A customer came in last Saturday and told me Spotify’s Discover Weekly had recommended the same Amyl and the Sniffers track three weeks running. Different playlist, same song. She said the algorithm seemed to think she only wanted “loud Australian women” and kept feeding her the same narrow slice. She’d come into the shop because she wanted something actually surprising.

I handed her the new Harmony album on Flightless. Then a Fabulous Diamonds reissue. Then a Japanese psych record by Kikagaku Moyo. She bought all three. None of them would’ve appeared in her algorithmically generated playlists because they had nothing in common except that they were all great and they were all in the shop.

That’s the difference between algorithmic recommendation and human curation. One pattern-matches against your existing behaviour. The other one reads you as a person, takes a punt, and occasionally changes your life.

How the Algorithms Actually Work

Let’s be honest about what music recommendation AI does. It looks at what you’ve listened to, what people with similar listening patterns enjoy, and what’s being promoted by labels and distributors. It crunches numbers. It finds correlations. It serves you options that are statistically likely to keep you listening, because listening time is the metric that matters to a streaming platform.

What it doesn’t do is understand why you liked something. It knows you played a Courtney Barnett album eighteen times last month. It doesn’t know that you loved the lyrical wit but were indifferent to the production style. So it recommends other jangly guitar music when what you actually wanted was sharp writing, which might lead you to hip-hop or folk or spoken word.

The algorithm optimises for engagement, not discovery. Those are fundamentally different goals. Engagement means keeping you comfortable. Discovery means making you slightly uncomfortable in a productive way.

Where Human Curation Wins

A good record store employee does something no algorithm can replicate. They listen to the actual conversation. You mention you’re into Eddy Current Suppression Ring and they ask whether you’ve heard UV Race. You say you’ve been listening to a lot of dub lately and they pull out an Adrian Sherwood compilation you’ve never encountered.

The connections are associative, lateral, sometimes completely unexpected. The best recommendations I’ve ever made have been the ones where the customer looked at the record and said “really?” and then came back a week later saying it was the best thing they’d heard in years.

Algorithms are risk-averse by design. They recommend what’s likely to work. A human recommends what might be brilliant. At Spank Records I’ve always had a wall of staff picks that rotates weekly — handwritten cards explaining why each record matters. That wall sells more per square foot than any other part of the shop.

Where AI Might Actually Be Useful

I’m not a Luddite about this. There are things AI recommendation systems do well, and some of them could benefit independent music retail rather than just streaming platforms.

For example, if you could build a recommendation engine trained on a specific store’s customer base and stock, that would be genuinely interesting. Not Spotify’s model of “people who listened to X also listened to Y” across a hundred million users, but a local model that understands what sells in a particular shop to a particular community. What Chapter Music releases do well with customers who also buy Poison City titles? That kind of cross-pollination within a curated context could be powerful.

I’ve been reading about companies like team400.ai working on AI tools for small businesses, and I think there’s a version of music recommendation that serves independent stores rather than competing with them. Not replacing the person behind the counter, but giving them better information about their own customers’ patterns.

The Real Problem

The bigger issue is that algorithms have trained a generation of listeners to expect music to come to them passively. You don’t search anymore. You press play and the algorithm delivers an infinite stream of vaguely appropriate background music.

That’s fine if music is wallpaper. But if it matters to you, passive consumption isn’t going to cut it. Walk into a shop. Flip through the racks. Ask someone what they’ve been playing lately. Pick up a record with artwork that catches your eye and take a chance on it.

The Tote, the Corner, the Curtin — Melbourne venues where you could walk in not knowing the band and walk out a fan. That’s the live music version of what record stores do. You put yourself in a position where surprises can happen. Algorithms eliminate surprise by design. Record stores and live venues are built on it.

I’ll take a knowledgeable person with strong opinions over the world’s most sophisticated recommendation engine, every single time. The algorithm might be right more often, statistically. But when the human is right, it hits different.