All articles

Is That Legal To Use AI Voice Clones in Your Songs?

My protégé-producer sent me a sample hook that sounds identically to The Weeknd but, of course, is not him. This is an AI clone of his voice, and the plan was to refer to the use of this AI in the song description and hope for the best. This plan will not work, and having seen too many clones ending up in takedown requests and threatening letters from lawyers, I wanted to finally set down the correct answer once and for all.

I use AI vocal technologies weekly, in my own voice and licensed models, and have spent the last three years researching the law, lawsuits and platform policies, which decide what will happen to the songs with AI voice clones. Here are the realities as of August 2026, without the hype and wishful thinking.

The Quick Summary

If the clone imitates the voice of a recognizable person and you do not have their consent, releasing such a song is illegal in most states in the USA and a criminal offense in Tennessee. It is also going to be removed from Spotify and other platforms due to the impersonation policy, long before any lawsuit appears.

However, if you clone your voice or use a licensed voice model from platforms like Elf.Tech, Kits.AI, or Voice-Swap, you are completely legal and ready for commercial release. The technology is not illegal; whose voice it is and whether this person has agreed to this is everything that really matters.

What Did the Fake Drake Really Prove

In April 2023, the TikTok user known as Ghostwriter977 released a track called Heart on My Sleeve with vocals sounding extremely similar to Drake and The Weeknd. These artists had nothing to do with this song. According to Billboard estimates, the track earned about 1.4 million streams in the United States in 4 days, earning around $7,500, while Axios reported around 15 million views on TikTok before the song was removed from all streaming services within a few daysafter the complaint filed by Universal Music Group to these platforms.

Here comes the most important detail, and the one I often cite. As NBC News noted back then, the song contained the producer tag of Metro Boomin, and the notice for the song removal on YouTube stated the copyright claim from Universal. This is also the same point that the legal blog Promise pointed out: the most obvious takedown hook was not the cloned voice at all, it was the unauthorized sampling. Copyright law gave the record label a fast legal tool. While the clones sat in the murky gray zone, that is why during the next three years the legislators tried to fix this gap.

Nobody filed a lawsuit against Ghostwriter977. They did not need this. The song disappeared from Spotify, Apple Music, YouTube, Tidal, Deezer, and TikTok within a week, along with any future streams it could have earned. This is the pattern: the legal question is debated in law reviews, while the practical question is answered by the takedown letter in 48 hours.

A Voice Cannot Be Copyrighted and That Goes Two Ways

Copyright is a protection of a fixed creative work: a recording, a composition, a lyric. Copyright does not cover the sound of the person’s voice as a concept. The section 114(b) of the Copyright Act even allows soundalike recordings, imitating the copyrighted recording. Thus, human impersonations on tribute albums were always legal in a strict sense of the copyright law.

These facts are confirmed in the first part of the report of the United States Copyright Office on the AI technologies and digital replicas, issued on July 31, 2024. The Office came to the conclusion that the existing law creates real gaps, recommended to Congress to create a new federal digital replica right for voice and likeness, and, as Jenner and Block’s analysis of the report states, clarified that Section 114(b) does not hinder the states from banning unauthorized voice replicas. In other words, you cannot rely on copyright to protect your voice, but you cannot hide behind the copyright and clone someone else’s voice.

But there are three situations where copyright still does bite. The first one, if the output of your AI contains an actual sample, as the Metro Boomin tag did. You are still infringing the copyright the old fashioned way. The second one, if the clone sings someone else’s melody or lyrics, and the composition becomes implicated regardless of whose voice performed this song. The third one is the biggest music law suit in this decade: the major record labels sued the song generators Suno and Udio in June 2024 for training their models on the record companies’ catalogs. The settlement between Universal and Udio in October 2025 included licensing for a new platform, Warner settled with both Udio and Suno in November 2025, while Sony is still fighting, according to reports by Billboard and The Hollywood Reporter. According to Music Business Worldwide, in June 2026 the American Federation of Musicians is suing Universal and Warner over these settlements, stating that the session musicians were not paid for a new use of their recordings. The training battle is far from being over and it decides which tools I am ready to use in my commercial releases.

