Songs Without Hands: AI music generation, honestly

In three years it went from a viral stunt to a courtroom fight to a boardroom handshake. Here is what the technology actually does, what is genuinely contested, and what it means for the people who make music.

In April 2023, an anonymous producer calling themselves Ghostwriter posted a track titled "heart on my sleeve" that paired startlingly convincing clones of Drake's and The Weeknd's voices over an original beat. It raced across TikTok and the streaming services within days, then disappeared almost as fast after Universal Music Group filed copyright claims to pull it down.¹² Neither artist had anything to do with it.¹

The song was a stunt. It was also a starting gun. In the years since, AI music generation has moved from curiosity to litigation to licensing — the tools got good enough that people use them every day, the lawsuits got serious enough to reshape an industry, and the major labels went from suing the companies to signing deals with them. This is a field guide to where things honestly stand.

01How it actually works

Generating audio is harder than generating text or an image, in part because sound carries an enormous amount of information every second. Most modern systems sidestep that by not working with raw audio directly. Meta's openly released AudioCraft toolkit is a useful, documented example: it pairs a neural codec called EnCodec, which compresses audio into a compact stream of discrete "tokens," with a model called MusicGen that learns to predict those tokens from a text description before decoding them back into sound.³ Other systems lean on diffusion, the same broad family of techniques behind many image generators.

The important caveat: the commercial leaders have not published the full details of how their models are built. Any confident, specific claim about exactly what happens inside Suno or Udio should be treated with a degree of skepticism — including claims made in their own marketing.

02The tools on the table

The product that pushed AI music into the mainstream is Suno, a Cambridge, Massachusetts startup.⁴ You describe a song in plain language and it returns a finished track — vocals, lyrics, instrumentation and structure — in well under a minute, with a free tier and paid plans for commercial use.⁵ Its March 2026 release, v5.5, added a feature called "Voices" that lets paying users record their own voice and sing through it, with a verification step meant to discourage impersonation.⁶ Its closest rival, Udio — built by a company called Uncharted Labs — works along similar lines.⁷

The platform giants have arrived too. Google DeepMind's Lyria model now powers music generation inside the Gemini app, where a text prompt or an uploaded image produces a short, vocal-capable track.⁸ Meta took a different road, releasing its AudioCraft models as open source so that researchers could train their own systems — and noting candidly that its music training data skewed heavily Western and English-language.⁹ That divide — closed commercial products on one side, open models on the other — runs through the entire field.

03The legal earthquake

The central fight has been about training data. On 24 June 2024, the Recording Industry Association of America, on behalf of Universal, Sony and Warner, filed twin copyright lawsuits — against Suno in federal court in Boston, and against Udio in New York.¹⁰ The labels alleged that the companies had copied sound recordings on a massive scale to train their models, and asked the courts for injunctions and damages.¹⁰ Suno's chief executive countered that the technology was designed to generate completely new outputs, not to memorize and regurgitate existing songs.¹¹ In their formal responses, the companies leaned on fair use, comparing model training to a person learning from the music they hear; the labels replied that copying an artist's catalogue to build a competing product is not "fair" at all.¹²

Then the ground shifted. Rather than wait for a verdict, the majors began cutting deals. Universal settled with Udio in October 2025;¹³ Warner settled with Udio in November and announced a licensed AI platform built on authorized music.¹⁴ Days later, Warner settled with Suno in a partnership that also saw Suno acquire Warner's Songkick concert-discovery service, with new, licensed models slated to replace the older ones.¹³ But as of spring 2026, Sony had settled with neither, and Universal and Sony's claims against Suno were still live¹⁵ — a reminder that the precedent-setting question of whether unlicensed training is legal has not, in fact, been answered. What is emerging in its place is a licensing regime: future models trained on catalogues the labels are paid for.¹³

04The human stakes

The legal story can drown out the human one. In April 2024, more than 200 musicians — among them Billie Eilish, Stevie Wonder, Nicki Minaj and the estates of Bob Marley and Frank Sinatra — signed an open letter through the Artist Rights Alliance calling the irresponsible use of AI an "assault on human creativity."¹⁶ Their objection was not the technology itself, which the letter explicitly allowed could be used well, but its predatory use: cloning voices and likenesses without consent, and flooding streaming platforms with AI output that dilutes the royalty pools working musicians rely on.¹⁷ A UK industry body put it more bluntly, describing unlicensed training as "music laundering."¹⁷

The Ghostwriter track had already shown how unsettling a cloned voice can be — and exposed a gap in the law itself.

An original song sung in someone else's synthetic voice raises questions of publicity and likeness that ordinary copyright does not neatly cover, which is part of why that 2023 track was so disorienting to the industry.¹ It is one thing to copy a recording; it is another to convincingly impersonate a person who never performed.

05Telling human from machine

One partial answer is provenance — being able to tell what was machine-made. Google embeds an inaudible watermark called SynthID into every track its Lyria model generates, encoded into the audio itself so it survives common changes such as MP3 compression, added noise, or speeding the track up.¹⁸ It is a genuinely useful tool for honest labelling at scale. It is not a force field. No watermark is perfectly tamper-proof, and provenance only helps if the platforms and regulators that sit downstream actually require it and check for it — which, for now, most do not.

06Where Wave lands

At Wave, we try to judge a tool the way OCXLY judges any tool: by asking who it serves. Used as an instrument — for sketching an idea, scoring a video, learning an arrangement, or getting an unsigned artist to a demo without a studio budget — generative music is plainly useful, and the people reaching for it are not the problem. The problems are specific, and they are addressable: training without consent, cloning without permission, distributing without disclosure, and paying the humans whose work made the model possible little or nothing.

The genuinely encouraging shift of the last three years is that the question has moved from "is this even legal?" toward "how do we do this with consent, credit and compensation?" The licensed platforms now being built are a flawed, partial version of that answer; watermarking is another. Neither is finished, and neither should be mistaken for a settled ethics. But the direction matters. Technology that respects the human mind has to begin by respecting the humans who made the thing it learned from. For music, that isn't a constraint on the future — it's the only version of it worth building.