KOHEN AVOCATS · PARIS, FRANCE
Grammys 2027: can AI-assisted music win a Grammy, and who owns the copyright

Hassan Kohen is a Paris-based lawyer admitted to the Paris Bar (Avocat au Barreau de Paris). Initial strategic analysis: €80 incl. VAT, personal reply within 24 hours.

100% confidential · Legal professional privilege · No obligation

Article generated by artificial intelligence, through a process designed and supervised by the firm

Sources: U.S. court opinions and official U.S. government publications, read for this article

Maître Hassan KOHEN, avocat au Barreau de Paris
Hassan KOHEN
Attorney at the Paris Bar

Grammys 2027: can AI-assisted music win a Grammy, and who owns the copyright

Speak to a lawyer 06 89 11 34 45

With Grammy nominations week putting new music on every screen, American songwriters, producers, and bedroom beatmakers face a practical question that the hype cycle never answers: if artificial intelligence helped make your song, do you actually own it? The Recording Academy’s members already use generative AI as a tool for song ideation, developing beats, and mixing tracks, according to comments the Academy and Universal Music Group submitted to the U.S. Copyright Office, summarized in the Office’s January 2025 report on AI and copyrightability (Copyright and Artificial Intelligence, Part 2). That report, the Office’s 2023 registration guidance, and three federal court decisions handed down between 2025 and 2026 now draw a usable map. The short version is encouraging but exacting: federal law protects what a human being actually authored, even inside an AI-assisted song, and it protects nothing that a machine produced on its own. A second line of new law answers the mirror question: when another company’s AI trains on your released music, courts decide fair use case by case, and verbatim commercial copying for a competing product just lost in a federal appeals court. This article walks through both lines, with the registrations, writings, and checks that follow from each.

The timing is not accidental. With the Academy announcing that the 2027 nominees will be unveiled on November 16, American search interest has swung toward timing questions such as when the ceremony takes place, while working musicians ask a harder question underneath: which uses of generative tools keep a song protectible, and which ones quietly surrender it. Federal law answers that harder question with unusual clarity this year, because the D.C. Circuit, a Pennsylvania federal court, and the Third Circuit each ruled on AI and copyright within the last eighteen months.

I. Can an AI-assisted song be copyrighted and registered?

Federal copyright law gives songwriters more room than the headlines suggest, and less than AI marketing implies. The discussion that follows is general in scope: it describes checks useful in comparable situations and makes no findings about Thomson Reuters’s, Ross Intelligence’s, Vedros’s, Sterling Group’s, or Thaler’s situations. Every rule below comes from a federal statute, the Copyright Office, or a federal court decision read for this article, and every rule stated here is federal law, not the law of any single state.

A. A machine cannot be the author, but a human using a machine can own the song

The controlling appellate answer arrived on March 18, 2025. In Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. Mar. 18, 2025) (reported at 130 F.4th 1039), the Court of Appeals for the District of Columbia Circuit decided March 18, 2025, the court considered a visual artwork that a computer system called the Creativity Machine produced on its own, with the system’s owner listed only as owner and the machine listed as sole author. The court affirmed the Copyright Office’s refusal to register the work, holding that “The Creativity Machine cannot be the recognized author of a copyrighted work because the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being.” A songwriter who types a detailed prompt and publishes the raw output owns no copyright in that output, but the same songwriter who writes the lyrics, composes the melody, and uses AI only to polish the mix owns every human-authored note. That distinction is the whole ballgame for Grammy-week creators: authorship follows the human hand, not the tool.

The statute underneath that holding is 17 U.S.C. section 102(a), which provides that copyright protection subsists, in accordance with this title, in original works of authorship fixed in any tangible medium of expression, and which expressly lists musical works among the protected categories (17 U.S.C. § 102(a)). Two words do the heavy lifting. Original means the work owes its origin to a human author and shows at least a spark of creativity; fixed means it is captured in a stable form, from a voice memo to a digital audio workstation session. An AI model is neither an author nor an owner under this section, which is why the Copyright Office’s long-standing practice, restated in Chapter 300 of its Compendium, insists on human authorship as the gateway to registration (Compendium (Third), Chapter 300).

