In a Stockholm Studio, the Music Industry's Last Human Session

On the desk in front of Lars Ekström, between a cracked coffee mug and a photograph of his daughter at her graduation, sits a score in manuscript paper. Pencil, not software. Thirty-two bars of strings for a commercial that will run during the European Football Championship in June. He has been working on it for three days. The client called yesterday to ask if an AI tool could do it faster.

Ekström is sixty-one. His studio is on the third floor of a converted textile factory in Södermalm, Stockholm's bohemian district, though these days the bohemia is mostly venture capital and wellness startups. The room smells of solder and old wood. There is a Neve mixing console from 1987, a Steinway upright that belonged to his father, and a rack of synthesisers that cost more than a decent car. He has worked here for thirty-three years. He is not sure he will work here much longer.

"I told them the truth," he says. "I said yes, the AI can do it faster. It can do it in four minutes. And then I asked them: do you want it to sound like thirty-two bars that took four minutes, or thirty-two bars that took three days?" He pauses. "They said they'd get back to me."

The Studio That Used to Be Full

Ekström learned his trade in the golden age of Swedish pop production, the late 1980s and 1990s, when Stockholm studios were exporting hits to every continent. He worked on Ace of Base sessions as a junior engineer. He arranged strings for Robyn's early albums. He scored television dramas for SVT, Sweden's public broadcaster. At his peak, in the early 2010s, he employed four full-time assistants and had a six-month waiting list.

The waiting list is gone. So are the assistants. In 2024, his invoiced hours dropped by forty-two per cent. In 2025, they dropped another thirty-one per cent. He now works alone, most days, in a studio built for five.

What changed was not his skill. What changed was the contract. In January 2025, Spotify announced its AI Music Licensing Framework, a set of terms allowing musicians and producers to opt their catalogues into machine-learning training datasets in exchange for upfront payments and fractional royalties on AI-generated works. Universal Music Group signed in March. Sony followed in May. Warner held out until September, then capitulated. By the end of 2025, an estimated seventy-three per cent of commercially available recorded music was available for algorithmic training.

The deals were structured to favour catalogue owners—labels, estates, legacy artists—not working musicians. A producer like Ekström, who owns no master recordings and works on commission, received nothing. But his clients now had an alternative: pay a subscription fee to a generative audio platform like Boomy, Soundraw, or Aiva, describe the mood and instrumentation they wanted, and receive a finished track in minutes.

Ekström pulls up an email from November 2025. The sender is a creative director at a Stockholm advertising agency he has worked with for twelve years. The subject line reads: "Budget pressures—need to talk." The body of the email is two sentences long. "We love your work, Lars. But we've been asked to explore cost efficiencies."

He did not reply immediately. He opened Aiva's website and typed in the brief from his most recent project: "Uplifting orchestral, 90 seconds, aspirational, strings and piano, building to crescendo." Four minutes later, the algorithm delivered three variations. He listened to all of them. They were competent. They were also, in a way he struggled to articulate, bloodless.

The Royalty System That Broke

Music has always been a strange economy. Unlike novels or films, which generate revenue through discrete sales or tickets, music generates revenue through performance and reproduction—concepts invented in the nineteenth century and codified into law in the Berne Convention of 1886. Composers and performers were promised payment every time their work was played in public or copied. Collecting societies were established to track and distribute those payments.

The system survived the transition from sheet music to records, from radio to streaming. But it was not designed for a world in which the music itself could be generated on demand by machines trained on existing works. Who owns a melody that sounds like Chopin but was written by an algorithm? Who gets paid when a listener asks for "something like Billie Eilish" and receives a synthetic track that mimics her cadence but contains none of her recordings?

The labels decided the answer was: the labels. Spotify's framework treats AI-generated music as a derivative work, not an original composition. Artists whose recordings were used in training receive a micro-royalty—typically a fraction of a cent per track. The platform and the label split the rest. Session musicians, arrangers, engineers, and producers receive nothing unless they negotiated a stake in the original master recording, which most did not.

Ekström is one of thousands caught in this gap. He estimates that his string arrangements appear on at least two hundred commercially released tracks. Many of those tracks are now part of training datasets. He has received nothing. When I ask if he considered legal action, he laughs. "Against who? Spotify has two hundred lawyers. I have a studio in Södermalm and a reputation I'd like to keep."

The Clients Who Apologise

In February 2026, Ekström received a commission from a Danish film production company. They needed an original score for a short documentary about Greenland's melting ice sheets. The budget was €8,000—half what he would have charged five years ago, but more than he had earned in the previous two months. He accepted immediately.

Three weeks later, the producer called. The broadcaster had requested a revised budget. Could Ekström work for €4,500? "She was very apologetic," he recalls. "She said, 'We tried AI first, but it didn't understand the emotional arc.' I told her I appreciated the honesty. Then I said yes, because I needed the work."

This is the new negotiating position: human labour as the fallback when the algorithm fails. Ekström is not alone. Across Europe and North America, the commercial music industry is bifurcating into two tiers. At the top: celebrity producers and composers whose names are brands—Hans Zimmer, Max Martin, Finneas. Their work commands fees that make AI irrelevant. At the bottom: everyone else, competing with software that improves every quarter.

