Training Data, Copyright Ownership, and the Ethical Case Against AI Music Generation

This argument carries real weight. It has history behind it, analogy on its side, and the reasonable suspicion that moral panics about automation rarely survive contact with reality. But history also shows something else: some technologies extract value from a commons before anyone calculates the cost, and by the time the accounting arrives, the extraction is already complete. The question worth asking about AI Music Generation is not whether it will displace musicians — it is whether the mechanism by which it produces music is ethically distinct from the machines that came before it. There is reason to believe it is.

Why Every Generation Has Heard This Argument Before

The case for AI music tools rests on a foundation that is, in historical terms, almost unassailable. The player piano automated what pianists performed. The drum machine automated what session drummers played. The digital sampler captured fragments of existing recordings and assembled something new from them. Each technology reduced the market for one class of musician while expanding it for another — producers, composers, sound programmers. The economy of music did not collapse; it reorganized around whatever the new tools made possible.

From this vantage point, Suno and Udio look like the latest iteration of a familiar pattern: tools that lower the barrier to creative expression. Someone who has a melody in their head but no instrumental training can now produce a finished track in a genre that previously required years of practice or expensive session work. The same underlying architecture that powers Text-to-Speech and Voice Cloning — systems trained to model human audio at fine levels of Prosody, capturing rhythmic and tonal nuance — has been extended to music composition and production. Platforms like Mureka have emerged with fewer reported legal entanglements than their predecessors, suggesting the industry is still mapping the range of ethically defensible approaches to training data.

The legal defense follows naturally: Music Copyright And AI jurisprudence has generally treated transformative use favorably, and training a model is not the same as distributing a recording. The model does not contain the songs; it captures statistical relationships in sound. New tools have always generated resistance from those whose expertise they threatened. The argument concludes: let the market sort this out, as it always has, and the musicians still working a decade from now will have adapted to what the tools allow.

The Mechanism Nobody Is Examining

And yet there is a structural difference hiding inside this democratization story, worth examining carefully before accepting the analogy.

The player piano read sheet music — symbolic notation abstracted from any particular performance. The digital sampler took identifiable fragments, and courts consistently required licensing for precisely that reason: the originals were traceable. The drum machine generated new sounds without absorbing existing ones.

Audio Diffusion models and Neural Audio Codec architectures were trained on the recordings themselves — the final creative product of musicians, sound engineers, and producers — ingested at scale, without consent, and without compensation. The US Copyright Office’s pre-publication Part 3 report on AI training (released May 2025) noted that this use “clearly implicates the right of reproduction” and concluded that commercial training on vast quantities of content to produce competing creative works goes beyond established fair use boundaries (US Copyright Office).

That phrase matters precisely: “competing creative works.” The drum machine did not produce songs that competed with the drummers it displaced. The player piano could not write new compositions. The current generation of AI music tools does what the training data did — it competes in the same market for the same listeners, producing outputs that are, to many consumers, functionally indistinguishable from the professional recordings on which the models were trained. This is not the same mechanism as prior automation. It is a different kind of extraction, and naming it clearly changes what follows from the historical analogy.

When the Evidence Switches Sides

Once the mechanism is visible, the historical precedent stops supporting the democratization argument and begins to support its inverse.

The record industry’s lawsuit against Suno was filed in June 2024, brought by Sony Music Entertainment, UMG Recordings, and Warner Records (RIAA). During discovery in November 2025, audio fingerprinting identified over 61,000 additional copyrighted songs in the training corpus alone (TechTimes). The scale here is not incidental — it is the product. These models work precisely because they ingested a comprehensive cross-section of recorded music, absorbing not just melody and harmony but the expressive signatures, timbres, and production aesthetics that individual artists spent careers developing.

Copyright status in 2026 reflects the same inversion. Works created entirely by AI without meaningful human creative input cannot be copyrighted under current US law — a position the Supreme Court effectively confirmed in March 2026 when it declined to hear the Thaler appeal (Rimon Law). If AI-generated music has no author, who profits from its circulation? Not the musicians whose work trained the model. Not the artists whose stylistic identity it absorbed. The economic flow runs from the accumulated commons of human musical labor toward the companies that organized the ingestion, with no reciprocal obligation built into the transaction.

