A Rare, Concrete Look at How AI Music Is Really Trained
The Suno AI scraping leak is a breach of source code and internal data that exposes how a leading music generator collected millions of songs, lyrics, and audio clips from online platforms to train its models, providing an unusually detailed window into AI training data practices and their copyright risks.
According to a hacker using the alias ellie.191, Suno scraped millions of songs and lyrics from YouTube Music, Deezer, Genius, stock libraries like Pond5, Jamendo, Freesound, the International Music Score Library Project, and even podcasts via RSS feeds, then shared details of those training libraries with a news outlet. The leaked material includes source code from 2023 and 2024 with explicit scraping instructions, mapping a pipeline that turns public-facing music services into AI training fuel. Suno has already admitted in legal proceedings that its systems were trained on “essentially all music files of reasonable quality that are accessible on the open internet,” totaling “tens of millions of recordings.” The leak does more than confirm that claim—it shows how it was operationalized, and why AI training data copyright is now the central fault line between generative AI ambition and artists’ rights.
From Allegation to Evidence: The Music Generator Lawsuit Grows Teeth
The most important shift is not technical but legal: this breach moves the fight over AI music from theory to documented practice. Suno has been the target of several major music generator lawsuits, with labels accusing it of training on millions of copyrighted songs. Until now, much of that dispute turned on broad claims—fair use versus infringement, innovation versus protection. The hacked code adds factual scaffolding the lawsuits previously lacked.
One leaked youtube_music file logs 2,013,545 ingested clips, giving a precise number to what had been a vague charge of large-scale copying. Dataset annotations list 113,879 hours of YouTube Music, 17,615 hours of Genius, 62,117 hours of Pond5, and 12,287 hours of Deezer, among others—a catalog measured in decades of listening time. “Allegations involving millions of songs and lyrics could add factual detail to the copyright fight major labels opened.” Labels also claim Suno engaged in unlawful stream ripping and circumvented technological measures on YouTube, accusations that go beyond copyright into anti-circumvention law. Independent authentication is still required before courts treat each instruction as genuine, but if validated, these files give judges more than abstract doctrine—they supply a playbook.
Inside the Scraping Pipeline: Why Artists See an Existential Threat
The leaked instructions do not describe casual scraping; they describe an engineered acquisition pipeline designed to feed an insatiable model. Comments in the code reference sources like “genius_hq, youtube_music, freesound, jamendo, imp, deezer, ytm_tagged,” with notes that “non-music will be filtered out,” revealing service-specific logic rather than opportunistic copying. Instructions reportedly from 2023 and 2024 cover collection from YouTube Music, Deezer, Genius, Pond5, Jamendo, Freesound, the International Music Score Library Project, and podcast feeds.
A single dataset file cites 113,879 hours of YouTube Music, 152,162 hours of tagged YouTube material, 62,117 hours of Pond5 music, 19,514 hours from the International Music Score Library Project, and 3,726 hours of Jamendo, among other figures, amounting to “at least decades worth of music.” Searches for a cappella tracks targeted vocal-only audio, while a separate plan aimed at about one million podcast hours via PodcastIndex, all routed through proxy services to distribute requests. For working artists and rights holders, this scale explains the anxiety: when a model inhales millions of tracks and lyrics, the line between inspiration and replacement blurs. As musician Kenneth Blum (Kenny Beats) put it, scraping like this feels like “stealing from countless struggling musicians” and “obliterating the work and dreams of artists.”
Fair Use, Circumvention, and the Coming AI Training Data Reckoning
Suno’s defense is the now-familiar refrain in AI training data copyright debates: training on publicly accessible files is fair use because the models create new outputs rather than substitutes for the original recordings. But the leak forces a harder question: can companies claim transformative fair use while allegedly bypassing technical protections and systematically ripping streams? Labels argue that Suno “obtained those copies in the first instance by unlawfully ‘stream ripping’” YouTube and “circumventing the technological measures” intended to prevent such copying, creating a circumvention issue that sits alongside the fair use fight.
The legal stakes are intertwined with business reality. Suno is one of the largest AI music generation tools and has already faced several major lawsuits from the record industry. Warner Music Group settled its case with Suno and entered a partnership in late November 2025, with licensed model development and opt-in use of participating artists’ names, images, likenesses, voices, and compositions. Future models are planned to compensate participating artists, making consent and payment part of their design. That deal signals where generative AI accountability is likely headed: not an end to AI training, but a shift from unilateral scraping to structured licensing and revenue sharing. The leak makes it harder for any model developer to argue ignorance or inevitability; it shows that acquisition mechanics are a choice, not a law of nature.
Beyond Suno: Industry-Wide Accountability After the Leak
It would be a mistake for other AI developers to treat the Suno AI scraping scandal as someone else’s problem. The hacked data is “a rare look at exactly how AI models and tools are built,” and it exposes methods that are likely mirrored, in spirit if not in detail, across the industry. The same incident reportedly let the hacker access customer information for hundreds of thousands of users, including contact details and Stripe payment information, even as Suno denies that sensitive personal data was compromised. Data-hungry models and security shortcuts make a dangerous pair.
Independent authentication of the leaked files will shape their weight in court, but regardless of outcome, the precedent is clear: courts, artists, and regulators will demand visibility into how generative systems are built, what they ingest, and how consent is—or is not—obtained. Suno has already shown a path by pivoting to licensed partnerships, scheduling the first music model developed with industry partners to roll out after June 2026. If the industry ignores this moment, it risks cementing a public narrative that generative AI is built on mass, unauthorized extraction. The alternative is harder but healthier: treat training data as a governed resource, not a free buffet, and build models that respect the people whose work makes them possible.
