Is making AI music legal?

The Three Legal Questions Behind Every AI Track

Ask ten people whether AI music is legal and you'll get ten different answers — all of them partially right. That's because "is AI music legal" isn't one question at all. It splits into three distinct legal questions, each with its own body of law, its own ongoing litigation, and its own shifting answer.

The stakes for getting this wrong are enormous. Copyright owners who sue over AI music can seek statutory damages of up to $150,000 per infringed work under 17 U.S.C. §504(c)(2), plus $2,500 per act of circumvention under the DMCA. And copyright owners are filing suits at an accelerating pace: more than 70 infringement lawsuits against AI companies were on file by the end of 2025, up from roughly 30 a year earlier, according to the Copyright Alliance's case tracker.

Here's how the three questions break down:

• Ownership: Can you actually own the music you generate? The U.S. Copyright Office refuses registration for purely AI-generated works, and courts affirmed the human-authorship requirement in Thaler v. Perlmutter. Human-assisted works may still qualify — but no bright-line test exists for how much human involvement is enough.
• Training: Was the tool itself built legally? This is where the Suno and Udio lawsuits live, and where outcomes diverge dramatically depending on how training data was acquired — legally purchased versus pirated.
• Distribution: Can you release what you made? Platform rules, voice-cloning and right-of-publicity claims, and streaming-fraud policies increasingly govern whether an AI track can reach listeners at all.

The Anthropic case illustrates why the training question matters so much. In Bartz v. Anthropic, the court found that training on legally purchased books was fair use — but training on more than 7 million pirated books was not, resulting in a $1.5 billion settlement, the largest copyright settlement in U.S. history. How data was acquired turned out to matter more than whether training happened at all.

At SkillzBench, we see the same pattern in the AI skills ecosystem: provenance and licensing determine risk, not the technology itself. That's why we're building license-checking and security scanning into our verification pillar — because "is it legal" increasingly starts with "where did it come from."

The bottom line for musicians: making AI music isn't illegal per se. But using unlicensed tools trained on copyrighted works carries substantial financial exposure, and the market is rapidly moving toward licensed models as the safe path forward.

Who Owns AI Music? The Human Authorship Problem

You can spend hours crafting the perfect AI-generated track and still walk away with no copyright at all. Under current U.S. law, ownership hinges on a single question: was a human the author?

The U.S. Copyright Office refuses registration for purely AI-generated works, requiring human authorship before a copyright attaches. That position got a judicial stamp of approval in Thaler v. Perlmutter, when the D.C. Circuit affirmed in March 2025 that the author must be human — "not the machine itself." Crucially, though, the court noted the human authorship requirement "does not prohibit copyrighting work that was made by or with the assistance of artificial intelligence," according to Holland & Knight's legal analysis.

That distinction — AI assistance versus AI generation — is where most creators now live. The Copyright Office laid out its position in a multi-part AI report, covering digital replicas in Part 1 (July 2024), copyrightability of generative outputs in Part 2 (January 2025), and AI training in Part 3 (May 2025), after receiving more than 10,000 public comments on its AI inquiry.

Here's the uncomfortable part: there is no bright-line test for how much human involvement is enough. No court has defined the threshold, which means two musicians using the same tool could get opposite answers about whether their work is protectable. The Copyright Office evaluates claims case by case, looking at what the human actually contributed versus what the machine produced.

For working musicians, that ambiguity makes documentation your best defense:

• Keep drafts, prompts, and revision histories that show your creative decisions over time.
• Record which elements you wrote, arranged, or performed yourself — lyrics, melody, instrumentation.
• Note where AI output was modified, rejected, or substantially reworked.
• Preserve stems and project files that demonstrate human arrangement choices.

The stakes go beyond registration. With major labels suing AI music companies and statutory damages reaching up to $150,000 per infringed work, proving your human authorship also helps establish your own original contribution if a dispute ever arises. And as platforms crack down — Deezer reports up to 85% of streams on fully AI-generated content are flagged as fraudulent — being able to show genuine human creativity increasingly affects whether your music gets monetized at all.

