalmas, M. (2020). Algorithmic Effects on the Diversity of Consumption on Spotify. The Web Conference 2020. DOI: 10.1145/3366423.3380281.
- Villermet, Q., Poiroux, J., Moussallam, M., Louail, T. and Roth, C. (2021). Follow the guides: disentangling human and algorithmic curation in online music consumption. ACM Conference on Recommender Systems. DOI: 10.1145/3460231.3474269. Author version linked.
- Holtz, D., et al. (2020). The Engagement-Diversity Connection: Evidence from a Field Experiment on Spotify. Author paper; see also the Spotify Research publication summary. The experiment concerns podcast recommendations.
- Porcaro, L., Gómez, E. and Castillo, C. Assessing the Impact of Music Recommendation Diversity on Listeners: A Longitudinal Study. We consulted the 2022 author preprint, arXiv:2212.00592v1. Subsequently published in ACM Transactions on Recommender Systems, 2(1), 2024.
- Kowald, D., Schedl, M. and Lex, E. (2020). The Unfairness of Popularity Bias in Music Recommendation: A Reproducibility Study. ECIR 2020. DOI: 10.1007/978-3-030-45442-5_5. Author preprint linked.
- Salganik, M. J., Dodds, P. S. and Watts, D. J. (2006). Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market. Science, 311(5762), 854–856. DOI: 10.1126/science.1121066.
- Tsay, C.-J. (2013). Sight over sound in the judgment of music performance. PNAS, 110(36), 14580–14585. DOI: 10.1073/pnas.1221454110.
- Mehr, S. A., Scannell, D. A. and Winner, E. (2018). Sight-over-sound judgments of music performances are replicable effects with limited interpretability. PLOS ONE, 13(9), e0202075. DOI: 10.1371/journal.pone.0202075.
- Kallitone About and Create — product descriptions, separate from the academic evidence above.
Research sources checked on 12–13 October 2026; Kallitone product information checked on 12 October 2026. This is a focused editorial review of selected primary studies, not a systematic review or meta-analysis. Several studies use streaming-platform data or involve platform researchers; their findings describe the methods, audiences and periods studied. Author versions and their publication status are identified in the source list.