I found my old music in an AI training dataset.

Quick context first. In a past life I made music. That’s actually what pulled me into data in the first place. I wanted to know where my fans were, which cities were spending real time with the songs, which zip codes to hit if I ever toured. Geo-demographics for ROI. That was the entry point. Eventually the analytics kinda ate the music and became the whole thing, and now I do this for a living instead.

Which is why the story I’m about to write lands a little sideways for me.

I heard about the whole thing the way most people probably did. The Atlantic broke a piece about millions of songs sitting inside AI training datasets, searchable by artist name. Instead of getting mad about it up front, I did what most engineers do. I went to check if my own catalog was in there.

It was.

That kicked off a much longer rabbit hole than I planned for. Into the actual paper behind one of these datasets, the company that built it, and a sequence of events over the following weeks that told me a lot about how these organizations behave once a spotlight lands on them.

What the Atlantic actually found

Reporter Alex Reisner’s recent “AI Watchdog” investigation identified four datasets circulating in the AI development community, together holding roughly 21.2 million tracks. The two biggest were LAION-DISCO-12M (around 12.6 million tracks, built by the nonprofit LAION) and Sleeping-DISCO-9M (around 9 million tracks, built by a group calling itself Sleeping AI). The Atlantic built a search tool specifically so artists could check their own catalogs against these datasets. I used it, and that’s how I confirmed my recordings, along with many friends of mine, were sitting inside Sleeping-DISCO.

Here’s the technical finding that stuck with me. Three of the four datasets don’t host audio at all. They’re structured pointer systems. Metadata plus links, mostly to YouTube, and the developers using them pair those link lists with automated download tools that bypass logins, ads, and monetization entirely to pull the actual audio. The “it’s just metadata” defense falls apart the second you understand what the metadata is actually for.

Going digging: what Sleeping-DISCO actually was

Once I knew my track was in there, I wanted to understand the thing itself, not just the headline. So I went to the source as an engineer. An arXiv paper titled Sleeping-DISCO 9M: A large-scale pre-training dataset for generative music modeling, published by a small team (Sleeping AI, Wynd Labs, and a researcher at L3S Research Centre) back in June 2025.

On paper it framed itself as a research contribution. The pitch was that previous open datasets (DISCO-10M, LAION-DISCO-12M) were practically YouTube video links and limited metadata, so Sleeping AI would do it better. Scrape Genius.com for full metadata (artist, album, producer credits, genre tags, lyric embeddings, YouTube links) across 648,118 artists and 169 languages, and release the whole thing to the community.

Underneath the framing, it was the same kind of structured pointer system the white paper called out. It didn’t host audio. It hosted links, plus enough metadata to make those links useful for automated retrieval at scale. The paper’s own comparison tables labeled this “audio: yes,” which is technically true and practically misleading. What’s actually stored is a way to get the audio without asking anyone.

The methodology section of the paper is where it gets genuinely uncomfortable to read as an artist. They built a custom Python spider using a library called cloudscraper, whose entire purpose is bypassing Cloudflare’s bot protection, then parsed the resulting HTML with BeautifulSoup to pull structured fields off Genius: song details, album and artist credits, record information. In plain terms, they engineered around the exact protections a site puts up to stop this kind of mass scraping, and describe it in the paper as a routine technical step.

Once they had the metadata, they still needed a way to point each song at actual playable audio, since Genius itself doesn’t host any. So they ran the song titles through embeddings, then used what they call the “Grass Foundation scraping pipeline” to search YouTube and match each song to the video with the highest title-embedding similarity. They checked the video title and description afterward to confirm it was a real match. The paper includes its own table of these matches, and some of the similarity scores are rough. One listed at 0.51 similarity paired “O2xGen Jjsnn” with a JJ72 video that doesn’t look like the same song at all. That’s the layer actually responsible for turning “we have some metadata” into “here’s a playable link to your song,” and by their own numbers, it isn’t especially reliable.

There’s one more detail buried in the paper that’s worth sitting with. They mention discovering two additional data fields while scraping: Genius’s own written annotations (their crowdsourced explanations of songs) and the lyrics themselves. Both got explicitly withheld from the public release, not for the artists’ sake, but because, in their words, “the exclusive rights are reserved for Genius to use them.” They understood clearly enough that lyrics and annotations carried real rights they couldn’t touch. They just didn’t extend that same reasoning to the underlying recordings the whole dataset points at.

The license on top of all this is CC-BY-NC-ND 4.0, no derivatives allowed. Which is its own quiet contradiction, since the entire stated purpose of the dataset is to be used as pre-training data. Training a model on something is about as clear a case of a derivative work as exists, and the paper never addresses that tension at all.

For context on what consent actually looks like in music: you can’t legally release a track that samples another artist’s recording without written clearance from the rights holder. That’s not a debated principle, not an edge case, it’s the law of the industry. Every artist in this dataset knows that rule. How many do you think were asked for written consent for any of this? I know I didn’t.

What I did next

Once my track was confirmed in there, I started digging more.

Checked the arXiv listing for the paper itself. Found it had been quietly withdrawn on June 22, 2026, one day after the Atlantic story ran. The stated reason was “an administrative error regarding submission authorship and institutional affiliation.” Not licensing. Not ethics. Not consent. An authorship technicality, timed suspiciously close to the story that put a spotlight on the dataset.

The other dataset is still sitting there, in the open

While all this was happening with Sleeping-DISCO, I went and pulled up LAION-DISCO-12M directly. The larger of the two datasets. The one Sleeping-DISCO’s own paper had positioned itself against. No withdrawal. No sunset notice. No 404. As of my last check, it’s fully live on Hugging Face. 12.3 million rows.

A few technical details on the dataset page are worth sitting with.

