Over the past few months, a number of label teams we work with been asking us the same question:
Do Spotify’s AI Playlists change how algorithmic discovery actually works?
With Spotify rolling out its new text-to-playlist feature — allowing users to generate playlists through natural language prompts — it’s a fair question. On the surface, the experience feels radically different: instead of passively receiving recommendations, users can now talk to the system.
So we decided to test it properly.
Putting Spotify’s AI Playlists to the Test
To remove as many variables as possible, we created a fresh Spotify account — no listening history, no follows, no signals, no prior taste profile. Then we spent time interacting with Spotify’s AI Playlist feature. Our goal was simple: build a “perfect Christmas party playlist” — all while gradually increasing prompt complexity to explore what the system is actually capable of.
“Build me a perfect playlist for an upcoming Christmas office party” works great: a clean, familiar mix of chart-topping Christmas hits, with all the usual suspects present.
But what if we don’t want to listen to the same old classics? As soon as we add a bit more nuance, first cracks start to show:
“Build a Christmas playlist that avoids the usual classics but still feels unmistakably festive” kind of — mostly — works. The Christmas vibes are there, but the system often reaches for B-sides, covers, or alternate versions of the same canonical tracks — technically avoiding “the classics,” while still leaning on the same musical references.
Okay, but what if we try to generate a more diverse, discovery-driven Christmas playlist that will showcase emerging artists across genres? This is where the system really starts to wobble.
Diversity is clearly one of the “levers” the AI agent can access, but balancing it with intent accuracy proves to be tricky. Ask the system to showcase diverse genres and cultures, and it tends to forget it’s supposed to curate a Christmas playlist. Reinforce the Christmas constraint, and it does something slightly ironic — producing a playlist auto-titled “Diverse Global Christmas” that features Dean Martin alongside Elvis Presley, Mariah Carey, and Michael Bublé.
After some trial and error, the most stable version of our “diverse” prompt had to be pretty simple:
This performs decently — but only once most additional constraints are stripped away. Even then, results skew toward higher-volume tracks, and adding further diversity conditions (languages, cultures, scenes) quickly causes the playlist to fall apart.
Finally, when we start using creating prompts in a more natural way — relying on cultural references or abstraction — the system often misses the mark entirely. For instance, asking for “Christmas music that could play in an A24 film” leads the system to latch onto surface keywords (“Christmas” and “film”), producing a generic playlist of Christmas movie music, while completely missing the reference to A24 as a specific aesthetic.
What this tells us about AI Playlists
The takeaway isn’t that the feature is bad — far from it. It’s that AI Playlists don’t introduce a new kind of musical intelligence.
What they do is translate natural language into the same dimensions Spotify’s recommender systems have relied on for years: genres, moods, artists, scenes, eras, popularity groups, and contextual tags. When prompts and user intent fit that vocabulary, results feel smart — as per Spotify’s own support page, “[best AI] playlists are generated with prompts that contain a combination of genres, moods, artists, or decades”
So no — AI playlists don’t change the underlying logic of recommendation and discoverability. But they do add another layer to the algorithmic consumption landscape, making recommender systems more influential in how users discover and consume music. They turn the recommender into something users actively talk to — which means the system’s internal representations (and how your music relates to them) matter even more.
Why this matters for artists, labels and institutions
In practice, this raises a familiar set of questions for artists, labels, and other recording stakeholders:
- How does the recommender system currently see my catalog?
- How much — and to whom — does it recommend my music?
- Which genres, moods, scenes, eras, and cultural “anchors” does it associate with my tracks?
- And what can I actually do to adapt my marketing strategies and investment decisions in an algorithm-mediated streaming economy?
Finding answers to these questions — and sharing them with the industry — has been at the core of what we do at Music Tomorrow from the start.
To make this more concrete, we’re currently offering a free artist analysis to anyone who creates a profile on Music Tomorrow. It’s a simple way to get a clearer picture of how streaming recommender systems see your music — and to identify opportunities to refine positioning, targeting, and overall release & investment strategies for your artists.
