A recommendation system can make your musical world smaller or larger using the same listening history.
If somebody repeatedly plays death metal, the system can respond by finding more death metal. That deepens a known preference.
It can also use the same history to recommend adjacent scenes, older influences, unfamiliar artists, another country’s version of the genre, or something structurally similar that the listener has never searched for.
Both outcomes are personalization.
Spotify describes its current Taste Profile as an interpretation of what a person likes based on what and how they listen. That profile helps shape Home recommendations and other personalized experiences. Spotify even lets users exclude tracks and playlists when a one-off listen would otherwise distort that profile. See Taste Profile and Spotify’s explanation of excluding tracks from it.
That control exists because listening behavior is not a perfect statement of identity.
Sometimes the children’s song is for the child.
Repetition and discovery are both design choices
Recommendation systems often face a tradeoff between exploitation and exploration.
Exploitation means recommending something close to what the system already knows works. Exploration means spending some recommendation space on uncertain material that may broaden the listener’s taste.
Spotify’s discovery products demonstrate both impulses. Discover Weekly uses listening history to personalize recommendations, while features such as Fresh Finds and editorial discovery playlists deliberately introduce less familiar material. In July 2026 Spotify described its weekly discovery playlists as tools for finding new releases, breakout tracks, and music beyond a listener’s existing rotation. See Spotify’s discovery-driven playlists.
So a personalized system is not automatically a musical filter bubble.
It can become one if similarity repeatedly wins over novelty.
Measure variety instead of guessing at it
Listening variety can be measured more carefully than asking whether recommendations “feel repetitive.”
Useful measures might include the number of distinct artists encountered, how often recommendations introduce artists never previously played, genre diversity, geographic diversity, catalog age, repeat rate, and the share of listening devoted to already-familiar music.
Even those metrics require interpretation. A person intentionally exploring one composer’s catalog may want less variety for a month. Another listener may explicitly want constant novelty.
The Algorithmic Reality issue is therefore not that a machine chooses songs.
Radio programmers, record stores, friends, DJs, critics, and record labels have always influenced what people hear.
The new difference is that the selector can continuously learn from the listener and rebuild the record shelf after every session.
Whether that shelf becomes a tunnel or a doorway depends on what the system is optimizing—and whether the listener can push back.