What was on the other end

I can still remember sitting on the floor of my parents' basement in front of a wood-grain Zenith CRT, when Nirvana's "Come As You Are" came on MTV. Watching it now, it practically drips early 90s. The music video literally features an underwater dream sequence with Kurt Cobain floating in slow motion that crossfades to a baby swimming in a pool. The song had a descending bass line, a chord change in the verse, and a watery effect on the guitar that I wouldn't know to call a chorus pedal for another twenty years. At some point I noticed I was cold. I looked down and saw the hair on my arms standing up. The song had done that.

Music has always worked this way. A song is one person drawing from their own lived experiences and rendering them in a way that names something the listener couldn't have named themselves. Doing that in public is a vulnerable act, and when the work is honest, the listener responds in ways they didn't choose. What happened to me on the floor of that basement has a name: frisson. It is the involuntary chills and goosebumps a song can produce, the body responding before the mind has decided what just happened. Frisson is the body recognizing that someone is alive on the other end of the wire.

Three rooms

The thing that exists on the other end of any song you respond to is a person who decided to put something specific into it. The years they spent living and learning are in the choices that produced the song, which is frankly the only reason it carries any weight. Generative AI puts a probability distribution where that person used to be.

In 1986, a young Adam Yauch was at Chung King Studios in New York, searching for a beat for a song Run-DMC had just freestyled the intro to. Yauch suggested flipping a reel-to-reel tape deck upside down, pressing record while their Roland TR-808 drum machine ran forwards, then flipping the tape back. The net effect was a backwards drum pattern. Mike D writes in the Beastie Boys Book about the moment they first played it back: "It was fucking magical. Up until that moment, there had never been a beat that sounded as funky fresh, dope, hype, or def." That backwards 808 became the foundation of "Paul Revere."

Fast forward a decade from New York to California. In the basement of Rare Records in Sacramento, DJ Shadow built Endtroducing..... on an Akai MPC60 II sampler and a Technics SL-1200 turntable. He had spent two years working upstairs at the store before the owner gave him keys to what he later called "a cavernous catacomb of disorganized and otherwise jettisoned vinyl." In Doug Pray's documentary Scratch, Shadow sits in those stacks and talks about pulling out specific records and feeling like he was meant to find them. AI can match a million samples by similarity in seconds, but it will never be able to discern the ones that matter most. Endtroducing was the culmination of a musician spending years sifting through discarded records and encoding them tastefully into his debut album.

Fast forward a decade after that, further down the California coast. J Dilla made Donuts on a Boss SP-303 sampler and a 45 rpm record player in his room at Cedars-Sinai Medical Center. He was being treated for complications from lupus. Stones Throw released the album on his 32nd birthday. He died three days later. Most of the beats on Donuts are under two minutes, and they were assembled between dialysis sessions on a sampler that could fit on his hospital tray. Dilla knew he was running out of time and used what he had left to finish his last album. A generative model can try and imitate his swing, albeit poorly. What it will never be able to do is assemble an emotionally charged album from inside a failing body.

What the imitation removes

AI systems can be trained on the texture of any of those records, and the same surfaces can be reproduced cheaply by people who have never been in any room remotely like these three. What these models cannot reproduce are the things those textures were derived from. Yauch's tape flip was a specific person's specific leap. Shadow's Endtroducing was the culmination of two years of one person searching through stacks. Dilla's swing was a man making his final album from a hospital bed, knowing it would be the last. The textures are the form those decisions took on the way to someone's ears. AI generates the form without the decisions, and the body responds to a wire that has no sender. The signal is a forgery, and no one is on the other end to answer for it. The response is real, and the listener is being moved by an absence.

Jack Antonoff posted to Instagram earlier this month that making music has become an ancient ritual, and that nothing is "more embarrassing than considering there is a way to optimize that process." Antonoff is right to call it embarrassing. The whole industry should be. Generative music exploits the body's response without participating in the system that produced it. It is an invasion at a psychological level, perpetuated by an industry whose listeners are five times as likely to feel concerned about AI as excited about it, and that has not stopped to think about what it is selling.

Who used to be on the other end

The reason Nirvana reached my parents' basement was that a Seattle radio station called KCMU, a listener-funded station run by volunteer DJs, had put them on the air four years earlier. In 1988, Kurt Cobain walked a copy of "Love Buzz" into the station and they did not play it. He had to call from a payphone and request his own song to get it into rotation. In 1994, Rodney Bingenheimer played an offbeat song called "Undone (The Sweater Song)" on Rodney on the ROQ, his Sunday night specialty show on KROQ, before any other station would touch it. After Rodney championed it, KROQ added it to regular rotation, MTV Buzz Bin followed, and Weezer's career was born. KCMU rebranded as KEXP in 2001. Both stations are still on the air.