The practical takeaway is simple: treat AI as an instrument, not a composer. If you want to own your music, be able to prove you made it.

That verification-first mindset applies beyond music, too. At SkillzBench, we're building tools that benchmark and validate AI agent skills the same way — because "the AI did it" is never a substitute for knowing exactly what contribution came from where.

From Lawsuits to Licensing: How the Suno and Udio Battles Are Rewriting the Rules

In June 2024, the three major labels — Universal, Sony, and Warner — sued AI music startups Suno and Udio for copyright infringement, and the industry braced for a courtroom war that would define generative music for a decade. Instead, the war largely became a negotiation table. By late 2025, Universal had settled with Udio (October 29, 2025) and Warner settled with both Udio and Suno in November, marking a decisive pivot from litigation to licensing that continues to reshape who can legally build — and use — AI music tools.

The terms of the Warner-Suno settlement matter for every creator. Under the deal, Suno agreed to phase out its current unlicensed models in favor of licensed versions, with artists controlling use of their "names, images, likenesses, voices, and compositions." Suno's first licensed models arrived in September 2026, according to LA Times reporting — though the rollout lagged the settlement by 114 days, and no company has disclosed artist payout amounts or how much control individual musicians actually exercise.

Not everyone settled. As of April 2026, Sony Music continues litigating against both platforms, filing a second lawsuit against Udio covering more than 30,000 recordings, and joining Universal in a new suit alleging Suno's latest model still infringes. Suno's licensing talks with UMG and Sony remain in limbo — "We have ongoing engagement, but there is no path forward with the current proposal," the company said. Independent artists have also organized separately, filing class actions of their own.

The most consequential legal lesson comes from outside music entirely. In Bartz v. Anthropic, Judge Alsup ruled in June 2025 that training AI on legally purchased books qualified as fair use — but training on over 7 million pirated books did not. Anthropic still paid $1.5 billion to settle the pirated-data claims, the largest copyright settlement in U.S. history. The takeaway: provenance of training data matters more than the act of training itself. Plaintiffs in the Suno case applied that logic directly, amending their complaint in September 2025 to allege the platform stream-ripped YouTube content, adding DMCA circumvention claims worth $2,500 per act on top of statutory damages up to $150,000 per infringed work.

For creators and platforms choosing AI music tools today, the practical risk picture is clear:

• Unlicensed tools carry exposure of up to $150,000 per infringed work, plus $2,500 per DMCA circumvention act.
• Licensed models (like Suno's post-settlement releases) shift that risk to the platform, though artist-consent questions remain contested.
• Sony's ongoing suits mean even settled platforms face continued legal uncertainty.
• Streaming platforms add their own enforcement layer — Deezer demonetized up to 85% of streams on AI-generated content flagged as fraudulent.

This is why verifying where a tool's training data came from is becoming as important as evaluating its output — the same "build first, clear later" risk that legal analysts warn is getting costlier. It's a philosophy SkillzBench applies to AI tooling generally: independent verification of what's under the hood, before you rely on what comes out of it. As entertainment attorney Jane Davidson put it, the labels settled because "they think they're going to profit more from licensing music to them rather than from the outcome of a lawsuit" — which means licensed AI music is now the industry's default future, even as Sony's courtroom fight decides how expensive the unlicensed past will be.

Distribution Risks: Voice Cloning, Streaming Fraud, and Platform Rules

Creating AI music in the studio is one legal question; releasing it into the world is another. Distribution is where creators collide with voice-cloning liability, platform fraud detection, and a patchwork of international rules that can make the same track legal in one country and infringing in another.

