‘Algorithmic fatigue’

The music app I’ve been using for the last 14 years recently decided that I’m obsessed with rainstorms.

When I last listened to my Spotify Release Radar playlist — which for years reliably curated a decent selection of newly released music informed by my favorite artists — the lineup quickly took a turn for the worse. After new songs by familiar indie pop names like Japanese Breakfast and The Marias, there was a five-minute recording of rain falling followed by a short, ambient instrumental by a little-known alternative rock artist. Then, for its remaining 25 tracks, my playlist delivered a succession of rainstorms, nature sounds, and brown-noise frequencies — not exactly the stuff of a killer music festival lineup.

When Spotify launched Release Radar in 2016, it promised “a weekly selection of the newest releases that matter the most to you.” Not only was the feature designed to delight music fans upon first listen, but it would get better over time, the company promised.

So why, nearly a decade later, was my latest Release Radar delivering 85% noise?

Sure, my girlfriend and I had recently started listening to “sleep sounds” — things like ocean waves and the hum of an airplane’s engine — at night, but that didn’t fully explain the issue. The algorithm hadn’t made such a crucial curation error before.

I’m hardly the first Spotify subscriber to notice that the gears of its music recommendation engine have gotten rusty. Over the past year, fans have taken to Reddit, LinkedIn, and other platforms to complain that curated, for you playlists like Release Radar, Discover Weekly, and Daily Mix have gone downhill, resulting in what many describe as an “echo chamber” that feels more repetitive or off the mark than it once did. Random tracks from white-noise playlists or kids’ music albums are popping up where they don’t belong, ruining the listening experience.

Molly Holder, Spotify’s senior director of product for personalization, disagreed that the quality of these curated playlists has declined. “People are discovering more new music and spending more time doing so,” she said in a statement, adding that the company listens to user feedback, continues to enhance personal recommendations, and that listener engagement metrics are up.

But after talking with former Spotify employees, it seems likely that a combination of layoffs and shifting business priorities has hollowed out the platform’s music discovery product. Essentially, in a quest for profitability, Spotify broke its algorithm.

One of the features that set Spotify apart from its intensifying competition over the years was the platform’s man-meets-machine approach to music curation, especially its personalized, algorithmically curated playlists. Discover Weekly promised to introduce listeners to new music that they might like, Daily Mix turned their tastes into themed playlists, and Release Radar highlighted freshly released tracks. For years, these features seemed to elicit near-universal praise online and gave Spotify a competitive advantage just in time to fend off competition from giants like Apple and Amazon. The music discovery features are likely one reason it’s been able to lead the global music subscription market. A recent EMARKETER report found that Americans spend over nine times as much time on Spotify as the next closest competitor. Lately, though, those same features have inspired frustration.

For many, Discover Weekly has gotten worse at delivering songs that align with their musical tastes, sometimes including songs they’ve already listened to on Spotify. Other users report that Release Radar misses new releases from artists they like or mistakenly includes tracks that came out weeks earlier. Release Radar has also started including less relevant music from what one Reddit user described as “random artists with under 50k listens” that seemed to come out of nowhere, while others bemoan the inclusion of what sounds like “AI garbage.”

It just continued until I finally said, ‘I can’t take it.’

The most audible groan of disapproval came with the arrival of Spotify Wrapped last year. While the personalized year-end recap playlist had historically sparked a tsunami of positive online buzz each year, the 2024 edition prompted a far more mixed reaction. Listeners said it lacked personality and interesting insights into their music habits.

Jeffrey Smith, a self-described “Spotify diehard” who leads marketing for the online music marketplace Discogs, has become so frustrated by the platform’s declining music discovery feature that he’s considering switching to Apple Music.

“Over the last couple of years, Spotify has met my needs less and less,” Smith said. “It’s not really reflective of my listening behavior as much as it is reflective of what they want me to listen to. It’s just a listening machine at this point, not a music platform.”

Smith’s affinity for Spotify started fading when he noticed one song — “Back on 74” by Jungle — kept popping up on his personalized Spotify playlists. While it’s possible that he had checked the song out at some point, it wasn’t something he enjoyed or actively listened to. Still, the song made its way from one personalized music recommendation feature to another and even started inspiring Spotify’s algorithms to play similar songs.

“It just continued until I finally said, ‘I can’t take it,'” Smith said.

As a music obsessive working at a major marketplace for vinyl records, Smith has no shortage of sources of quality music recommendations. But the days of Spotify augmenting his organic music discovery appear to be over. If he keeps his Spotify subscription at all, he says it will be for the podcasts and audiobooks the company has added to its premium tier.

I’ve found myself in a similar conundrum. After largely dismissing the 2015 launch of Apple Music as a then-satisfied Spotify subscriber, I recently decided to give Apple’s music service another try. Since its debut, Apple Music has prioritized human editors over algorithms for music curation. Where it does use data-driven personalization, the results are pretty solid, and crucially, free of rainstorms. While Spotify’s new audiobook library makes it tempting to stay, its playlists no longer feel like the best in the business. Once my free trial of Apple Music is over, I’m planning to make the switch.

Music subscription services like Spotify spend an extraordinary amount of money — last year, the company said it spent $10 billion — to license their massive music catalogs from record labels, publishers, and other rightsholders. These high costs make it more challenging to turn an enduring profit, something Spotify only managed to do for the first time in 2024.

From Wall Street’s perspective, Spotify is killing it as year-over-year metrics like subscribers and revenue continue to grow. But in the quest to please investors, some ex-Spotify employees think it has abandoned too much of the human element that makes music and the culture around it so special.

