If you’re a recording artists or musician, you have most likely heard some form of this pitch:

“Don’t worry. If you don’t like how your music is used for AI, we can always take it back out later.”

That sounds fair.

The hard part is that, today, it is not that simple. And it’s so not simple, that statement may as well be untrue.

That is why we created the  Machine Unlearning Research Hub  at MusicTechPolicy.com—to help musicians understand what machine unlearning is, what it can do today, and what it may be able to do tomorrow.

Think of it like baking a cake

Suppose you bake a cake with flour, eggs, butter, milk, and sugar. The eggs make the cake what it is.

Once the cake comes out of the oven, you cannot remove just the eggs. AI training works in much the same way.

When an AI developer “trains” a model (often on illegally acquired recordings as we are seeing in the litigation), it does not simply store copies of recordings in a folder. It changes billions—or even trillions—of internal values based on every recording it has been trained on and extracted from those training tracks or its “training pipeline”. Each recording and the values the AI extracts from the recording helps shape the finished model.

That’s why it knows what you mean when you ask it to play an Eric Clapton solo or a John Bonham drum fill. That is also why taking one recording back out later is so difficult.

Surely, you say, they didn’t build a dataset that cannot be corrected and only gets larger with human personal rights and copyrights? Well, they kind of did. Not that different from Google Street View. There’s a long “Hotel California” tradition with these people.

So what is machine unlearning?

Machine unlearning is a field of research that asks a simple question:

Can an AI system forget what it learned from a particular piece of data without rebuilding the entire model from scratch?

The quick answer today is not very easily, no matter what they tell you. Researchers all over the world are working on that question.

The results are a mixed bag. A cynic—who me?—might say they’re not trying very hard. But some techniques show real promise. Others work only under very limited conditions with largely toy datasets.

For today’s largest music models, there is still no widely accepted way to guarantee that every artifact of one recording—or one artist—has been completely removed.

That does not mean machine unlearning will never work. It means it is still a research problem.

How much can it fix today?

The honest answer is:

Some things. Not everything.

Researchers have developed methods that can reduce a model’s reliance on particular data or make it much less likely to produce certain results. That’s not 100% or zero depending on how you look at it. But it’s better every year and those advances matter.

But reducing influence is not the same thing as proving that a recording has been fully removed from a commercial AI model. For musicians, that difference is obviously important.

Will it get better?

Almost certainly.

Machine unlearning is receiving significant attention from universities, technology companies, and governments. Why? Well, for one thing there are countries that have the “right to be forgotten” that Google fought so hard against in Europe and especially Germany. There’s also other privacy laws that all of the AI models are essentially out of compliance with, so it’s only a matter of time until there’s a real effort to do something about it.

Five years from now, today’s methods will almost certainly look primitive. Ten years from now, the tools may be far better still. But no one can honestly promise today that future improvements will solve every problem created by yesterday’s training.

That is why decisions made now still matter. Let’s be honest, the best way to stay out of AI is to never be included in AI.

What does this mean if someone wants to license your music?

Ask yourself one simple question:

If I change my mind in three years, how will my music come back out?

Do not settle for vague promises.

Ask specific questions.

• Will my recordings be kept in a separate training set?
• Can they be removed without rebuilding the model?
• What machine unlearning process will be used?
• Has that process been independently tested?
• How will you prove my music was removed?
• What happens if the technology cannot fully remove it?
• Who pays to remove it?

If the company cannot answer those questions today, it is worth understanding that they are asking you to rely on technology that may not yet exist in the form they hope it will. Or as Blanche Dubois said in Streetcar Named Desire, “I have always relied on the kindness of strangers.” Feel good yet?

That does not necessarily mean you should refuse every AI license.

It does mean you should know exactly what risk you are accepting.

This is why we built the Machine Unlearning Research Hub

Machine unlearning is moving quickly, and new research papers appear almost every week. Some represent genuine breakthroughs, others improve only narrow parts of the problem.

Our goal is not to tell musicians what decision to make. Our goal is to help musicians make informed decisions based on the best available evidence. If someone asks you to license your music for AI, you deserve to know one thing before you sign:

Can they realistically give it back if you ask?

Today, that question deserves a careful answer—not a marketing slogan. That is what the Machine Unlearning Research Hub is here to explore.