Spotify just expanded its Prompted Playlists feature to include podcasts, and I’ll be honest, this is one of those rare AI features that actually solves a real problem instead of creating new ones.
The feature was already available for music since December, but podcasts are a different beast entirely. Music discovery is relatively low-stakes. You skip a bad song recommendation and move on with your life. Podcast discovery, though? That’s a commitment problem. Episodes run 30 minutes to 3 hours. You can’t just sample 15 seconds and decide. The barrier to entry is high, and most podcast apps handle discovery terribly.
The Algorithm That Waits (And Verifies Humor?)
When I tested this thing, I asked for a playlist of Dungeons & Dragons actual play shows with a mix of popular series and hidden gems. Spotify’s AI sat there “verifying humor” for several minutes. I have no idea what that means from a technical standpoint, but I appreciate the honesty about whatever computational gymnastics were happening behind the scenes.
The first result had a critical flaw that anyone working with sequential content would immediately spot. It pulled random episodes from various shows. If you’ve ever tried jumping into episode 47 of a D&D campaign, you know this is useless. The whole appeal of actual play podcasts is following ongoing narratives. Dropping someone into the middle defeats the purpose.
But here’s where the feature redeemed itself. I edited my prompt to specify “only first episodes of campaigns or shows,” and the second generation was legitimately good. It surfaced Dimension 20 and Critical Role (obvious picks), but also smaller shows like Tales from the Stinky Dragon and Join the Party that I hadn’t encountered yet. That’s the sweet spot for recommendation engines: validating your existing taste while introducing genuinely new content at the margins.
The Summaries Don’t Suck
Each episode came with an AI-generated explanation for why it fit my prompt. I’m usually skeptical of machine learning summaries because they tend to be confidently wrong in subtle ways. These were vague but accurate. Describing Dimension 20: Fantasy High as “funny with real stakes” is about as generic as you can get, but it’s not misleading. For discovery purposes, that’s acceptable.
The technical challenge here is actually harder than it looks. Summarizing a 2-hour unscripted podcast episode requires understanding tone, pacing, and context that doesn’t exist in a transcript. Either Spotify has access to better podcast metadata than I assumed, or their model is doing some interesting analysis on audio features and conversational patterns.
What interests me more is the refresh functionality. You can set these playlists to regenerate daily or weekly, which transforms the feature from a one-time recommendation tool into an ongoing discovery mechanism. That’s a subtle but important shift. Most recommendation systems are reactive. You finish something, they suggest the next thing. This is proactive, almost like having a friend who drops podcast recs in your inbox regularly.
The Podcast Discovery Problem Nobody Solved
The podcast ecosystem has a discovery problem that streaming music never really had. Music platforms could lean on decades of metadata, genre classifications, and collaborative filtering. Podcasts don’t have that infrastructure. Genre tags are inconsistent across platforms. There’s no equivalent to the music genome project. And podcasts are fundamentally more diverse in format, length, and style than songs.
Spotify has been trying to own the podcast space for years now, throwing money at exclusive deals and acquisitions. This feature feels like an attempt to solve discovery through computation rather than curation. Whether that works long-term depends on whether their models can actually understand what makes a podcast good, not just what keywords match your prompt.
The feature is still beta and only works in English for Premium users in select countries. That’s a lot of qualifiers, which suggests Spotify is being cautious about scaling this. Smart move, honestly. Content recommendation is one of those problems that looks solved until you deploy it at scale and realize your algorithm has weird biases or edge cases.
I still don’t plan on using Prompted Playlists for music because I’m perfectly capable of finding new songs on my own. But for podcasts? This might actually change how I discover new shows, assuming it doesn’t degrade into the algorithmic slop that plagues every other recommendation engine once the novelty wears off and the engagement metrics start driving product decisions.