Back in times before the AI era, the courts found that a deliberate imitation of the famous voice is a tort regardless of whether any copyrightable recording was copied. Back in 1988, the Ninth Circuit decided in Midler v. Ford that Ford’s ad agency violated a tort under the California law when it hired a soundalike to sing in a commercial after Bette Midler refused. Then she won a lawsuit and received $400,000. The logic was quite simple: a distinctive voice belongs to the person that it identifies.

Tom Waits went further. The radio advertisement for Doritos featured a singer hired specifically for his Waits’ imitation, and the Ninth Circuit’s Waits decision in 1992 upheld the verdict with $375,000 of compensatory damages and $2 million of punitive damages for voice misappropriation, trimming only the $100,000 duplicative false endorsement award. Punitive damages, for imitation of the voice. In 1992.

They were the cases of the human impersonations and product ads, and it may seem that releasing an AI-clone of someone’s voice for commercial purposes is different, because it is music, not an advertisement. But I would not stake my career on this distinction. Your monetized release trades on the cloned artist’s identity, courts have always considered commercial exploitation broadly and the statutes below were adopted to prevent any attempts of squeezing AI voice clones into a 1988 framework.

Tennessee Criminalized Voice Cloning

The Ensuring Likeness Voice and Image Security Act, or the ELVIS Act, was signed on March 21, 2024 and became effective on July 1, 2024. This is the first state law making a person’s voice, explicitly including the simulation of this voice that can be recognized easily, a protected property right. Scholars in the Vanderbilt Journal of Entertainment and Technology Law described it as the template of a future federal law.

The consequences are important. According to the analysis of the law by Holland and Knight, the violations allow a civil lawsuit and also can be criminally charged as a Class A misdemeanor, punishable with up to 11 months and 29 days of imprisonment and fines up to $2,500. The statute also holds not only the person who uploaded the AI clone, but the distributors of the tools whose primary function is producing the unauthorized clones. It passed 93 to 0 in the Tennessee House and 30 to 0 in the Senate.

This statute covers Tennessee residents, and this is the practical detail: a huge number of American recording artists live in and around Nashville. Clone a country singer, a Christian artist, or half of the session vocalists in the business and there is a fair chance that you chose the person that is covered by this law. I see Tennessee as the floor of the possible enforcement, not an outlier.

California, Illinois, Federal Statute and Some Real Momentum

California signed two digital replica laws on September 17, 2024. As Proskauer’s summary describes, the AB 2602 invalidates contract language allowing a company to use a digital replica of the performer’s voice or likeness in lieu of the performance, unless the uses are explicitly specified in the contract, or the performer had an attorney or a union representing them, starting January 1, 2025. The AB 1836 bans the production and distribution of digital replicas of deceased performers without consent of their estates, starting with $10,000 statutory damages per violation. The state of Illinois amended its Right of Publicity Act to cover digital replicas and passed a separate Digital Voice and Likeness Act focusing on contract terms, as Covington’s Global Policy Watch reports. New York and Louisiana have their replica statutes as well, though the Copyright Office mentioned that both contain exclusions of some voice simulations, that is why the map looks like a patchwork.

The federal solution is the NO FAKES Act, and 2026 is the year when it gained momentum. The bill was reintroduced on May 20, 2026 as S.4591 and H.R.8915, its fourth version since the 2023 discussion draft. As Manatt’s client alert mentions, this version has a new element: a notice-and-takedown procedure modeled on the DMCA, so that a person whose upload is flagged can formally challenge the removal. According to the bill summary, it is a licensable federal property right of every individual’s voice and visual likeness, that survives after death, plus a notice-and-takedown procedure for unauthorized replicas, while preserving the state laws that existed as of January 2, 2025.

