Why AI Tracks Put Music Supervisors in an Impossible Position

By Chris Castle of Music Tech Policy

The Framing Problem

Anyone who lived through the Dot Bomb era is reluctant to be branded a “Luddite” by the same people who were using technology to enrich themselves with one of the great free rides. But the humans who own and direct AI are doing far more than forcing a reluctant industry to embrace a new distribution channel. Unlike streaming, downloads, or social media, generative AI is not merely changing how music is delivered. It is changing who gets paid, who bears the risk, and whether anyone can reliably determine where an AI output came from in the first place.

Platforms like Suno and Udio are asking music supervisors, studios, and brands to accept unprecedented chain-of-title uncertainty in exchange for short-term convenience, while shifting legal, financial, and reputational risk onto the productions that use AI-generated tracks.

The creative community may not understand everything about the major artificial intelligence platforms (aka “hyperscalers” or anyone who buys GPUs from Nvidia and CPUs from Micron) or even everything about the music generating models like Suno and Udio. What we do know is that these models were very likely all trained on stolen goods. Our stolen goods. The risk burden should not be on a music supervisor to catch them if you can; the burden should be on the AI platform to raise the confidence level and trust matrix to a market clearing level. They have, so far, failed in that endeavor.

The theft accusation is not just hyperbole; this has been proven in court, under oath, subject to the crucible of cross examination. It’s been proven of Anthropic and Meta with books, and Udio just admitted to scraping YouTube for music. Does anyone think that the hyperscalers like Anthropic and Meta did it with books but not with music, photographs, or your baby pictures posted on social media? Does anyone think that Udio scraped YouTube but Google did not?

Or have these platforms, who are intensely lobbying for a retroactive safe harbor for their training practices, all been doing more or less the same thing, in more or less the same way at more or less the same time? Massive, intentional theft, which begs the question: If they all are stealing, how is it that they all reached that same conclusion at roughly the same time? And also begs the question “If this isn’t criminal copyright infringement prosecutable by U.S. Attorneys, what is?”

There is a tendency in the AI music debate to frame the piracy issue as a binary conflict between “innovation” and “resistance to technology” that wraps greed in an American flag of “because China”. That’s the cheap shot defensive narrative employed by the hyperscalers (and adjacent data center builders like the Canadian investor Kevin O’Leary). But when it comes to music generative AI, that framing misses the operational reality confronting one of the most practical and risk-sensitive groups in the broader entertainment industry: music supervisors.

For music supervisors, the AI challenge is not only about piracy. It is logistical, contractual, and—perhaps most significantly—what is insurable.

Like music publishers and labels, supervisors are being pitched AI-generated tracks for use in film, television, advertising, games, and streaming productions at an increasing rate. In only some cases, the tracks are openly identified as synthetic. In most others, they are outright misrepresented, or pitched through aliases, shell artist profiles, or generic production libraries with little provenance information attached. The problem is not simply whether the music sounds good or fits the need. The problem is that the existing sync licensing economy was built around identifiable human authorship, chain of title, and copyright ownership. AI-generated music destabilizes each of those assumptions simultaneously.

Why the Sample Analogy Fails

Composers and producers mistakenly think AI tracks are analogous to samples circa 1993, before routine clearance processes solidified. Traditional samples can usually be cleared because there is an identifiable song copyright owner, master owner, and licensing chain. Tracks generated by platforms like Suno are qualitatively different than old-school samples because the underlying training inputs, embedded influences, and potential rights holders may be intentionally obscured and are therefore unknowable without expensive litigation. Supervisors and productions “licensing” an AI track may face infringement risk without any practical mechanism to identify, locate, or clear the affected parties.

Unlike traditional samples, AI outputs may create what amounts to a rights black box. Sample clearance can be difficult, but it is rarely impossible because the relevant owners can usually be identified. With AI-generated tracks, the potentially affected rights holders may be unknowable from the outset, leaving supervisors exposed to claims they cannot practically investigate, clear, or insure against.

That creates an impossible position for supervisors.

The Supervisor’s Role as Risk Allocator

At the front end, supervisors are expected to clear rights and protect productions from downstream claims. Every professional music supervisor understands that a sync placement is not merely a creative decision, it is also a risk allocation exercise.

In addition to getting the music, supervisors have to clear the rights to the music or work with production staff or a licensing department to do so. Productions rely on supervisors to confirm that the music being delivered can actually be licensed, monetized, insured, and defended. AI complicates every step of that process.