In the UK, the Musicians' Union reported a thirty-four per cent decline in arranger employment between 2024 and 2025. In Germany, the Deutscher Komponistenverband logged a forty-one per cent drop in new commissions for film and television scores. In the United States, the American Federation of Musicians surveyed its membership in November 2025: sixty-two per cent of session musicians reported income declines of more than twenty-five per cent compared to 2023.

The collapse is not total. Live performance remains largely immune—audiences still pay to see humans play instruments in real time. Bespoke composition for prestige projects—feature films, high-budget advertising, orchestral commissions—still commands human fees. But the middle has fallen out: the bread-and-butter work of scoring commercials, corporate videos, podcast intros, YouTube content, and low-budget television. That work has moved to algorithms.

The Question No One Can Answer

I meet Anna Bergqvist in a café near Stockholm Central Station. She is a copyright lawyer specialising in music and a lecturer at Uppsala University. She has spent the past eighteen months trying to answer a question that neither Swedish law nor European Union law has resolved: if an AI produces a piece of music by analysing thousands of human compositions, who owns the result?

"The current legal position is absurd," she says. "EU copyright law requires human authorship. A machine cannot be an author. So technically, AI-generated music should have no copyright protection at all—it should be public domain." She stirs her coffee. "But the platforms and labels have written contracts that treat it as a derivative work owned by whoever operated the algorithm. No court has tested this yet. When it does, the entire licensing framework may collapse."

Several cases are moving through the courts. In the United States, a class-action lawsuit filed in October 2025 by session musicians against Spotify and the major labels argues that AI-generated music constitutes unauthorised derivative use of their performances. In the UK, the Ivors Academy—a songwriters' advocacy group—has filed a complaint with the Competition and Markets Authority alleging that Spotify's licensing framework constitutes an anti-competitive carve-up of the market.

But litigation takes years. In the meantime, the market has moved on. Bergqvist estimates that AI-generated music now accounts for approximately eleven per cent of all new tracks uploaded to streaming platforms each month. Spotify declined to confirm the figure but did not dispute it.

The Score He Will Not Sell

Back in Södermalm, Ekström shows me the score he is working on. Thirty-two bars, strings and piano, pencil on manuscript paper. I ask why he still writes by hand. He could use notation software—Sibelius, Finale—which would be faster and easier to share with clients.

"Because this is mine," he says. "When I write by hand, I'm the only one who has touched it. The client gets the recording, but the score stays here. It's the last thing I own."

He pauses, then adds: "And because if I put it into software, the software company owns the data. They can train their algorithm on it. I know how this works now."

This is the other shift: a generation of creative workers learning, too late, that the tools they used to make their work could also be used to replace them. Notation software, digital audio workstations, sample libraries—all of it collects data. Avid, the company behind Pro Tools, updated its terms of service in August 2025 to allow "aggregate anonymised usage data" to inform product development. Steinberg, which makes Cubase, followed in October. Both companies declined to specify whether that data would be used for generative AI training.

Ekström no longer uses Pro Tools. He records directly to tape when he can, then transfers to digital only at the final stage. It is slower. It is more expensive. It is, he insists, the only way to keep his work his own.

The Call He Has Been Expecting

On the morning we meet, Ekström receives an email from the advertising agency. The subject line: "Re: Football spot—final decision." He reads it aloud: "Lars—thanks for your patience. We've decided to move forward with an AI solution for this project. Budget pressures. Hope we can work together again soon."

He closes the laptop. For a long time, he says nothing. Then he picks up the manuscript score—three days of work, thirty-two bars, pencil on paper—and places it carefully in a drawer.

"I knew this was coming," he says. "I just didn't think it would come this fast."

I ask what he will do now. He looks around the studio—the Neve console, the Steinway, the synthesisers, the shelves of scores and session tapes accumulated over three decades. "I don't know," he says. "Maybe I'll teach. Maybe I'll score films no one pays me for. Maybe I'll get very good at explaining to my daughter why her education fund is smaller than I promised."

He pauses. "Or maybe the courts will decide this was all illegal and the licensing deals will collapse and we'll get our industry back." He does not sound like he believes it.

What the Algorithm Cannot Hear

Before I leave, Ekström plays me two recordings. The first is the AI-generated track his client chose instead of his score. It is competent, professional, unobjectionable. It builds to a crescendo at exactly the right moment. The second is a demo he recorded of his own arrangement—strings, piano, thirty-two bars.

The difference is subtle. The AI version is smoother, more predictable. Ekström's version has small imperfections—a violin entry half a beat late, a piano chord voiced slightly differently than convention would dictate. These are not mistakes. They are choices.

"The algorithm optimises for average listener preference," he explains. "It has been trained on a million recordings, and it knows what works. But it doesn't know why a particular listener, at a particular moment, might need something slightly wrong. That's what I'm trying to do—write something slightly wrong in exactly the right way."

I ask if he thinks most listeners can hear the difference. He smiles, sadly. "No. That's the problem."

On the desk, the manuscript score sits in its drawer. On the shelf, the photograph of his daughter smiles from her graduation day. Outside, Stockholm is moving into spring. Inside, Lars Ekström is trying to calculate how many more sessions he has left before the studio becomes unaffordable and the work he has done for thirty-three years becomes something only machines are paid to do.