• Pascal's Chatbot Q&As
  • Posts
  • The Legal Reckoning Facing AI Firms. The investigation by the International Confederation of Music Publishers (ICMP) as revealed by Billboard, has been labeled "the largest IP theft in human history."

The Legal Reckoning Facing AI Firms. The investigation by the International Confederation of Music Publishers (ICMP) as revealed by Billboard, has been labeled "the largest IP theft in human history."

ICMP’s evidence — implicating AI firms such as OpenAI, Meta, Google, Microsoft, and others — exposes the widespread unauthorized scraping and use of copyrighted music, lyrics, and even album artwork.

Author

Pascal Hetzscholdt
September 04, 2025

by ChatGPT-4o

Introduction: The Music Industry's Smoking Gun

The recent investigation by the International Confederation of Music Publishers (ICMP), as revealed by Billboard, has been labeled "the largest IP theft in human history." ICMP’s evidence — implicating AI firms such as OpenAI, Meta, Google, Microsoft, and others — exposes the widespread unauthorized scraping and use of copyrighted music, lyrics, and even album artwork to train generative AI (GenAI) systems. With datasets allegedly sourced from platforms like YouTube and Spotify, and with LLMs producing outputs replicating protected content, AI firms may now face an unprecedented wave of litigation.

This moment represents a tipping point not just for the music industry, but for all content sectors — including books, images, academic works, journalism, and film — whose creators are watching closely. Below, we unpack the legal risks AI firms face in the U.S. and beyond, what they could have done differently, whether remediation is still possible, and what broader consequences are likely to follow.

What the Evidence Consists Of

The evidence compiled by the International Confederation of Music Publishers (ICMP) spans a wide array of sources and clearly demonstrates the unlicensed and systematic use of copyright-protected music in AI training. Collected over two years and verified through open-source intelligence, leaked documents, and expert analysis, the dossier includes:

  • Scraped Datasets: Private datasets showing that AI music apps like Udio and Suno scraped YouTube for songs without permission, including mass-scale harvesting of lyrics and audio.

  • Model Outputs: Analyses of LLMs such as Meta’s LLaMA 3 and Anthropic’s Claude, which produce outputs reproducing copyrighted lyrics from songs by artists like Beyoncé, Bob Dylan, Ed Sheeran, Kanye West, and others.

  • Admissions by AI Firms: A direct admission from OpenAI that its music-generation tool Jukebox was trained on copyrighted songs from The Beatles, Madonna, Drake, and more. Similarly, Google’s Gemini chatbot acknowledged that its MusicLM model likely trained on protected music content.

  • Legal Filings: Court documents from ongoing lawsuits against companies like Anthropic, which show evidence of both input (training) and output (generation) involving unlicensed copyrighted lyrics.

  • Leaked and Labeled Training Sets: Data from companies such as Runway and Google’s AudioSet, which reveal internal organization of scraped music files by artist, genre, tempo, and track — indicating intentional and commercial-grade ingestion of copyrighted material.

  • Visual Art Evidence: Midjourney’s outputs include AI-generated replicas of iconic album covers by Gorillaz, Dr. Dre, and Bob Marley, reinforcing concerns that visual copyright infringement is also widespread.

  • Double Standards in ToS: ICMP also highlights the contradiction in tech firms’ own terms of service, which prohibit scraping their platforms, even as these same companies scrape others’ content for training purposes.

This collection of multi-source, cross-verified evidence demonstrates not only scale and intent, but also the commercial application of unlicensed creative works — a critical factor in triggering global litigation and regulatory attention.

A. The “Fair Use” Defense Under Pressure

In the U.S., most AI firms have leaned on the doctrine of fair use to justify using copyrighted material in training data. Courts have historically offered some latitude here, especially in the realm of transformative use. However, AI model training has pushed this to the limit.

  • A California judge recently acknowledged that it remains an “open question” whether using copyrighted content for AI training qualifies as fair use — indicating legal uncertainty.

  • Meta and Anthropic won technical motions related to fair use in author cases, but the rulings were narrow. For example, Judge Vince Chhabria explicitly stated that Meta’s victory “does not stand for the proposition that [training] is lawful” — only that the plaintiffs’ arguments were insufficient.

This implies future cases — especially those with stronger evidence like the ICMP dossier — could result in different outcomes, especially if courts perceive harm to the market for the original work (e.g., if AI replaces human lyricists or composers).

B. Lawsuits Already Underway

  • Major publishers (UMPG, Concord, ABKCO) are already suing Anthropic for infringing song lyrics.

  • All three major music labels are suing AI music generators Suno and Udio.

  • Other cases include Getty Images vs Stability AI, and OpenAI/Microsoft vs New York Times.