Provenance and the Ownership Vacuum

The first issue is provenance. A supervisor receiving a human-created track can usually determine who created it, who owns the master, who controls the publishing, whether samples were used, whether performers consented, and whether the work is registered with a PRO for cue sheets. I’m not saying this is always easy, but it can be done. Even when rights are fragmented, there is still a recognizable rights infrastructure. With synthetic tracks, that infrastructure may not exist at all.

There are no work-arounds. For example, the tenancy-in-common analogy does not resolve the problem. Copyright co-owners can generally grant nonexclusive licenses of an undivided interest, subject to a duty to account (and no economic “waste”). But that regime presupposes an actual copyright estate with identifiable co-owners. AI tracks may instead present an ownership vacuum: contractual access rights from the platform, possible partially human-authored contributions, unresolved third-party infringement claims against the platform (including putative class actions), and no clear undivided copyright estate to license. There is no ascertainable set of fractional owners, and no one to account to.

Plus an often overlooked caveat on TIC law is that a tenant in common may not grant a license that results in substantial destruction or “economic waste” of the common asset. A unilateral AI license may fall within this prohibition where the licensed use irreversibly extracts value from the work that cannot be replenished, impairs existing or future licensing markets, or otherwise diminishes the interests of non-consenting co-tenants.

AI platforms say things like, we just provide statistical inferences and similar vague statements, we leave all these details to you who toil in the vineyard. Then they wait for the reaction like they’ve just said something smart rather than nightmarish. In the world of “tech bros”, maybe they have. But not in a rights-respecting universe.

Training Data and Unresolved Fair Use Claims

Many generative AI systems are trained on enormous datasets of dubious provenance containing scraped reproductions of other people’s copyrights, including recordings and compositions. Several AI companies have publicly acknowledged using copyrighted works in training datasets, often under aggressive theories of the “fair use” affirmative defense that remain unresolved in courts as of this writing. In other words, the AI platform knows they don’t have a license, they merely have an excuse, and there’s so much money involved it is worth it to them to take the chance of being caught.

That means a supervisor may be asked to place a track generated by a system whose underlying legality itself is actively being litigated, often by the very people who might make a claim against the supervisor or her clients.

Adding to the uncertainty, some platforms defend their scraping and training practices by claiming they only scraped “publicly available” content. But “publicly available” is not a recognized legal standard that confers a right to copy, reproduce, or create derivative works. A song streaming on Spotify, a recording posted to YouTube, a composition indexed by a search engine, or a Facebook post all may be publicly accessible in a practical sense that they are not behind a paywall or other barrier, while remaining fully protected by copyright and privacy laws. The phrase functions as rhetorical cover rather than legal justification—and no platform has disclosed its training corpus with sufficient specificity for anyone to verify the claim or assess the resulting infringement exposure with any degree of legal certainty.

For example, as Complete Music Update has reported, Mikey Shulman, the CEO of Suno, has stated that the music industry’s resistance to AI innovation stems from a “fixed pie mentality” (whatever that means) while he simultaneously admits to using copyright-protected music in his company’s AI training data—a practice he describes as “stock standard” and one that “every AI company does.”

Even if the generated output itself does not directly infringe an immediately identifiable work, a music supervisor is still left asking difficult questions about ownership, provenance, infringement exposure, E&O insurance, indemnity, and downstream liability. Often these are the same questions that are being litigated in some of the largest copyright infringement cases of all time that have revealed in discovery just how massive the infringement really is “under the hood.”

The risk compounds when supervisors are pitched an AI-generated track that has previously been rejected or removed by a streaming platform for copyright or other violations—often unbeknownst to the supervisor. That takedown creates a documentable paper trail. If the same track or one generated by the same model under similar conditions is subsequently placed in a production, the prior removal may constitute constructive notice of infringement risk, potentially converting what might otherwise be innocent infringement into willful exposure. Just saying.

Production Music Disruption

The production music library business may face particularly acute disruption from generative AI. Historically, libraries created value by investing in composers, musicians, metadata, curation, and rights administration while offering supervisors a relatively reliable and easy-to-clear chain of title. AI threatens to flood the market with low-cost alternatives that mimic many of those functions without providing comparable certainty regarding provenance, ownership, or infringement exposure. Ironically, the more copyright and E&O risk associated with AI-generated tracks, the more valuable traditional production libraries may become. In a market increasingly saturated with uncertainty, a reliable catalog with rights cleared, identifiable authorship, and insurable chain of title may become premium products rather than commodities.