According to the S&P Global, the Senate Judiciary Committee passed it unanimously in June 2026, gaining the support of the labels, Hollywood and major tech companies. RIAA endorsed the bill, stating in the Senator Blackburn’s announcement that 92 percent of Americans are concerned about AI deepfakes. As of early August 2026 I write this text, the bill is not yet the law, but I do not try to predict what Congress may do. However, the tendency is obvious: unauthorized voice cloning is on the way of gaining a federal cause of action. It is not only an American tendency. According to Soundplate’s producer guide, it is a 2024 Bombay High Court decision that protected the voice of the singer Arijit Singh under the personality rights, after his catalog was used to train the clone models without consent.

Platforms Remove Faster Than Any Court Does

Even if you will never enter a courtroom, the platforms decide whether your song is alive or dead. In September 25, 2025 Spotify announced its impersonation policy declaring that vocal impersonation is allowed only if the impersonated artist has authorized it. Read the policy carefully: it is the clone whether you pretend to be the artist or not, it is the track where the voice is obviously recognized as belonging to the other artist, even if you do not mention their name in the metadata. The declaration that your song is an AI-version of Drake does not help you. There are two narrow exceptions for parody, and that’s it.

This announcement revealed the scope of the cleanup operation. According to Music Business Worldwide, Spotify removed more than 75 million spammy tracks in the previous 12 months, while TechCrunch announced the new DDEX disclosure standard allowing artists to specify whether AI was used for vocals, instrumentation, or post-production, with 15 labels and distributors signed at launch. DDEX is the common music metadata standard and AI disclosure becomes the regular field in it, and this is a genuine good news for all of us who use AI honestly.

Your distributor is another chokepoint. TuneCore, DistroKid, and others ban impersonation in their terms, and repeated strikes may lead to the removal of the entire catalog, not only one track. The tools themselves are the targets: the British recorded music trade body BPI threatened to file legal action in March 2024 against Jammable, formerly called Voicify, as Music Business Worldwide reported. The site offered more than 3,000 cloned voice models including Rihanna, Drake, and Taylor Swift, while RIAA previously mentioned this service in its notorious markets submission. The Rihanna’s model disappeared silently after the pressure. My blunt advice: do not build anything using cloned celebrity voices from this type of sites. The output is legally toxic and the sites themselves may not survive the release period.

Releasing Music into Europe Just Became More Difficult

If your music is distributed in the EU countries, there is a new deadline you have to respect. On August 2, 2026, five days before I write this text, the European Commission began to enforce the transparency rules of the AI Act. The AI-generated or altered audio that realistically depicts a real person becomes the deepfake and requires labeling. The synthetic content requires machine-readable marking so that the detection tools can recognize it.

As Cooley’s briefing on the rules states, the penalties can reach 15 million euros or 3 percent of worldwide annual turnover, while the heaviest duties fall on the AI providers and professional deployers. Also, there is the transition period till December 2, 2026 for some marking obligations. The artistic works have less strict rules on the presentation of the disclosure, but the era of silently passing off the AI-generated vocals as human, in the Europe at least, is over. Practically, this leads everyone to the same habit: disclose the use of AI in the credits and metadata.

The most interesting part, because the AI clones of voices are indeed a legitimate tool if the consent is legitimate. Here are the ways I use them and recommend to use them, with some real numbers.

Cloning your voice. This is legal everywhere, and it is the main use case. I run my own voice model for double tracking, harmony stacks, and demoing toplines outside of my vocal range. One detail to mention if you are signed: read your contract before you license your voice to anyone, because the label contracts may cover the vocal recordings broadly, and the AB 2602 of California was adopted exactly to limit this kind of stretching of the boilerplate contract language.

Using the proper consent, the Randy Travis style. In May 2024, Randy Travis, who lost his singing ability due to a stroke in 2013, released Where That Came From using an AI model, as CBS News reported, that was generated from 42 of his vocals from his catalog, applied to a guide vocal from singer James Dupré, with Travis participating in every step. Rolling Stone reported that Warner even introduced a new credit, vocal bed, for Dupré’s contribution. This is the entire recipe of doing it the right way: the consent, compensation, credit, and disclosure.