The streaming economy is already showing symptoms of this asymmetry. Spotify introduced a minimum threshold of 1,000 streams per year to qualify for royalties — a policy explicitly motivated by AI-generated tracks exploiting the payout system. Warner and UMG reached settlements with AI companies in late 2025 — Warner with Suno in November, UMG with Udio in October — but the Sony lawsuit against Suno remained active as of June 2026, with a hearing scheduled for July 2026 (Chartlex, TechTimes). Since those settlements, both platforms have undergone significant operational changes — Suno restricting downloads and deprecating earlier models, Udio transitioning to a walled-garden architecture that prevents audio export.

Platform changes (as of late 2025):

• Suno: Pre-settlement models deprecated; free-tier downloads removed; paid-tier monthly download caps now apply.
• Udio: Transitioned to walled-garden model; audio exports no longer available; commercial use terms changed.

Creativity Is Not a Training Set

Thesis: The ethical problem with AI music generation is not that it automates creative work, but that it does so by consuming — without consent — the irreplaceable creative labor of specific human beings, and then distributing the economic benefits to those who built the machine.

The player piano did not take the pianist’s work without permission; it played notation the pianist might have been paid to produce or chose to share publicly. The digital sampler was eventually forced, through litigation, to license the fragments it borrowed. The AI music model was built on an act of systematic appropriation that has no precise prior equivalent in the history of recorded music — and the companies that built it continued raising capital through the legal challenge without pausing to resolve the underlying question. Suno raised $400 million in additional funding in June 2026, while the Sony lawsuit was still pending (TechCrunch). The funding did not wait for the courts. The courts did not wait for the funding. The extraction and the valuation proceeded simultaneously, each largely indifferent to the other.

Who Is Not at the Negotiating Table

The settlements reached in late 2025 are easy to read as resolution. Warner settled. UMG settled. Licensing frameworks are being structured. The market, it might seem, is finding its equilibrium.

But the parties at that table were major record labels and venture-backed technology companies. The veteran session musician is not at that table. The independent songwriter with a modest streaming catalog is not at that table. The artist from a country whose recorded music reached global platforms without legal representation is not at that table. When Warner settled with Suno, one element of the deal included Suno acquiring Songkick from Warner — a transaction between two corporations that left the actual creators of the training data in precisely the same position as before: uncompensated for what was taken, unrepresented in the terms of what followed.

The NO FAKES Act, which cleared the Senate Judiciary Committee unanimously in June 2026, signals legislative recognition of a narrower version of this asymmetry — creating a federal right for individuals to control their voice and likeness in digital replicas. It addresses the symptom of voice reproduction without touching the underlying dynamic of training data appropriation at scale. The EU AI Act’s requirement for documentation of consent from rights-holders is more structural, but its enforcement mechanisms in the music industry remain undeveloped. What looks from a distance like market resolution is, on closer inspection, a negotiation among institutions with substantial power who share an interest in not setting too precise a precedent.

Who bears the cost of imprecise precedent? Not the institutions at the table.

What Would Make This Wrong

The argument presented here has a weak point, and intellectual honesty requires naming it.

I have treated the displacement of professional musicians as both real and attributable to AI training practices. But the causal chain is genuinely contested. Streaming already decimated recorded music income before AI music generation existed at meaningful scale. Musicians have always faced substitution from cheaper alternatives. It is possible — not certain, but possible — that AI music will expand the total market for audio in ways that eventually create new forms of income for human creators, much as streaming created new listening occasions even while reducing per-play revenue.

The argument would also require revision if compensation frameworks developed quickly enough and broadly enough to provide genuine benefit-sharing — not the narrow label deals of late 2025, but systems that reach independent artists and provide ongoing compensation proportional to how much their work shaped model outputs. Influence-based attribution remains nascent and unproven at scale. But if it matured into something functional, the ethics of the situation would shift meaningfully. The ethical case against AI music generation as currently practiced is not a case against the technology itself. It is a case against the method of its construction and the distribution of its benefits — and both of those things are, in principle, changeable.

The Question That Remains

The machines that came before did not ask permission either. But they also could not replicate what they displaced with the completeness that AI music systems can now achieve. If we accept that consent matters when creative work is used to build something that competes directly with its source, then the negotiation now underway between labels and technology companies is addressing the consequences of an action already completed — with the people most affected largely absent from the terms.

What does it mean to settle a debt when the people who are owed were not consulted about what it was worth?