Podcasts may present a special use case. While clearance practices in the podcast industry can be informal, that informality often disappears once a podcast becomes successful. As audiences, advertising revenue, licensing opportunities, and acquisition interest increase, copyright owners frequently revisit past uses and demand explanation if not compensation. At that point, provenance, chain of title, and documentation become critical. A producer who relied on an AI-generated track with uncertain origins may find it difficult—or impossible—to demonstrate that all necessary rights were obtained, creating legal and business risks precisely when the podcast has become most valuable. Good thing that’s never happened before.

Errors & Omissions Insurance

Most film and television productions must deliver extensive rights documentation to distributors, studios, broadcasters, streamers, and financiers. Delivery packages frequently require chain-of-title records, licenses, cue sheets, publishing splits, performer consents, and representations that the production does not infringe third-party rights. A supervisor who knowingly places AI-generated music may not be able to provide reliable answers to any of those questions.

That uncertainty becomes especially serious in connection with Errors & Omissions insurance which is one of the cornerstones of risk allocation in the world of third-party rights. AI clearance has been a topic in E&O underwriting circles for several years, and underwriters are likely to recognize exactly what is being asked and to have already incorporated AI-specific clearance procedures into their underwriting questionnaires. Based on current market practice, nearly all such procedures present significant obstacles to placement except in very narrow circumstances if at all. (And since delivering E&O coverage may be a delivery obligation of the producer, failing to get bound may also result in failing to get a delivery installment of a minimum guarantee or license fee.)

A workable E&O solution would likely require a bespoke AI model built specifically for rights-cleared commercial production use. If what goes in is already cleared for the bespoke AI model, what comes out can be relied upon, at least theoretically. In practice, that means a closed training corpus consisting entirely of owned, licensed, or public-domain recordings and compositions, combined with provenance tracking, output similarity testing, and contractual indemnities.

It is not enough that the model is closed—one must be able to prove it is closed, both to the E&O carrier and in court if challenged. And one way to prove a particular recording is not an output is to be able to prove it was never an input, which requires rigorous attention to provenance and training.

But that creates a difficult economic question. The commercial attraction of generative AI has largely been scale: ingesting vast libraries of existing culture to generate outputs with broad stylistic range, fine-tuned by developers and users. Once the model is limited to a carefully licensed corpus, the system becomes much smaller, more expensive, and creatively constrained.

And even then, copyright problems may persist. Newly generated outputs could still raise derivative-work, substantial-similarity, style imitation or ownership disputes. So the industry may ultimately discover that a fully cleared AI music ecosystem is technically possible but commercially unattractive compared to traditional commissioned humans writing music and existing licensing markets of pre-AI works.

Because you could just hire a composer and musicians. Problem solved. There’s a thought.

Terms of Service as Quitclaim

It bears emphasis that the following analysis focuses on Suno’s Terms of Service, but each platform presents its own contractual framework. According to press reports, Udio and possibly others have evidently agreed to “walled garden” licensing arrangements with major rights holders that may severely restrict what users can do with tracks created on those platforms. The terms vary significantly across services, and each must be independently reviewed.

E&O insurers evaluate and price legal risk. Traditionally, music clearance risk has been relatively quantifiable. But AI outputs create novel categories of uncertainty that insurers do not yet fully know how to evaluate, much less underwrite.

One of those categories is what I call “TOS Risk.” AI platforms frequently introduce this problem contractually through poorly drafted click-through Terms of Service or “TOS”. Most users don’t read the lenghty TOS but underwriters do or may ask questions about it of production counsel or executives, and indirectly of supervisors on the front lines. This is where the Terms of Service structure begins to resemble something closer to a quitclaim than a traditional copyright assignment—the platform conveys whatever interest it may have, without warranting that any particular interest actually exists.

For example, companies like Suno state that, to the extent they own rights in generated outputs, those rights may be retained by Suno or “assigned” to users depending on the subscription tier the user has paid for. The Suno FAQs categorically state that “If you make music using the Basic (free) plan, Suno is the owner of the songs.” No explanation, no option to acquire rights, no discussion of how Suno can “own” anything the user creates, no question of what if the user already had a publishing or record deal, just the bald assertion.

The Suno TOS continues to state that “If you make songs while subscribed to the Pro or Premier plan, you own the songs. Further, you are granted a commercial use license to monetize the songs.” No explanation of the distinction, no explanation of the word “songs” (do they mean recordings, songs, both?) and no explanation of who is granting these commercial use rights.

But the same Terms of Service simultaneously disclaim guarantees whether copyright protection actually exists in the output at all. In practical terms, the platform is effectively saying: whatever rights we may have, if any, we assign to you—while reserving broad limitations, disclaimers, and platform controls. But only if you pay for a higher tier of service on the service we built by scraping the Internet.