Elf.Tech, if you want the Grimes’ voice. In 2023, Grimes decided to open license her voice, stating “feel free to use my voice without penalty.” This means the free access to the voice model in exchange for 50 percent of the master royalty and the GrimesAI feature credit. The Forbes reported on the partnership with TuneCore, allowing the distribution. However, before you start, read TuneCore’s published rules for the GrimesAI releases: the 50 percent split is applied no matter how much of the song uses the voice, stems must be downloaded from the official platform, no YouTube Content ID, no covers of copyrighted songs. My personal opinion: the half-split for a single adlib is very expensive, so treat the clone as the featured artist and use where the feature earns its money.

Kits.AI, if you want a licensed voice library. Kits offers artist voice models that are, as Kits claims, licensed ethically directly from the vocalists, who earn revenues through its Kits Earn program. In case of the free tier, you’ll get one custom voice slot and 15 minutes of conversions, and according to the official pricing page, the cheapest option is $9.99 a month, scaling to around $60 a month on the highest tier. Just so you know – and this goes for anyone using it – the list of artists is top-notch session-singer quality, not celebrity impersonators, and that’s the idea. If marketing of a tool focuses on celebs, it probably isn’t the right tool.

Voice-Swap, for credited featured artists. This UK-based platform collaborates with active vocalists such as Robert Owens, Ayah Marar, Angie Brown, and Liam Bailey, and its founders told Billboard theirs was the first platform to pay royalties for each conversion (not just future streams). You start with 60 seconds of free credits and go with a subscription plan, and according to Deepgram’s description of the service, each commercial release will require a per-song license that artists commit to replying to within 48 hours. It’s also a VST plugin that integrates the whole thing into Logic, Ableton, or FL Studio, which is how I test it.

Another person’s voice, including a friend’s. Put it in writing. A single-page release that covers the use for training, the specific song, credit, and compensation takes you twenty minutes to draft and will save you a friendship or a lawsuit. Remember to keep provenance: session files, stems, the licensing email, the receipt. When a distributor or a label questions a vocal and they do, this folder is your answer.

Three Misconceptions That End Up Costing People Their Careers

Misconception one: It’s OK if I don’t monetize. Wrong on two accounts. Tennessee’s law doesn’t wait for advertising context the way the older cases did, the Copyright Office has recommended including non-commercial distribution, and the NO FAKES initiative targets replicas regardless of whether they earned any money. And platforms take down clones no matter if it ever made a dollar. Monetization only worsens your case; its absence doesn’t help.

Misconception two: I labeled it as AI, so I’m covered. Disclosure is not consent. Spotify explicitly covers uploads presented as AI versions of an artist, and putting “AI” in front of “Drake” means you knew perfectly well which identity you traded on. Parody gets you some leeway both on Spotify’s exceptions and under the First Amendment, but an earnestly serious song in cloned voice is not parody and courts understand the difference.

Misconception three: Fair use will protect me. Fair use is a copyright doctrine. Right of publicity and the replica laws are not copyright, so they are not answered by fair use. I hear this from producers weekly, and this is the most common advice in this sphere, despite being fundamentally wrong.

Four Questions to Ask Yourself Before Uploading a Song

Whose voice is it? If a random listener will recognize a specific living or dead artist, you need to get permission.

Can you produce the permission? Platform license, a signed agreement, and official programs such as Elf.Tech count. Screenshot of someone’s tweet or Discord message, or voice model from a clone site don’t.

Where could this lead you? If the impersonated artist is a resident of Tennessee, California, or Illinois, or your song will reach EU listeners, you’ve got yourself the most stringent laws currently active.

Did you disclose? Fill in the AI fields your distributor offers under DDEX standard and credit licensed voices as the license requires. Disclosure will not legalize a clone you weren’t allowed to use, but not doing it can tank your legal release.

What Would I Do This Week?

If you want to experiment with AI voice clones, you may start with those whose legality is settled. Tonight, train a model on your voice using Kits’ free tier and create a harmony stack with it. In case you need a recognizable featured vocalist, run a hook through Elf.Tech and set up the 50% split correctly or ask for Voice-Swap license and actually read it. Then set up your disclosure fields with your distributor so that this becomes a habit before it becomes mandatory.