Doug Ford was a Spotify executive who oversaw editorial playlist curation from 2013 to 2018. He arrived via Spotify’s acquisition of Tunigo, a Swedish company that specialized in expert-curated music playlists based on different genres and moods. The next year, Spotify acquired Echo Nest, a music data startup founded by MIT grads that used a mix of machine learning and data filtering to build Pandora-style music recommendation algorithms. Both teams built out Spotify’s man-meets-machine music discovery system, which helped the company successfully weather new competition from tech giants like Apple and Amazon as it started to explore the possibility of going public.

You need to be a successful business in order to offer this utility to people. But they’ve discarded some of the human aspects of it, for sure.

“That was a really beautiful moment in Spotify’s history,” Ford said. “That mix of the intentional algorithmic and human curation to make a really deep product was great.”

The results spoke for themselves. Spotify said the 2015 launch of Discover Weekly yielded 1.7 billion streams in its first five months and generated a weekly wave of social media buzz when listeners’ Discover Weekly playlists were refreshed.

“Our goal with Discover Weekly was to make something that felt like a friend or someone who knew you well was making you a mixtape,” Spotify’s product lead at the time, Matt Ogle, told me in an interview. To accomplish that feeling, the algorithm drew heavily from how real flesh-and-blood music enthusiasts were curating playlists on the platform.

“That’s why that thing was so good,” Ford said. “It was taking all the inputs equally and letting things happen normally with the audience. Now it feels a little bit different.”

Ford said things started changing in the lead-up to Spotify’s 2018 initial public offering and the ensuing pressure to prove to investors that profitability was possible in the notoriously tight-margined music streaming business.

“The culture changed because it had to become a business,” Ford said. “You need to be a successful business in order to offer this utility to people. But they’ve discarded some of the human aspects of it, for sure.”

One way the company shifted was by finding cheaper content to use. As the journalist Liz Pelly outlines in her book “Mood Machine,” Spotify’s pre-IPO years saw an increasing reliance on what is known internally as “perfect fit content,” a euphemism for cheaper-to-license audio and mood-specific stock music that is optimized for longer listening sessions. As more listeners turned to streaming services like Spotify for “chill vibes” playlists and background music for studying, the company realized it could save money on royalty costs by populating those playlists with the cheap stuff. That would certainly explain the astonishing number of rainstorm audio tracks being released in a given week.

Ford watched with dismay as the platform became flooded with generic mood music. Around the same time, he said, respected colleagues started leaving the company. After what he describes as “a particularly bad stretch” of feeling frustrated by what he saw as a dramatic culture shift, Ford accepted an offer from YouTube to lead content for its music subscription service in 2018.

In the years since, Spotify’s music programming strategy has only skewed more heavily toward automation. Pelly writes that by late 2023, the number of Spotify employees working on editorial music curation (about 200 people globally) was less than one-third the size of the team focused on algorithmic curation and personalization, according to an internal org chart. Notably, that head count was before Spotify laid off 17% of its total workforce in December 2023, a drastic cost-cutting measure that CEO Daniel Ek later admitted “did disrupt our day-to-day operations more than we anticipated.” Holder said that there are now over 130 employees working on editorial music curation.

One of the more palpable victims of Spotify’s 2023 cuts was Glenn McDonald, the company’s data alchemist (his real job title), who came from the original Echo Nest team and designed and built much of the original data infrastructure and algorithms powering Spotify’s music recommendation engine. He designed the system that identifies a song’s genre, which was crucial to effectively sort and recommend music and understand listeners’ tastes.

“The genre system was human-guided,” McDonald said. “After they laid me off, they replaced it with a system that is not human-guided. It’s just machine learning. It looks at patterns of words in the titles and descriptions of playlists. That’s objectively worse.”

There’s been a tradeoff in favor of high metrics, long listening sessions, and music that’s just ‘good enough.’

McDonald said the ripple effects of this change could be felt across Spotify’s listening experience. Music was often misclassified, like when Swedish folk-metal bands were suddenly lumped in with all other folk artists. He said it took Spotify engineers a year to fully deactivate and replace the genre system he built, according to data output from the company’s public API. One of the many data-filtering tools he built during his tenure at the company was a set of filters designed to fine-tune the music recommendations presented by Discover Weekly and Release Radar — and crucially, prevent things like nature sounds and ambient noise frequencies from showing up where they don’t belong.

“They would have been flagged as anomalies on those dashboards, and I would have followed up with the teams in charge of those recommendations,” McDonald explained. “Without those dashboards, probably nobody is even watching for these issues.”

Holder said that Spotify’s approach to classifying genres and subgenres has “evolved” from the legacy system built by McDonald and other Echo Nest engineers, but declined to get into technical specifics of how it works now or what advantages the new system introduces.

“We’re always laser-focused on continually enhancing personalized recommendations,” she said. Recent updates include improvements to genre accuracy and Discover Weekly recommendations, she added.

Ford said the changes at Spotify are part of a broader industry shift away from costlier human curation. “There’s been a tradeoff in favor of high metrics, long listening sessions, and music that’s just ‘good enough,'” he said. “It’s happening everywhere.”

The result is what Ford calls a “rampant wave of algorithmic fatigue” among listeners, who are clamoring for more meaningful cultural experiences as online platforms become increasingly hyper-automated.

While the occasional bug and ongoing iteration are to be expected in any product, some longtime subscribers find it odd that Spotify would need to fix or enhance features that worked so well in the first place. My Release Radar delivering mostly white noise and rainstorms, for instance, feels like a fundamental failure of what that feature used to do so well. Over the course of reporting and writing this story over the last few weeks, I’ve checked back with my Release Radar. Each week, it has a few new songs, but it’s still mostly rainstorms.

John Paul Titlow is a freelance journalist who writes about technology, digital culture, travel, and mental health.