A Song You Didn’t Ask For

On Thursday, Sept. 10, a WHRB 95.3 FM DJ played a song that fellow radio member Shahmeer Baqir ’29 had never heard before. By the next day, he had listened to it twenty times.

For most modern listeners, finding a new song means opening Spotify and letting the platform’s algorithm do the work. Tools like Discover Weekly, Release Radar, AI DJ, personalized mixes, and a web of tailored recommendations promise to predict what we might like before we press play. This past July, Spotify introduced a beta feature that lets users literally talk to the app about what they want to hear, from requesting unfamiliar artists to narrowing playlists by genre or mood.

Over time, Spotify has become remarkably good at knowing what its listeners like. The company uses billions of behavioral signals, including listening history, to create an increasingly detailed picture of each user’s taste. In March, Spotify introduced Taste Profile, which lets users see and directly shape the profile the company has built around their listening habits.

Yet, in an age of convenience, a more organic way to discover music can be found on Harvard’s campus.

Baqir is a member of WHRB’s Record Hospital, which plays genres like underground punk and indie rock, and co-directs its Jazz department. For his radio shows, he finds music in several places: Instagram, friends’ recommendations, and occasionally a self-directed rabbit hole.

“I’ll basically just pick a country off the top of my head and do a bit of a deep dive into what’s happening in the scene there,” he said.

Algorithms can still play a role. After he has built half a set, Baqir occasionally puts it into Spotify and looks at the platform’s recommendations. Still, he does not consider the algorithm the foundation of his radio programming.

“I think there’s just a lot more information and a lot more intentionality built into it,” he said of human DJs.

That distinction is increasingly relevant as Spotify moves beyond simply recommending existing music. At its May 2026 Investor Day, the company described its future as a shift from “curation and recommendation into an era of generation,” driven by what it calls a Large Taste Model trained on the aforementioned behavioral signals.

The appeal is immediately apparent. There is too much music in the world for anyone to sort through alone. Recommendation systems on streaming platforms can sift through enormous catalogs and narrow them down to a manageable number of possibilities. Since its launch, Spotify says that Discover Weekly has generated more than 100 billion streams, while Release Radar reaches roughly nine million weekly listeners.

However, narrowing the scope of possibilities also narrows the potential for discovery. A recommendation system begins with what it knows about you, while a human DJ can begin somewhere else entirely.

Liya Naod ’29, WHRB’s live show director and a member of Record Hospital, finds much of her music through Spotify and Instagram. The latter has become particularly useful for following Boston’s underground music scene, where flyers and show announcements circulate online before bringing listeners together in person.

For Naod, just seeing the same band’s name repeatedly can be enough to create curiosity. “It gives a band a lot of name recognition when you see the same band’s name on a lot of different flyers,” she said.

Naod also sees a difference in what happens when a DJ begins assembling music.

“I think a human DJ can recognize a lot of cohesion in ways that algorithms can’t necessarily create,” she said. A set can be built around an obscure subgenre or a connection between songs that exists solely because the DJ notices it, a judgment that is hard to reduce to listening data.

For Marwa Sahak ’29, another member of Record Hospital, this is a real challenge. Sahak hosts “Mideast Emo,” a show focused on rock and metal music from the Middle East, a niche requiring her to search beyond obvious recommendations.

“To find rock music from the Middle East has been really hard,” Sahak said. She searches social media and asks friends from different countries for recommendations.

“I would prefer to get recommendations from friends,” she said.

Her approach points to a limitation that researchers have identified in recommendation algorithms. Personalization can improve recommendation accuracy while making music diversity harder to maintain, creating a potential “filter bubble” in music exploration.

Spotify acknowledges that recommendations cannot be built solely on popularity or established habits. The app’s human editors curate playlists such as New Music Friday and Fresh Finds, with more than 30 editors working on Fresh Finds alone.

Record Hospital takes this philosophy of resisting popularity seriously by deliberately expanding their discovery efforts. According to Baqir, “there’s a listener cap on the music that we can play on air, so we don’t play any artists with above 100,000 listeners.” The restriction forces DJs to search for music not yet validated by mass popularity.

Listeners can create the same kind of constraint for themselves. Instead of asking an algorithm to narrow thousands of songs down to something they like, they can deliberately start somewhere unfamiliar and see where they land.

The appeal of those experiences is not that they are more convenient—they are less efficient —but that’s exactly why they can feel different. Spotify can still be an amazing tool for discovery, and the point is not to stop using the app; it is to stop treating personalization as the only way to find something new.

Pick a country and search for its music. Follow a local venue on Instagram. Ask a friend what they have been listening to recently. Walk into a record store without a plan. Turn on a radio station and stay for a song you would not have chosen yourself. The best recommendation might be the one you could not have predicted.

Rohan Tyagi ’29 (rohantyagi@college.harvard.edu) stumbled across Tuareg Rock music while writing this piece—he is a fan.