And in case you’ve got a cloned celebrity vocal in the track you love, replace it now, when it is on your terms, rather than later when it is on theirs. I am not a lawyer, and if there are money or actual disputes involved, you should definitely talk to a lawyer specializing in music business. But after three years of watching this field, the trend became clear to me: the artists who get burned by voice clones were those who asked for forgiveness. The artists building careers with clones asked for permission.

Sources

Common questions

Is it illegal to use an AI voice clone of a famous singer in a song?

In most of the United States, yes, under right of publicity laws, and in Tennessee it is a Class A misdemeanor under the ELVIS Act with up to 11 months and 29 days in jail and fines up to $2,500. California and Illinois have digital replica statutes too, and the federal NO FAKES Act passed the Senate Judiciary Committee unanimously in June 2026. Even where the law is unsettled, platforms remove such tracks within days.

Can a voice be copyrighted?

No. Copyright protects fixed works like recordings and compositions, not the sound of a person's voice, and Section 114(b) of the Copyright Act even permits soundalike recordings. That cuts both ways: you cannot use copyright to protect your own voice, but you also cannot hide behind copyright to clone someone else's, because voice misappropriation is handled by right of publicity and the new replica laws instead.

What happened to the fake Drake song Heart on My Sleeve?

Released in April 2023 with vocals mimicking Drake and The Weeknd, it earned about 1.4 million US streams in four days and around 15 million TikTok views before Universal Music Group had it pulled from every streaming service within a week. Nobody was sued. The quickest takedown hook was actually an unauthorized Metro Boomin producer tag, showing that copyright still applies when a clone track contains real samples or someone else's melody or lyrics.

Does Spotify allow AI voice clones?

Only when the impersonated artist has authorized it. Spotify's September 2025 impersonation policy covers any track where the voice is recognizably another artist's, whether or not you name them in the metadata, and labelling a song as an AI version of an artist does not help. Parody gets narrow exceptions. Distributors like TuneCore and DistroKid also ban impersonation and can remove an entire catalog after repeated strikes.

Is it legal to clone my own voice with AI?

Yes, everywhere, and it is the main legitimate use case for double tracking, harmony stacks and demoing toplines outside your range. One caution for signed artists: read your label contract before licensing your voice to anyone, since some contracts define vocal recordings broadly. California's AB 2602 was passed specifically to stop boilerplate contract language from covering digital replicas without explicit terms.

Where can I get a licensed AI voice for a commercial release?

Elf.Tech offers Grimes' voice for a 50 percent master royalty split and a GrimesAI feature credit, distributed through TuneCore under specific rules. Kits.AI provides ethically licensed session quality voices from $9.99 a month, with a free tier of one custom voice and 15 minutes of conversions. Voice-Swap works with credited vocalists, pays royalties per conversion, requires a per song licence for commercial releases and runs as a VST plugin.

If I label my song as AI generated, am I protected?

No. Disclosure is not consent. Spotify's policy explicitly covers uploads presented as AI versions of an artist, and writing AI in front of a famous name shows you knew whose identity you were trading on. Disclosure through DDEX fields is still required and good practice, but it cannot legalize a clone you had no permission to use.

Does fair use cover AI voice clones?

No. Fair use is a copyright doctrine, while voice clone claims arise under right of publicity and digital replica laws, which are not copyright and are not answered by fair use. The article calls this the most common and most fundamentally wrong piece of advice circulating among producers.

Do I need permission if I am not monetizing the song?

Yes. Tennessee's law does not require a commercial context, the Copyright Office has recommended covering non commercial distribution, the NO FAKES Act targets replicas regardless of earnings, and platforms remove clones whether or not they made a dollar. Monetization makes your position worse, but its absence does not make it safe.

What are the EU rules on AI voices in music?

Since August 2, 2026 the European Commission enforces the AI Act's transparency rules, under which AI generated or altered audio realistically depicting a real person counts as a deepfake and must be labelled, with machine readable marking so detection tools can identify it. Penalties can reach 15 million euros or 3 percent of worldwide turnover, mostly aimed at providers and professional deployers, with a transition period until December 2, 2026 for some marking duties. Artistic works get lighter disclosure rules but not exemption.