This article is based on episode 2 of the AD Podcast, Audio Dream’s podcast about the music industry.
In the first episode, AI was only touched on briefly. Here, it takes center stage: how generative AI has transformed music over the past three years—much faster than most people realize. We look at four layers: how the tools work, voice cloning and identity, the ongoing legal battle, and where business models are headed.
One date effectively opened this conversation: April 2023, when “Heart on My Sleeve” went viral using AI-cloned voices of Drake and The Weeknd in a song neither had ever recorded. The track was taken down after the label intervened, but the breakthrough was clear: with the right tool, anyone could convincingly clone a global artist’s voice. The following three years became a race to create rules, defensive technology, and business models.
How Generative Music AI Works (and What It Can Already Do)
Tools such as Suno and Udio are models trained on vast amounts of existing music. They learn patterns—how pop beats are structured, how a chorus builds, and how a timbre behaves in a genre. You enter a style, a mood, and sometimes finished lyrics, and the tool generates a complete track with vocals, instruments, and mixing in under a minute.
The radical difference is that every earlier tool—synthesizers, editing software, samplers—required musical, programming, or operating skills. Generative AI removes that requirement: the barrier to entry has fallen from “years of musical study” to “knowing how to write a prompt”. And the quality is no joke: many AI-generated tracks can fool a trained ear today, at least for the first few seconds.
The First Practical Problem: Volume (and Fraud)
Before any philosophical discussion about creativity, there is the problem of sheer volume. Spotify removed more than 75 million spam tracks in one year — most of them linked to mass AI generation: short, duplicated tracks with misleading titles, created to capture fractions of a cent in royalties at industrial scale. This is not “an artist using AI to compose”; it is large-scale fraud, with thousands of tracks uploaded every day. Spotify responded with a spam filter that looks not at the sound but at upload patterns: bulk uploads, duplicate titles, and tracks that are too short.
Legitimate Uses (An Important Distinction)
- AI as a working tool: a producer generating a draft arrangement in seconds before a live recording; a jingle composer creating ten variations in an afternoon; an independent artist adding a string layer they could not afford to record. It speeds up the path to a creative decision—it does not replace it.
- Functional music: background tracks for videos, podcasts, call-center hold music, and store playlists—functional uses where the demand is for “cheap background sound.” This market (royalty-free libraries and stock music) already existed; AI has made it faster and cheaper.
The real tension lies at both extremes: mass-upload fraud and directly imitating a specific artist without consent. One fact shows how quickly things changed: Suno and Udio barely existed as consumer products before 2023, yet in less than three years they faced action from the three largest record companies in the world at the same timeNeither CDs, MP3s, nor streaming followed this curve. The technology arrived and became a mass-market product before rules, contracts, and licenses were built—the root of today’s legal turmoil.
Voice Cloning and Artist Identity
Who owns your voice? Technically, it takes only minutes of reference audio to clone an artist’s voice today. The market has reacted along two opposing paths, both active:
- Total resistance: o ELVIS Act, passed in Tennessee in 2024, it was the first U.S. state law to explicitly protect a person’s voice against AI cloning as personal property, alongside their name and image. It covers living people and estates; unauthorized use can lead to civil lawsuits or criminal charges. It is no coincidence that the law began in Nashville.
- Licensing: the singer Grimes launched the Elf.Tech in 2023. Anyone can record a vocal and transform it into her synthetic voice. Creative use is free, while commercial releases split 50% of master royalties with the artist. By mid-2026, more than 160 tracks had been officially released through the program.
Neither path has “won,” and they will probably coexist, serving different kinds of artists. But celebrity controversies often leave out another group: studio musicians, voice actors, and jingle singers — professionals who make a living lending their voices, often without dedicated lawyers or a record label fighting for them. They are the first to hear, “Why pay a session singer when a similar voice can be generated for a tenth of the price?” Musicians’ unions and voice-acting groups are already negotiating consent clauses for using voices to train AI—protections that barely existed a few years ago. Another side of the story is genuinely compelling:
Another genuinely interesting aspect is: voice cloning with explicit consent already allows an artist who lost their voice to illness to keep releasing music with their own voice digitally preserved—or a deceased artist, with the estate’s permission, to “take part” in a collaboration decades later. The technology is essentially the same as in “Heart on My Sleeve.” The difference is consent, contracts, and compensation: three words that determine whether a use is legitimate innovation or a violation.
