Style imitation vs person imitation
Copying a style that reproduces no protected work is largely lawful in the US. Replicating a specific person's face or voice is where liability starts.
"In the style of" is the most-asked legal question in AI content, and the answer is mostly reassuring in a way that surprises people. Style itself is not the thing copyright protects. Copyright protects works. In the United States, imitation of a style that reproduces no protected work remains largely lawful.
The mistake is stopping there. The dividing line that matters is not between "inspired by" and "copied from." It is between imitating a style and replicating a person, and those two run on completely different legal machinery. One is a copyright question with a fairly forgiving answer. The other is a personal-rights question with a much sharper one.
This is general information, not legal advice.
Two different systems, not two points on one scale
| Style imitation | Person imitation | |
|---|---|---|
| What is at issue | Whether protected expression was reproduced | Whether a specific individual's identity was used |
| Governing law | Federal copyright, uniform nationwide | State right of publicity, ~50 divergent regimes |
| Who can complain | The rightsholder in a specific work | The individual, or their estate |
| The test | Does the output reproduce protectable expression from a work | Is the person recognisable |
| Typical outcome for a clean case | Generally lawful | Generally requires consent |
Read that right-hand column carefully. There is no federal right of publicity in the US, so "person imitation" resolves under state law that varies by state. Tennessee's ELVIS Act made voice a protected personal right from 1 July 2024. California's AB 2602 and AB 1836 took effect 1 January 2025 covering performer contracts and deceased personalities respectively. The voice and likeness law overview maps the current pieces.
The practical upshot: a legal review that only asks "did we copy anything" answers half the question, and it is the easier half.
Where style stops being style
Three specific ways a style prompt drifts into something else.
A protected character. This is the most common failure and it is not a grey area. A style can be evoked without reproducing anything. A character cannot — the character is the protected expression. Disney, Universal and Warner Bros. filed against Midjourney over outputs depicting protected characters; the cases were consolidated in the Central District of California on 4 November 2025 and are in discovery, with expert disclosures scheduled for October 2026. Nothing has been decided. What is worth taking from it now is where the plaintiffs aimed: not at the training corpus, at the outputs. That is the user-facing end.
Logos, trade dress and packaging. A brand's visual identity is protected by trademark, which has nothing to do with copyright and does not care whether you copied a work. Generated packaging with a recognisable mark on it is a trademark problem even when everything about the render is original. Our note on legal copy and barcodes on generated packaging covers the adjacent version of this.
A soundalike that is the person. A style of singing is a style. A voice that listeners identify as a specific artist is that artist's voice, and under a statute like the ELVIS Act it does not matter that you never used their name or their recording. This is the sharpest edge on the whole map because it looks like style and is not.
The music case is worth understanding precisely
Music is where the two systems get confused most often, because the copyright clearance apparatus is so visible that people assume it covers everything.
It does not. Content ID matches recordings and compositions. It does not match style. A track that reproduces neither a protected recording nor a protected composition generally passes through it — which tells you exactly nothing about whether the voice on it was cleared. Voice imitation is a personal-rights question and lives outside the matching system entirely. YouTube's synthetic singing voice detection reaches toward that gap, but it is a platform tool rather than a legal test, and it should not be treated as comprehensive.
There is also no merits ruling anywhere on AI music training. UMG settled with Udio in October 2025, and Warner settled with both Udio and Suno in November 2025, each paired with forward licences for opt-in artist catalogues. Sony has settled with neither, and UMG v. Suno continues with Suno defending on fair use. A settlement is a commercial agreement, not a holding — the reason a signed label deal is not a cleared model is that the licensed product and the shipping product are frequently different things.
Lawful is not the same as distributable
Even a clean style imitation can be economically penalised, and this is the part that catches teams who cleared the legal question and stopped.
Deezer tags AI-generated tracks and excludes them from editorial and algorithmic playlists. That is a distribution outcome applied to content that broke no law. Spotify adopted DDEX AI-disclosure fields in credits, announced September 2025, and states that disclosure does not affect royalties or recommendations — but it bans unauthorised voice clones and impersonation outright on its own terms.
Two platforms, two different economics, both operating independently of what any statute says. Platform policy is contractual, it moves faster than legislation, and it is the thing that actually determines whether the work reaches anyone. Musicians planning around this will find the practical framing in our guide for musicians.
Prompt and pipeline hygiene that keeps you on the safe side
None of this argues against style work. It argues for building it out of the right inputs.
- Describe attributes, not names. "Flat vector illustration, four-colour risograph palette, heavy grain, 1970s technical-manual feel" is a style specification. A living artist's name in the prompt is a shortcut that imports their entire body of work as the reference, and it is the input a plaintiff will point at. Style keywords for AI image prompts covers the attribute vocabulary.
- Use your own references. A reference image you own or licensed anchors the look without invoking anyone. This is also how you get repeatability, which name-dropping does not give you.
- Prefer a fine-tune on owned material for a durable look. A LoRA trained on material you own is both the cleanest provenance story and the most stable output. Style references vs fine-tunes covers where the crossover sits.
- Separate "synthetic person" from "replica of a person" in your brief. A fully synthetic individual raises disclosure duties but no publicity claim, because there is no rights holder. A replica of a real person raises both. Keep those categories distinct at brief stage rather than discovering the difference in review.
- Review outputs for characters and marks, not just for style. The failure mode is a generated frame with a recognisable character or logo in the background of an otherwise original shot.
There is one more consequence worth knowing. Under US law, prompts alone — however detailed — do not make you the author of an output. Human-authored selection, arrangement and modification are protectable. So a pure style-prompt generation may be lawful to make and simultaneously weak as an asset you can stop anyone else from copying. If you need a defensible brand look, the fine-tune-plus-editing path gives you both a cleaner provenance story and something with actual human authorship in it. Style transfer for a brand look and the copyright and safety guide cover the two halves.
FAQ
Can I prompt "in the style of" a living artist?
Legally, style itself is not what copyright protects, and an output that reproduces no protected work is generally not infringement. That said, naming a living artist is the highest-risk phrasing available: it pulls their specific works toward the output, it is preserved in your prompt logs, and it invites a claim you would rather not defend. Describe the attributes instead. You lose nothing and gain repeatability.
Is a soundalike singer legal?
A style of singing is not owned by anyone. A voice that listeners identify as a specific artist is a different matter, and Tennessee's ELVIS Act reaches imitations that merely sound like the person, with no need for their name to have been used. If the point of the performance is that it sounds like them, treat it as a personal-rights question, not a copyright one.
Does Content ID clearing my track mean it is cleared?
No. Content ID matches recordings and compositions. It does not evaluate style, and it does not evaluate whether a voice was cleared. Passing it means you did not reproduce a matched recording or composition. Nothing else.
Do platform AI rules apply even to lawful content?
Yes, and they apply faster. Platform terms are contractual and enforced without a court. Deezer excluding AI-tagged tracks from editorial and algorithmic playlists is an economic consequence attached to lawful work, which is why platform policy belongs in the plan alongside the legal analysis rather than after it.