A recent federal decision shows that authorship filter in action. In Laatz v. Zazzle, No. 5:22-cv-04844-BLF (N.D. Cal. May 9, 2025), the United States District Court for the Northern District of California decided May 9, 2025, the court used the Compendium to assess authorship as applied to the work at issue, citing Thaler v. Perlmutter, 130 F.4th 1039, 1042-43, 1047. The court used the Compendium to assess authorship as applied to the work at issue, and that authorship filter is exactly what Grammy-week songwriters must satisfy before registering an AI-assisted track.

That gateway has a paperwork consequence the Office spelled out in March 2023. Under its AI registration guidance, effective March 16, 2023, applicants must be candid about machine-generated material: “AI-generated content that is more than de minimis should be explicitly excluded from the application,” using the limitation-of-claim section, and no applicant should list an AI system or its vendor as an author (88 FR 16190). For a songwriter, the practice is concrete. Register the lyrics you wrote, the melody you composed, and the arrangement choices you made; describe and exclude any extended passage the generator produced untouched. An honest application protects the human share of the song and keeps the registration defensible if anyone ever challenges it.

The Thaler litigation also shows how narrow the losing position was, which is good news for musicians. The applicant insisted the machine itself was the sole author and claimed ownership only as the machine’s owner, first as a work made for hire. The Office denied registration, a federal district court granted summary judgment to the Office, and the court of appeals affirmed on the same ground: the statute demands a human author in the first instance. Nobody in that chain questioned that a human who meaningfully shapes AI output can own the result; the courts simply had no such human authorship before them, because the applicant disclaimed it.

The statute protects two distinct musical objects, and the distinction matters for AI workflows. Section 102(a) lists both musical works, the composition, and sound recordings, the fixed performance, as separate categories (17 U.S.C. § 102(a)). A singer-songwriter therefore holds potentially layered rights: the song as written and the recording as performed and produced. AI can touch either layer without destroying either, provided a human authored the layer being claimed. And section 102(b) draws the outer boundary for prompt enthusiasts: “In no case does copyright protection for an original work of authorship extend to any idea, procedure, process, system, method of operation, concept, principle, or discovery, regardless of the form in which it is described, explained, illustrated, or embodied in such work.” A prompt is an idea or instruction, even a brilliant one. Protection starts where human expression starts.

B. Prompts alone do not make you an author, but selection and arrangement can

The most common misunderstanding of 2026 is that a clever prompt equals authorship. The Copyright Office studied that exact claim and rejected it. Based on currently available technology, the Office concluded that “prompts do not alone provide sufficient control” over generated output, while confirming that human authors own the expression perceptible in AI-assisted output, including creative selection, coordination, or arrangement of material and creative modifications of it (Copyright and Artificial Intelligence, Part 2). Typing a one-line request for a sad country ballad about a truck is an idea, and ideas were never copyrightable. Choosing the third generated chorus, rewriting its bridge, stacking harmonies you voiced yourself, and sequencing the track order of an EP are authorship of the classic kind, exercised through new tools.

Federal courts confirmed that line again in May 2026. In Vedros v. Sterling Group, No. 4:24-cv-02183 (M.D. Pa. May 14, 2026), the United States District Court for the Middle District of Pennsylvania decided May 14, 2026, a photographer sued a company that lifted his photograph onto its commercial website and then invoked artificial-intelligence authorship concepts to argue the photo deserved no protection. The court rejected the argument outright, explaining that Thaler holds that “the human authorship requirement does not prohibit copyrighting work that was made by or with the assistance of artificial intelligence,” 130 F.4th 1039, 1048-50, and entered judgment for the photographer after finding the company’s verbatim commercial copying was not fair use. For musicians, the application is direct: AI assistance in your workflow does not forfeit your copyright, and a commercial outfit that simply republishes your work unchanged cannot borrow AI-authorship talk as a shield.