The Authorship Paradox

Suno’s own language underscores the instability of the arrangement. Suno expressly distinguishes between “ownership” and “copyrights,” stating that a user may “own” songs generated under a paid plan while simultaneously warning that the material “may not be eligible for copyright protection.” Even more strikingly, Suno states that “writing the prompt does not constitute the creation of the song.” Which immediately draws the question, says who?

Those statements from Suno create profound implications for sync licensing and downstream rights administration. If writing the prompt does not create authorship in the lyrics or composition, then on what legal basis does the prompt create a durable property claim to the broader musical output? The platform appears to be conceding that prompting alone may not satisfy the human authorship requirements traditionally necessary for copyright protection—as reflected in the U.S. Copyright Office’s guidance requiring more than de minimis human creative control—while simultaneously attempting to create a separate contractual ownership framework through Terms of Service.

That leaves supervisors in an extremely uncomfortable position. A production may receive a commercial use license and a platform-level assurance of “ownership” but without certainty that any enforceable copyright or other property right exists underneath the transaction? What does that even mean? The resulting structure looks less like traditional copyright ownership and more like a contractual allocation of access and monetization rights layered on top of legally uncertain outputs—a radically different proposition from a conventional sync license involving identifiable human authorship, customary representations and warranties, indemnity, and a stable chain of title.

The Collective Rights Administration Risk

AI-generated tracks may create another second-order problem that also receives far less attention than copyright infringement: the integrity of the metadata systems on which performance royalties depend. Performing rights organizations distribute billions of dollars annually based on cue sheets, work registrations, writer information, publisher claims, and other metadata submitted by rightsholders and productions—and music supervisors. That system assumes that the parties listed on a cue sheet actually possess identifiable ownership interests in the underlying work.

With AI-generated tracks, that assumption may break down. If the authorship, ownership, or chain of title for a work cannot be independently verified, a production may nevertheless submit the track to a PRO for cue-sheet reporting. Once entered into the rights-management network, the track can generate performance royalties, be matched to registrations, and participate in royalty distributions despite unresolved questions regarding who, if anyone, possesses valid rights in the underlying composition or recording. (This is also true of streaming mechanicals and the Copyright Royalty Judges don’t seem very interested in addressing the issue.)

The resulting governance problems could be significant. Competing ownership claims may emerge years after distribution. The obvious problem is that AI platforms, users, publishers, performers, or previously unidentified rightsholders may assert conflicting interests in the same work. Royalty payments may have already been distributed and spent, forcing PROs into expensive disputes, reversals, indemnity claims, and administrative investigations. Because PRO databases were designed to resolve competing claims among identifiable human creators and publishers—not to determine whether a work generated by a machine possesses a valid ownership claim at all—AI-generated tracks risk introducing uncertainty at the very foundation of the metadata system.

At scale, the problem becomes systemic. A single disputed work can usually be managed. Tens of thousands of AI-generated tracks with uncertain provenance, ambiguous authorship, and conflicting ownership theories could create a metadata governance crisis, undermining confidence in cue-sheet reporting, royalty allocation, and the accuracy of the databases on which collective rights administration depends. In an industry already struggling with unmatched works and incomplete metadata, introducing large volumes of commercially exploited tracks with unverifiable ownership information risks compounding existing problems rather than solving them.

The one thing you can count on is that the hyperscalers could care less.

The copyright system has historically assumed that metadata errors are accidental. Generative AI introduces the possibility that the metadata itself may be fundamentally unknowable. Or even intentional in a statistically significant number of cases.

That’s a different category of problem than missing songwriter splits or administrative mistakes. It’s a governance problem at the level of the rights system itself.

Reasonable Reliance on Platform Assurances

Users might consider whether the platform’s affirmative ownership assurances support a reasonable reliance claim. However, such a claim faces a fundamental obstacle: the same Terms of Service that promise “ownership” simultaneously disclaim that copyright protection exists in the output at all. That internal contradiction may support an argument that ambiguity is construed against the drafter (contra proferentem) in consumer-friendly jurisdictions, particularly where ownership language appears prominently while disclaimers are buried in dense TOS click-through provisions. But for sophisticated commercial users like music supervisors—who are expected to evaluate IP risk independently—a court is far less likely to find that reliance on platform-level assurances was objectively reasonable. And do you really want to hang your hat on that one?

The Practical Problem

The result is that supervisors and productions may—at best—find themselves holding a conditional bundle of contractual rights defined by an AI platform that has no interest in solving the very downstream clearance problems the platform created. Larger ownership questions are likely to remain unresolved for years as litigation involving AI training datasets, copyrightability, derivative infringement, and artist claims continues to wind through discovery, appeals, and potential settlements. Eventually we will get around to litigating the TOS.