The same kind of taste moved online with blogs. Stereogum, Pitchfork, Gorilla vs Bear, and a hundred smaller sites ran daily posts, with the Hype Machine aggregating everything into one feed. Between them they broke bands like Fleet Foxes, The xx, and Tame Impala. Labels eventually realized the blogs were doing their marketing for them and started feeding them MP3s with individual access codes so they could track who leaked first. Every link in the chain was a person with a name and a point of view.

Even the platforms did this work for a long time. Apple ran iTunes Single of the Week from 2004 to 2015. Every Tuesday the iTunes editorial team picked one song and gave it away free for the week, with a name attached to every choice. I would come home from school, get off the bus, and use my parents' dial-up internet to download whatever they had picked. Those songs are all still in my Apple Music library two decades later. Apple quietly killed the program in January 2015 in the run-up to Apple Music's launch, but the editorial team survived the transition and still curates much of what plays on Apple Music today.

A problem Spotify built

While Apple Music kept its editors, it isn't inherently anti-AI, nor should it be. Apple's AI crossfade feature, shipped in 2024 as AutoMix, applies AI to song transitions rather than song selection. Spotify has largely taken the opposite approach, gaining streaming market share by dismantling most of the chain in the process. This is the standard pattern Cory Doctorow calls enshittification, and Spotify is one of his canonical examples. When AI-generated music began flooding the platform in 2025, listeners said they did not want it, but Spotify had already spent a decade optimizing for engagement rather than authorship. The bots arrived through a door the platform had been intentionally holding open for years.

Spotify's response in April 2026 was the Verified by Spotify badge, a small green checkmark given to artists with "consistent listener activity and engagement over time," "good standing with Spotify's platform policies," and "signals of a real artist" on the profile (concert dates, merch listings, linked social accounts). Every one of those criteria is itself an algorithmic signal. The company that intentionally removed human judgment and taste from curation is paying contractors to verify which artists are real. The badge lives on the artist profile page, a screen most listeners never open while listening. Spotify compared the program to nutrition facts. Nutrition facts arrived after the food got worse, not before. The badge is more AI deployed to fix a problem prior-generation AI created. The recommendation engine that broke discovery in the first place is still running underneath.

Counter-currents

I just finished Mixtape, the new game from Beethoven & Dinosaur and Annapurna Interactive. It follows three teenagers in 1990s Northern California on the last night before they leave for college. The protagonist's stated dream is to be a music supervisor. The game was built by making a literal mixtape of Devo, Joy Division, Iggy Pop, and The Smashing Pumpkins, then letting the crescendos and lulls dictate the story's shape. It is the inverse of an AI-generated playlist. Director Johnny Galvatron picked his personal greatest hits, and the songs shaped the story.

The other week I walked past a billboard near KEXP's Seattle Center studios. The station that put Nirvana on the air almost four decades ago is now buying outdoor advertising to say some things are better without AI. KEXP was reporting on AI music siphoning streams from real artists in November 2024, almost two years before any formal Spotify response. The chain of human curators that decided what would reach a basement television in 1992 is, at least on one street corner in Seattle, making its argument out loud.

A smaller chain

The version of the chain I rebuilt for myself is small. Clickwheel is a personal tool I wrote that takes my music library and pushes it wherever I decide: my iPod, Apple Music, or Plexamp on my phone. Every play, no matter the destination, flows back into Rewind, my personal data layer.

Clickwheel runs as a CLI and is reachable through an MCP server, so I can drive any of its commands directly inside a Claude conversation. Claude makes the curation work easier, but the taste stays mine. "Pick a road trip mix that fits under 4GB and leans toward 90s alt rock" is a sentence I can type now, and a question that used to take me an afternoon of dragging files into Apple Music. Rewind has every song I've listened to going back to college. Claude lets me read that history like a journal, surfacing patterns I would never piece together on my own. None of this is grand. It is one person trying to keep a small piece of the chain intact, because the chain is part of what made the music meaningful in the first place.

What stays

"Come As You Are" came out in January 1992. The wood-grain Zenith in my parents' basement is long gone, but I still get goosebumps running up my arms when a bass line sneaks up on me. That happens because an artist somewhere dreamed up that chord change, and because someone felt it was worth their effort to help it reach me. None of that is replaceable.