The scramble is on to help unscramble the mess that is AI-generated music

The thing people forget about AI-generated music is that it’s not very intelligent. It’s just very good at guessing.

This is how it works:

Songs are fed into an AI program such as Suno, Udio, Soundful, Boomy, Musicfy, Playbeat and a dozen others. The program logs billions of data points about how each song is constructed, using that information to grow the knowledge and breadth of its music-making algorithm.

When a user enters a prompt, say, “Create an avatar for a female-fronted indie band that sounds similar to The Strokes, The White Stripes and Arcade Fire and then generate an anthemic song worthy of an encore,” the program sorts through what it’s learned about such material.

It starts with one note, and then, based on statistical probability using an analysis of the song styles mentioned in the prompt, picks the next note. And so on and so on until the song is finished. Nothing new was created; the program just dug for gold in human-made music that had already been created. It’s sophisticated regurgitation, a closed loop that will never result in any innovation. We are not in danger of facing an AI Paul McCartney or Bob Dylan.

Cool tech, but there’s a legal problem. No one asked artists for their permission to allow AI programs, known as LLMs or Large Language Models, to learn from their copyrighted intellectual property. In other words, their work was stolen, taken without compensation, to create a business that will make money directly from their talent and labour.

Yes, there have been some lawsuits, but in recent months, there has been some rapprochement with both Universal Music Group and Warner Music — two of the three major labels — which reached settlements with Udio.

Both labels promise big things for AI and music in 2026. The music companies say that new licensing agreements will open up new revenue streams for artists. Some lawyers say that this technology, and this kind of music scraping, is unstoppable and that the only solution is to negotiate a legal framework for the use of human-made music with the AI companies.

Get breaking National news

For news impacting Canada and around the world, sign up for breaking news alerts delivered directly to you when they happen.

Proponents say that while artists can theoretically opt out of training licenses, the labels claim, they’ll just be leaving money on the table. (Sony, the third major label, is hanging back for now and is continuing its litigation.)

Artists are rightfully suspicious. They remember the line they were fed on streaming, and how it was going to be a financial windfall for everyone. AI-generated music is a much bigger threat to their livelihood and the control over their intellectual property. It’s feared that AI will suck away their talent and give nothing in return. As Canadian singer-songwriter Mac DeMarco told The Globe and Mail, “Soon we’ll all just be batteries, like in The Matrix.”

What if artists have had their music scraped already? How would they know? And what if they object to having some wannabe Skynet composer surreptitiously learning how to take their jobs by using their own music against them?

There are solutions. A company called Musical AI has just received a new round of funding for its attribution model. The company says its technology will “determine what materials in the training dataset influenced generative AI outputs, down to precise percentage breakdowns.” They say that it can help artists “monitor, take down, and sunset usage of works they own.”

Then there’s Hirundo, an Israeli company that’s into machine unlearning. I spoke with the company’s Ben Luria, who likens the tech to the neuralyzer, the flashy memory-erasing device in Men in Black. It can unscramble the datasets and the eleventy-billion lines of code and pick out what music shouldn’t be there.

“As generative AI systems become more capable, they are increasingly colliding with copyright law — especially in music and other creative fields,” says Luria.

“Artists are finding AI-generated material that appears to borrow from their voices, styles or catalogues without consent. At the same time, AI developers face a different problem: even when misuse is identified and acknowledged, the technology itself offers few practical ways to correct it once a model has already been trained.

“Hirundo’s view is that this has trapped the industry between two unsatisfying extremes. On one end is the idea of full takedown — shutting down models entirely until retrained without the unlicensed content, which is neither practical nor likely to be enforced. On the other is a de facto free pass, where violations are acknowledged but effectively left unresolved. Hirundo argues that what has been missing is a technically viable middle path: a way to remediate specific problems inside trained AI systems without dismantling them altogether.”

AI does not store music as individual files. The best analogy is that the information is distributed across billions of artificial neurons. AI cannot delete a specific piece of training in an Eternal Sunshine of the Spotless Mind sort of way. The tiniest bits of every song are entangled with everything else, just like our real-world memories get intertwined with other memories. It’s like looking at a fully finished cake and thinking, “I want to take out one teaspoon of sugar.”

Hirundo says it has figured out how to unscramble an egg quickly, efficiently and cheaply. Remediation is their goal. Work continues with AI developers, labs and enterprises that build and deploy LLMs so that this tech will enable them to expunge specific copyrighted works, personal data and other protected IP down to the last artificial neuron.

Luria says, “Creators should be able to say, ‘I didn’t consent to this,’ and there should be a practical way for companies to actually fix it — not just apologize for it.”

We’ve been through this before. Remember the great sampling freak-out of the late ’80s and early ’90s? It was another case where technology had raced ahead of the law, and it took a little while to sort things out. Now there are protocols, regulations, and oversight concerning who samples what and who should get paid when a piece of their music is used in the creation of another.

The AI LLM training situation is just like that, except exponentially more impactful. It’ll take a few more years for things to settle down — if, of course, the robots don’t take over first.

&copy 2026 Global News, a division of Corus Entertainment Inc.