Voice cloning is the sharpest exposure. U.S. law doesn't clearly protect a performer's voice at the federal level, but the U.S. Copyright Office's digital replica report recommended a federal digital replica law, and class actions like Jason Isbell et al. v. Suno signal growing liability for outputs that evoke real artists without consent (LA Times reporting). Under the Warner-Suno settlement, artists will control use of their names, likenesses, voices, and compositions — a recognition that identity rights are now a core part of AI music deals.

Streaming platforms are policing AI content themselves. Deezer reports roughly 60,000 fully AI-generated tracks arriving daily — about 39% of all deliveries — with up to 85% of streams on that content found to be fraudulent, demonetized, and stripped from the royalty pool (Music Business Worldwide). Even where no law is broken, platform fraud policies can erase revenue entirely, making distribution rules a business risk as much as a legal one.

Jurisdiction matters just as much as platform policy. Key divergences include:

• In GEMA v. OpenAI (Nov 2025), the Munich Regional Court ruled for GEMA — the first worldwide copyright ruling against OpenAI — and a Munich court later ruled against Suno in a separate GEMA case.
• UK and EU policymakers are exploring transparency obligations and collective licensing models for AI training data.
• In the U.S., statutory damages can reach $150,000 per infringed work, plus $2,500 per act of circumvention under DMCA §1201.
• Sony continues litigating against both Suno and Udio, including a second Udio suit covering 30,000+ recordings — so U.S. rules remain unsettled.

The practical takeaway: legality is jurisdiction-dependent and platform-dependent. A track that clears one country's rules may fail another's, and a fully lawful upload can still be demonetized if a platform flags it as fraudulent AI content. For teams evaluating AI tools and skills more broadly — including the agent skills SkillzBench is building to verify, license-check, and security-scan — provenance is the same underlying question: know where the inputs came from before you ship the output. As legal analysts put it, "build first, clear later" is becoming a riskier path as regulation and litigation catch up.

Your Compliance Playbook: Five Steps to Make AI Music Safely

Knowing the rules is one thing; building a workflow that keeps you out of court is another. The good news is that a handful of disciplined habits covers most of the legal exposure in AI music today.

Step 1: Choose licensed tools. With Suno phasing out unlicensed models under its Warner settlement and Sony still litigating against both Suno and Udio, training-data provenance now defines your risk. Unlicensed tools carry exposure of up to $150,000 per infringed work in statutory damages, plus $2,500 per act of circumvention under the DMCA.

Step 2: Document human authorship. The Copyright Office refuses registration for purely AI-generated works, and the D.C. Circuit affirmed in Thaler v. Perlmutter that the author must be human. Keep records of your creative contributions — lyrics you wrote, arrangements you chose, takes you edited — because no court has yet defined how much human input suffices.

Step 3: Never clone voices. Right-of-publicity class actions and the Copyright Office's digital replica recommendations signal growing liability for outputs that evoke real artists without consent. The safest rule is simple: don't generate anything that sounds like a named artist.

Step 4: Check platform policies before you upload. Deezer reported roughly 60,000 fully AI-generated tracks per day — about 39% of daily deliveries — and found up to 85% of streams on that content fraudulent, demonetized and removed from royalty pools. Distribution risk now lives in platform fraud policies, not just copyright law.

Step 5: Monitor the litigation. The ground is still moving, and fast:

• Sony's second Udio suit covers 30,000+ recordings, and new suits allege Suno's latest model still infringes.
• German courts ruled against both OpenAI and Suno in GEMA cases, so the answer differs by jurisdiction.
• Over 70 infringement lawsuits were filed against AI companies by end of 2025, up from about 30 a year earlier.

Underlying all five steps is a trust-verification mindset: don't assume a tool is safe because it's popular, verify it. That's the same philosophy SkillzBench applies to AI agent skills — security-scanning, license-checking, and version-pinning every skill before it reaches an agent, rather than taking a vendor's word for it. Whether you're picking a music generator or an agent skill, the discipline is identical: prove the tool's provenance before you build on it.