Two more statutory mechanics complete the picture. First, when you build on AI-generated material the way a producer builds on a sample, section 103 of the Copyright Act limits the new protection to what you added: the copyright in a compilation or derivative work covers only the later author’s contribution, not the preexisting material (17 U.S.C. § 103). Your reharmonization, your edited vocal comp, and your track sequencing are yours; the raw generator output underneath stays outside the claim. Second, registration remains the ticket to federal court. Section 411(a) provides that no civil action for infringement of the copyright in any United States work shall be instituted until preregistration or registration of the copyright claim has been made (17 U.S.C. § 411(a)). Register the human-authored version early, before the dispute, because the courthouse door stays shut until the Office has acted on your application.

The Copyright Office has already applied this framework to a real registration, and the example translates directly to music. In the well-known Zarya of the Dawn matter, described in the January 2025 AI report, the Office registered a comic book’s human-authored text together with the author’s selection, coordination, and arrangement of the written and visual elements, while refusing the Midjourney-generated images standing alone. A songwriter’s equivalent is straightforward: the verses you wrote and the order in which you sequenced eleven tracks can be registered even when a generated texture sits inside track seven, as long as the application claims only the human contributions and excludes the rest.

The exclusion mechanics are deliberately simple. In the Limitation of the Claim section of the application, under Material Excluded, the applicant gives a brief description such as the chorus or artwork generated by artificial intelligence. Vague or missing exclusions create the exact vulnerability a challenger would exploit years later, when memories of the studio session have faded and only the certificate speaks. Section 103(b) reinforces the same modesty from the other side: protection in a new arrangement or compilation reaches only the new matter the author contributed, never the underlying material as such. Claim big and the registration overreaches; claim precisely and each human layer stands on its own foundation.

II. When someone trains AI on your released music

The second question runs in the opposite direction and matters just as much during awards season, when catalogs circulate widely: what happens when an AI company ingests your released songs as training data? Federal law answers with fair use, a case-by-case defense codified at 17 U.S.C. section 107, and a September 2026 federal appellate decision applying it to AI training in the strictest commercial setting (17 U.S.C. § 107).

A. Fair use is decided factor by factor, and direct commercial competition loses

On September 30, 2026, the Third Circuit gave AI developers their clearest appellate loss to date. In Thomson Reuters v. Ross Intelligence, No. 25-2153 (3d Cir. Sept. 30, 2026) (official slip opinion), the Court of Appeals for the Third Circuit decided September 30, 2026, a legal publisher sued an AI startup that had copied thousands of its editorial headnotes to train a competing legal-research tool. The court held that “We therefore hold that Thomson Reuters’s materials are copyrightable and that ROSS’s use was not fair,” affirming partial summary judgment for the publisher. A musician whose released tracks are scraped to train a competing music generator stands in a comparable posture to that publisher: verbatim copying of expressive material to build a direct commercial substitute is the weakest possible fair-use position, and the first factor, purpose and character of the use, weighed heavily against the copier precisely because the startup aspired to compete with the very works it took.

The opinion walks through each of the four statutory factors rather than treating AI as a trump card: purpose and character of the use, nature of the copyrighted work, amount taken, and effect on the market for the original. Creative expression sits at the core of what copyright protects, wholesale copying of that expression for the same functional purpose points against fairness, and a defendant’s commercial motive counts. None of that guarantees any particular songwriter would win a training-data case, because music cases will differ on how much was taken, how transformative the output is, and what licensing markets exist. But the framework is now appellate law from 2026, and it rejects the idea that feeding copyrighted work into a model is automatically excused because the technology is new.

The statute the Third Circuit applied has not changed since 1976, and its text is worth reading whole: “In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include,” first, the purpose and character of the use, including its commercial or nonprofit character; second, the nature of the copyrighted work; third, the amount and substantiality of the portion used; and fourth, the effect of the use upon the potential market for or value of the work. No single factor controls, and courts weigh them together against the facts of each dispute, which is why training-data outcomes will vary with what was copied, how much, for what competing purpose, and with what market harm.