The Legal Battle: Who Owns What AI Has Learned
Em June 2024, the RIAA sued Suno and Udio on behalf of Universal, Sony, and Warner. The companies were accused of training their models on huge volumes of copyrighted recordings without a license. This is probably the most important recent case in music intellectual property: the outcome will literally determine the cost of training AI on copyrighted music. The story then split into two parallel paths:
The story then split into two parallel paths:
- Resolved through licensing: Warner reached a deal with Suno, replacing its lawsuit with a licensing relationship. Universal did the same with Udio in October 2025, going further by announcing a joint AI music platform built around catalog licensing. Udio also reached agreements with Merlin (independent labels, early 2026), publisher Kobalt (spring 2026), and the NMPA (the U.S. music publishers’ association, June 2026)—creating a licensing framework around much of the relevant catalog. joint AI music platform built around catalog licensing. Udio also went on to reach agreements with Merlin (independent labels, early 2026), publisher Kobalt (spring 2026), and the NMPA (the U.S. music publishers’ association, June 2026), creating a licensing framework around much of the relevant catalog.
- Still in court: a Sony Music continues its case against both companies. In May 2026, Universal and Sony asked to add more than 61,000 recordings to the lawsuit against Suno after digital audio identification revealed training on millions of protected tracks. The U.S. court moved the core question—whether training qualifies as “fair use”—to 2027. more than 61,000 recordings to the lawsuit against Suno after digital audio identification revealed training on millions of protected tracks. The U.S. court moved the core question—whether training qualifies as “fair use”—to 2027. 2027
The dispute is not limited to the United States: in July 2026, important decisions involving Suno were issued in Munich and Boston. Different jurisdictions and rules are creating a regional patchwork rather than one global answer. The core argument is that AI companies say their models “learn patterns” without literally copying (transformative use), while record labels say the training is commercial-scale copying that competes with original music. It is one of the most expensive questions in technology today.
In practice, licensing is already shaping the market: a licensed tool (such as Udio after its deal with Universal) is likely to have a broader catalog and less risk of losing features through a court ruling. A tool still in litigation (Suno, with Universal and Sony) operates in a gray area. In parallel, European regulation will require AI-generated content—including audio—to be technically labeled by AI Act will require technical labeling of AI-generated content—including audio—by August 2026.
Authenticity, Watermarks, and What Comes Next
While courts decide, the industry is building technical tools to distinguish human-made content from AI:
- Audio watermarking: an inaudible technical signature embedded in sound, detectable by software—one example is SynthID (Google), which survives compression and format conversion.
- Signed metadata (C2PA): rather than marking the sound, it marks the file—an “origin certificate” that travels with it and documents authorship, licensing, and AI involvement.
- Credit disclosure (DDEX): since September 2025, Spotify has adopted the standard for labels and distributors to declare AI use in vocals, instruments, or post-production. Declared and licensed AI is not penalized; the platform prohibits unauthorized cloning, deepfakes, and impersonation.
An honest technical caveat: a watermark is a strong signal, but its absence proves nothing — not every tool watermarks what it generates, and someone with the right skills may remove a mark. Watermarking is one part of a trust system, not conclusive proof on its own. This is also driving a cultural shift:
This is driving a cultural shift: human authenticity is becoming a premium product — labels and artists are using claims such as “recorded by humans, with no artificial editing” as a selling point, much as the food industry uses “organic” or “artisan.” This is not nostalgia; it responds to fans who value knowing a real person is behind the sound. The same idea is reaching audio equipment: independent certification, published bench measurements, and transparent technical reviews— the more uncertain the content, the more valuable every verifiable source of trust becomes.
Three plausible scenarios, without overreaching: (1) licensing becomes the dominant model, with remaining disputes focused on value rather than principle; (2) watermarking and credit disclosure become standard requirements on major platforms; and (3) human-made authenticity grows into its own category alongside AI-generated music, just as collectible vinyl coexists with algorithmic playlists. The market will probably accommodate all of them.
Frequently Asked Questions
Will AI destroy music?
Neither on its own—it will neither destroy nor save music. AI is forcing the entire industry to decide quickly which rules for trust and compensation will apply.
Is cloning an artist’s voice legal?
Without consent, no—and increasingly so. Laws such as the ELVIS Act treat a person’s voice as protected personal property, and unauthorized use can lead to civil lawsuits or criminal charges.
Who won the legal battle between record labels and AI companies?
There is no final ruling yet. Half the industry has moved toward licensing (Warner + Suno, Universal + Udio), while Sony remains in court. A fair-use decision is expected in 2027.
How can you tell if a track was made by AI?
Audio watermarking (SynthID), signed metadata (C2PA), and credit disclosure (DDEX) can help—but the absence of a watermark does not prove human authorship.
Is it wrong to use AI to compose music?
It depends on how AI is used: as a tool in a genuine creative process, it is legitimate; mass fraud and imitation without consent are violations.
Consent, Contracts, and Compensation
The thread connecting everything is always the same: consent, contracts, and compensation determine whether using AI in music is legitimate progress or a violation—the underlying technology is practically identical on both sides. For everyday listeners, this is no longer abstract: a track in your playlist may be AI-generated, AI-assisted, or entirely human-made, and technical ways to verify which will become more common. Follow the AD Podcast deep dives on the Audio Dream blog — and if you want to listen to music through equipment that delivers on its promises, contact us.