Each of these could set landmark precedents if they survive dismissal and proceed to full trial — especially when plaintiffs present concrete evidence of direct copying, verbatim outputs, and market harm.

A. European Union – Stronger for Rights Holders

The EU's AI Act and existing copyright law provide far more robust protections than U.S. fair use:

  • It mandates transparency regarding training data.

  • It honors rights reservations (opt-outs or exclusive licensing).

  • It applies irrespective of where the data was sourced — meaning offshore scraping won’t escape EU scrutiny.

  • ICMP has actively lobbied EU officials and shared evidence — likely bolstering future enforcement efforts.

Firms like OpenAI and Meta, whose services reach EU users, may thus face cross-border liability, including fines, injunctions, or market access restrictions.

B. United Kingdom and Other Common Law Jurisdictions

While the UK does not currently offer a fair use exception similar to the U.S., it has resisted stronger copyright protections for training data. However, the Getty v Stability AI case in the London High Court may set a new precedent, especially given Getty's claim of direct copying of millions of images.

Canada, Australia, and Japan are reviewing their AI and copyright frameworks, and could pivot quickly in response to pressure from their music industries or international alignment with the EU.

III. How AI Firms Could Have Prevented This

  1. Proactively Licensing Content
    Firms had the option — and in many cases were offered opportunities — to license lyrics, compositions, and performances through existing rights organizations. These include mechanical rights societies, PROs, and publisher aggregators. They chose not to.

  2. Using Synthetic or Licensed Datasets
    Rather than scraping YouTube or Spotify, AI developers could have built datasets from public domain works, Creative Commons sources, or licensed music catalogs from willing partners.

  3. Transparency and Disclosure
    Most firms refused to reveal training data sources, citing trade secrets — yet internal documents reveal that datasets included highly specific metadata (song names, genres, tempos, etc.), suggesting deliberate collection.

IV. Can This Still Be Fixed — Or Is It Too Late?

It is not too late, but remedial action will be costly and complex:

  • Licensing retroactively: Firms may need to negotiate “make-good” licenses and royalties for past usage — similar to retroactive mechanical licensing settlements in music streaming.

  • Dataset audits and unlearning: AI firms will likely be forced to conduct third-party audits of training data and may have to “unlearn” outputs based on infringing inputs — a technically challenging and ethically fraught process.

  • Transparency mandates: Future compliance with laws like the EU AI Act will require AI makers to disclose data sources and obtain rights — limiting their ability to rely on opaque scraping practices.

The longer firms delay, the worse the exposure: damages could be trebled for willful infringement under U.S. law, and loss of market access is a real threat under EU frameworks.

V. Consequences When Other Sectors Join the Fight

The ICMP dossier is just the beginning. Other content sectors are likely to follow, bringing with them:

A. Lawsuits from Book, Film, Academic and Journalism Sectors

  • Books: Authors Guild and others have already sued; academic publishers like Wiley, Springer Nature, and Elsevier may follow if research content was scraped.

  • News: NYT's lawsuit against OpenAI/Microsoft is a leading case that could redefine news scraping legality.

  • Film/TV: Disney and Universal have taken aim at Midjourney for image theft, but music video and film soundtracks may open another front.

  • Education/Academia: AI outputs that mimic journal articles or textbooks raise substantial copyright and integrity concerns.

B. Strategic Coalitions and Cross-Sector Evidence Sharing

ICMP's model — combining open-source investigation, leaked datasets, and commercial impact assessment — will likely be adopted by other rights holder groups, making litigation more coordinated, well-resourced, and global in reach.

C. Regulatory Backlash and Platform Penalties

  • Access restrictions: AI tools may be banned in certain markets unless they comply with licensing regimes.

  • Reputational damage: Perception of tech hypocrisy (wanting free use of others' content while forbidding scraping of their own) will trigger public and political pushback.

  • Loss of training ground: Rights holders may cut off future access to valuable content, diminishing model quality and undermining commercial viability.

Conclusion: “License or Desist” — A New Paradigm for AI Companies

The ICMP investigation is more than just a legal exposé — it’s a blueprint for the music industry’s strategic response to AI copyright infringement. The evidence is extensive, the stakes are existential, and the message is clear: AI companies cannot continue to extract massive value from copyrighted content without consequence.

What started in the music sector will ripple into every content industry. For AI developers, investors, and executives, the warning signs are blinking red: fix your datasets, license your inputs, disclose your sources — or prepare for regulatory and legal escalation on a global scale.

It’s no longer about whether infringement occurred — but about whether the AI industry is ready to take responsibility.

·

25 AUG

Apple vs. Oppo: A Case of Trade Secrets Theft and Its Implications for AI Litigation

·

31 AUG

Elon Musk’s IP Hypocrisy — Trade Secrets, AI Theft, and the Fantasy of Abolishing IP Law