The scraping method is a deliberate, documented escalation. LAION’s own writeup explains that they started with the same 18-artist seed list the original DISCO-10M authors used, and found it “not sufficient.” It only surfaced 90,007 artists and 5.4 million songs. So they built a bigger seed list from YouTube Music’s own top-song and genre charts (45,218 artists), then let a recursive crawl run through YouTube’s “Fans might also like” graph “for as long as we are able to find new artists.” That’s not passive data collection. That’s an explicit decision to keep expanding the crawl until it hit a wall, specifically because the first pass wasn’t comprehensive enough.

Downstream training is listed openly, not hidden. Right on the dataset page, under “Models trained or fine-tuned on this dataset,” there are two live entries. An audio-to-audio model called EAR_VAE and a text-generation model, one updated as recently as a few weeks before I looked. Compare that to Sleeping AI’s insistence, in their sunset notice, that “no generative AI models were ever trained on this metadata.” LAION isn’t even trying to hide it. The citation graph does that work for them, automatically, in public.

And here’s the detail that actually made me laugh out loud. The dataset card ends with a polite request: “This work was inspired by DISCO-10M, consider citing them if you use this dataset.” An organization that scraped the metadata of 250,516 artists, many of whom were never asked, never credited, and never paid, is asking the AI community to please remember to credit the researchers if they build on top of it. Academic citation etiquette, applied rigorously in exactly one direction. The artists whose songs made the dataset worth anything get a YouTube URL and a play count. Spotify pays those same artists around $0.004 per stream, if that helps anchor the scale. The people who assembled the pointer list get a request for attribution.

The sunset notice, and why I don’t buy it

Here’s the notice, in full, as posted on their site:

“The Sleeping-DISCO dataset has been permanently decommissioned and completely purged from all Sleeping AI infrastructure and operational servers. This absolute erasure was executed proactively by our team in response to emerging public controversies, localized targeting behaviors, and escalating incidents of doxing that compromised the personal security, privacy, and individual liberties of independent authors. Sleeping AI explicitly clarifies that Sleeping-DISCO was strictly a metadata repository; at no point did the dataset host, index, or distribute any copyrighted audio, video, or external multimedia content, nor were any generative AI models ever trained on this metadata. By completely eliminating the repository, Sleeping AI has removed the vector being used to target independent creators, rendering any further disputes or data removal requests entirely moot.”

Reading it closely, a few things jump out.

“Rendering any further disputes or data removal requests entirely moot” is the actual point of the notice. That seems to me the entire point of the notice.

“No generative AI models were ever trained on this metadata” directly contradicts their own paper. The entire premise of the arXiv submission, the one just withdrawn on an unrelated technicality, was that this was, quote, “a large-scale pre-training dataset” built explicitly for training generative music models. Can you market something for a year as training data and then, after press scrutiny, claim no training ever happened?

The technically-true parts are doing the most obscuring. Yes, the dataset was metadata and links, not hosted audio. That was true from day one, and it’s exactly the loophole the Atlantic’s investigation called out. Link-based datasets are still fully functional as training pipelines when paired with the download tooling that developers actually use.

What “pulled down” doesn’t mean

Being clear-eyed about what actually changed here, because “decommissioned” sounds a lot more final than it is.

It stops new downloads from that specific repository. It does nothing about the copies already taken during the roughly year the dataset was live and, per the Atlantic’s reporting, widely downloaded.
If any model was in fact trained on data retrieved through this dataset, that training already happened. Deleting the source doesn’t un-train a model.
The same organization’s other datasets, including ones built from Udio- and Suno-generated songs, were, as of my last check, still live and being actively updated. This wasn’t a wholesale shutdown. It was one specific, high-visibility liability getting cauterized.

What I can’t prove

One honest caveat before I close this out. Being in one of these datasets is not the same as being confirmed in a trained model’s weights. My track is documented sitting inside Sleeping-DISCO. Whether any specific generative model actually ingested it, and whether any specific output was influenced by it, I can’t prove from the outside. What I can prove is that a pointer to my song was in a pipeline explicitly built for that purpose, released under a research-adjacent label, and downloaded by unknown parties for unknown uses over the course of a year. The presence is documented. The downstream use, on a per-track basis, is not, and I want to name that gap out loud rather than let it hide inside the rest of the argument.

This is just getting started

None of this is happening in a small market. Suno AI, one of the most visible generative-music companies out there, was valued at around $500 million in mid-2024. By 2026 it’s sitting at roughly $5.4 billion. A 10x jump in under two years, for a company whose entire product depends on models trained on massive corpora of existing recorded music. Demand for training data doesn’t shrink on that kind of curve. It gets more aggressive, better funded, and more insulated from consequences. Sleeping-DISCO getting pulled down is not the end of anything. It’s the shape of what a new industry looks like when the incentive to acquire training data at scale is measured in billions of dollars.

Why I’m writing this down

I’m not a policy expert and I’m not a lawyer, and I don’t have a tidy conclusion to offer. This is a dataset built explicitly to train AI on real, well-known and small independent artists’ work. A year of it being freely downloadable. A press story that made the mechanism impossible to ignore. A paper withdrawal on a technicality. That seems less of an apology and more of a ‘Hey, people see this now’.

Weird thing to sit with, though. The whole reason I ever got into data in the first place was to figure out where my music was landing. Where the fans were. Which markets to hit. It was upstream analytics on my own catalog, run by me, for me. Ten years later a group I’ve never spoken to scraped that same catalog into a pointer list, called it a research contribution, and shipped it out to the training-data community to produce multimodal models that generate replicating sounds and make multi-billion dollar valuations.

Whatever happens with any individual takedown, having your own record of what you found and when might be the only piece of this whole process that’s actually in your control.

Sources

The views expressed in this post are my own personal opinions only.