My favorite season is Fall. Upstate New York is known for its apples, and I love everything about them. The picking, the eating, the cider, the cider donuts. Apples are amazing by themselves, but they can be turned into dozens of things. Pies, crisps, caramel apples, the list goes on.

The podcasting ecosystem is a lot like a massive apple orchard in the Fall. Tons of apples, ripe and ready for harvesting and to be used in a variety of ways. As a consumer of apples, I can look at them and enjoy them all I want, but as soon as I pick one, I've got to pay for it. I can't just walk out of the orchard with a bag full of someone else's crop. The farmer wants to be compensated.

This is exactly what companies are doing with your podcast to train their AI models. Walking away with a bag full of apples they didn't pay for.

Bags Full of Apples

Before I started writing this, the assumption was that AI models were being trained on non-licensed data. I hadn't truly investigated whether this was real or just industry anxiety. It didn't take long to find out.

Muyan-TTS is an open-source text-to-speech model built specifically for podcast scenarios. Its creators trained it on over 100,000 hours of podcast audio collected from open-source datasets and what they describe as a "proprietary podcast collection." If you assume the average episode is 45 minutes, that's roughly 133,000 episodes worth of human speech. The model was released under an Apache 2.0 license, meaning anyone can use it commercially. No evidence exists that any podcast creator was asked, informed, or compensated.

The Emilia dataset contains over 101,000 hours of speech collected from "various video platforms and podcasts on the Internet," spanning talk shows, interviews, debates, and commentary. Again, publicly available with no licensing structure in sight.

Muyan and Emilia are the examples we can see. The ones we can't see are almost certainly larger. OpenAI, Suno, Udio and Apple have all been sued for consuming text, music, images, and videos they allegedly had no rights to. Lawsuits have since turned into licensing conversations, but podcast audio has been remarkably absent from these conversations.

No Farmstand

Lawsuits turning into licensing conversations occurred because the infrastructure for licensing different kinds of media has been in place for a long time. When AI training became a new use case, the music industry had collection infrastructure, legal precedent, and institutional muscle to bargain with.

Podcasting has no equivalent. There is no organization that tracks who owns what podcast audio. There is no registry of rights. There is no standard licensing agreement for AI training use. There is no mechanism for a podcast creator to opt in or opt out of their audio being used this way. There isn't even a reliable way to know it's happening.

Ripe for the Picking

Voice AI models need large volumes of natural, conversational human speech. Podcast audio is naturally expressive, topically diverse, often multi-speaker, and available in enormous quantity. It's the largest corpus of conversational human speech on the internet, and almost none of it was recorded with the expectation that it would be fed into a machine learning pipeline.

The qualities that make a podcast good are the qualities that make the audio valuable as training data. Authentic tone, emotional range, conversational rhythm, the natural back-and-forth of a real interview. The better your show, the more useful your voice is to someone training a model to sound human.

According to the Podcast Index, there are more than 159 million episodes. All of them can be readily consumed without consent or compensation, and no infrastructure exists to manage either.

Back to the Orchard

This isn't an argument against AI. The companies and creators suing over training data aren't anti-AI. They want to be compensated for the time and energy they put into their creative work.

Podcast creators deserve the same. Not protection from the future, but participation in it. That means someone has to build the infrastructure that makes participation possible. Podcast creators need a way to say yes, or no, and a way to get paid.

Right now, podcast audio is the raw material for a multi-billion dollar industry, and the people who made it are the only ones not at the table. One hundred and fifty-nine million episodes. That's a lot of bags of apples no one’s paying for.