Three cautions keep this holding in proportion. The appeal was interlocutory, reviewing a partial summary judgment rather than a final trial verdict, and it binds federal courts in the Third Circuit while persuading elsewhere. Music-training disputes will present different facts: models that quote thirty seconds in a transformative collage differ from models that memorize entire choruses for on-demand generation, and courts will examine the licensing markets that exist for training data, which are evolving month by month. The decision removes the strongest blanket defense, not the need to prove each case.

B. The checks that protect a catalog before and after release

The Vedros decision doubles as a checklist for enforcement. There, the infringer used the entire photograph for the same decorative purpose on a commercial site, offered no evidence of a different transformative purpose, and showed no meaningful market justification; the court granted the photographer’s infringement motion and denied the fair-use defense as a matter of law. Translate that to music: full-track ingestion for a commercial model, with the resulting tool competing for the same listeners or licensing dollars, replicates the losing pattern almost exactly. Save dated evidence of your releases, keep the project files that show your human contributions, and monitor where your catalog appears, because the factors reward creators who can document originality, commercial purpose of the taking, and market substitution.

Two formalities deserve attention while the song is still yours alone. First, as noted above, register before you need to sue; section 411(a) makes registration the precondition to a federal infringement action, so an unregistered viral hit is leverage you cannot yet spend (17 U.S.C. § 411(a)). Second, put creative splits and transfers in signed writing. Section 204(a) provides that a transfer of copyright ownership, other than by operation of law, is not valid unless an instrument of conveyance, or a note or memorandum of the transfer, is in writing and signed by the owner of the rights conveyed (17 U.S.C. § 204(a)). Handshake beat leases, texted promises of producer points, and unsigned split sheets fail this test. The Thomson Reuters and Vedros rulings bind only the circuits and districts that issued them and turn on their own facts: training-data disputes over music will be litigated work by work, model by model, and licensing market by licensing market.

Collaboration paperwork matters as much as registration paperwork. Most AI-assisted songs involve several humans: the topliner, the producer who ran the model, the session vocalist, the mixer who comped the final. Under section 204(a), handshake allocations of those shares are worthless until someone signs them. A one-page split sheet signed before release, naming each contributor, each share, and who administers the work, converts a future courtroom argument into an administrative task. Beat leases with AI-generation warranties, producer agreements identifying who owns model-assisted stems, and band agreements updated for generative tools all serve the same function: they prove the human chain of title that Thaler demands and that Thomson Reuters rewards.

Documentation habits decide close cases. Keep the digital audio workstation sessions with their edit histories, bounce the intermediate mixes that show which bars you rewrote by hand, and preserve dated messages in which collaborators approve the final comp. In Vedros, the photographer prevailed in part because the taking was total and commercial while his ownership was documented and undisputed; creators who can show the same two things, authorship on their side and wholesale appropriation on the other, start every enforcement conversation ahead. Where a platform offers opt-out or licensing programs for training data, use them in writing and keep the confirmations, since courts examining market harm will ask what licensing avenues existed and who used them.

Conclusion

Grammy week rewards songs that move people, and federal law still reserves its rewards for the people who make them. A machine cannot be listed as your co-writer, prompts alone do not create authorship, and anything more than a minimal machine-generated passage should be excluded from your registration application. What you wrote, selected, arranged, and transformed is yours to register, and registration is what opens the federal courthouse door. On the other side of the trade, companies that copy expressive work wholesale to train competing commercial tools now face a 2026 appellate holding that such use is not fair, decided factor by factor under the same four-part test that will govern music cases. Keep the project files, sign the splits, register early, and read each new decision for what it actually held rather than what the headlines claim.

None of this requires abandoning new instruments. The banjo, the sampler, and the digital workstation each provoked fears that the machine would replace the musician, and each ended as a tool in human hands. The 2025 and 2026 decisions say the same about generative models: bring your judgment, leave your fingerprints on every bar, write down who owns what, and register what you made. Do that, and the law that protected songwriters for two centuries protects this song too.

📄 Official circular