For productions operating under delivery deadlines requiring fast turnaround on clearances (which is usually all of them), indemnity obligations, and E&O requirements, that uncertainty is not theoretical. It becomes an immediate practical problem. Music supervisors are not technology regulators. They are professionals tasked with getting projects delivered safely and efficiently. But the current AI environment increasingly asks them to absorb unresolved legal risk that properly belongs upstream—with AI developers, platforms, and investors.

Remember the inputs/outputs issues. An AI-generated output may not sound substantially similar to any identifiable song, yet the training inputs used to create that output may themselves have been copied or exploited without authorization. As a result, the legal risk may arise not from the finished track alone, but from the provenance and lawfulness of the materials used to build the underlying model.

There is a deeper irony here. Suno CEO Mikey Shulman has argued that AI will “democratize” music creation by lowering barriers to entry and expanding access to creative tools—which is about 10% of the story. Yet many of the most important commercial music markets such as film, television, advertising, games, and other professionally supervised productions depend upon verifiable ownership, enforceable chain of title, insurable rights, and predictable licensing relationships—the other 90% of the story. Mikey Schulman’s product not only does not help with that 90%, I would go so far as to say Suno offers false hope to its users and embeds a core issue in its outputs.

If generative AI outputs cannot satisfy those requirements, supervisors, insurers, distributors, and production counsel may increasingly gravitate toward a smaller universe of vendors capable of providing reliable provenance and rights assurances. The result could be the opposite of democratization: a market in which direct access to reliable high-value commercial placements becomes concentrated among established composers, catalogs, and vendors who are able to demonstrate verifiable ownership rather than merely offering low-cost generation tools.

That is not resistance to innovation. It is basic professional risk management that results in hiring composers and musicians. Or should.

AI Music Clearance Checklist for Supervisors

The following checklist is derived from the risk factors identified above. It is intended as a practical starting point for music supervisors or underwriters evaluating whether an AI-generated track can be safely placed in a production and is not intended as legal advice

Provenance and Authorship

Can you identify one or more natural persons (humans) who exercised creative control over the track sufficient to constitute human authorship?
Is the generative AI platform and model used to create the track identified and documented?
Can you determine what training data the model was built on, and whether it included copyrighted works? (For products like Suno, this answer is likely “no”)
Is the training corpus closed (consisting entirely of owned, licensed, or public-domain material), and can that be independently verified?
Has the track been run through similarity testing against existing copyrighted recordings and compositions?
Has the platform claimed its training data was merely “publicly available,” and if so, has anyone verified that claim with sufficient specificity to assess infringement exposure?

Ownership and Chain of Title

Does the licensor claim copyright ownership in the track, or only contractual/platform-level “ownership”?
Is the track eligible for copyright registration, and has it been registered or is registration pending?
Is there a complete chain of title from creation to the party granting the sync license?
Are there any unresolved third-party claims, pending litigation, or competing ownership assertions affecting the track or the platform?

Licensing Terms and Indemnity

Does the license grant include customary representations and warranties regarding non-infringement of third-party rights?
Does the licensor provide a contractual indemnity covering infringement claims arising from the track’s AI-generated elements or training data?
Is the indemnifying party financially capable of standing behind the indemnity if a claim arises?
Are the platform’s Terms of Service consistent with the specific license being granted, or do ToS disclaimers limit or contradict the license terms?

Insurance and Delivery

Has the production’s E&O insurer been notified that AI-generated music may be used, and has the insurer confirmed coverage?
Can you satisfy the AI-specific clearance procedures in the E&O underwriting questionnaire?
Can you provide complete chain-of-title documentation, cue sheets, publishing splits, and performer consents for the delivery package?
Can you make the standard representation to distributors, studios, and financiers that the production does not infringe third-party rights?

Risk Assessment

Is the AI platform or its training data the subject of pending or threatened litigation that could affect the track’s legal status?
Has the track (or similar output from the same model) previously been rejected, taken down, or flagged by a streaming platform or distributor on copyright grounds?
If the user is relying on the platform’s ToS ownership assurances, has counsel evaluated whether such reliance would be deemed objectively reasonable for a sophisticated commercial licensee?
If any of the above items cannot be affirmatively resolved, has the production been advised of the residual risk in writing?
Has a determination been made as to whether the unresolved risk should be borne by the production, the AI platform, or another upstream party?

Why AI Tracks Put Music Supervisors in an Impossible Position