AI music generation: how AI makes songs, explained

In plain words

You type a mood, a style and maybe some lyrics. A minute later you have a song: instruments, melody, even a voice singing your words. Apps like Suno made this popular from late 2023.

The AI learned by listening to a huge amount of recorded music, along with descriptions of it. It picked up how beats repeat, how a chorus returns, how a sitar or a bassline usually sounds. It then makes new audio that follows those patterns and fits your request.

It is fun and useful for jingles, videos, games and sketching ideas. It is also controversial. Record labels say some tools trained on their songs without permission, and as of October 2026 courts are still deciding what is allowed.

Everyday example

A wedding band that plays any request

A good wedding band has played thousands of songs. Shout "something like an old Bollywood duet, but faster" and they improvise a tune in that style on the spot. They are not playing one existing song, but everything they play comes from what they have heard. Music AI works in a similar way.

Where the example breaks: Band members understand the couple, the moment and the meaning of the lyrics. The AI only predicts sound that fits patterns in its training. It can also echo real songs or voices too closely, which a band would know to avoid.

Go deeper

How it actually works

Music is hard for AI: a few minutes of audio is millions of numbers, and songs need structure that lasts, like a chorus that comes back.

Turn sound into tokens
Raw audio is squeezed by a neural audio codec into a much shorter sequence of tokens, a bit like words for sound. Google's MusicLM (January 2023) used two kinds: semantic tokens for long-term structure and acoustic tokens for sound quality.
Link words and music
MusicLM also used a separate model, MuLan, that learned which text descriptions go with which clips, similar to how CLIP links images and captions. This lets a typed prompt like "calm violin with a distorted guitar riff" steer the music.
Generate the sequence
A transformer predicts the next sound tokens, one after another, or a diffusion model refines noise into audio. Either way, a decoder turns the result back into a sound wave. Commercial apps such as Suno and Udio have not published full details of how they work.
Voices, and the rules around them
Many apps also sing the lyrics you type. Because voices can be imitated, platforms are adding rules: Spotify said in September 2025 that vocal impersonation is only allowed when the impersonated artist has authorised it.

Computer music roots: 1957: the Illiac Suite, a string quartet composed with a computer at the University of Illinois, generally seen as the first computer-composed score
Raw audio with singing: April 2020: OpenAI's Jukebox, which took about 9 hours to render one minute of music
Mainstream apps: Suno widely available from December 2023; MusicLM paper January 2023
Scale, as of July 2026: The streaming service Deezer said more than half of the tracks uploaded to it each day were AI-generated, averaging about 90,000 a day in June 2026

In the US, the Copyright Office said in January 2025 that prompts alone do not make you the author of an AI output, so a purely AI-made song may not be protected by copyright. See AI and copyright.

Is it fair to train music AI on existing songs without permission?

One view

Yes, it is learning

Suno argues in court that training on recordings is "quintessential fair use". Supporters say a model learns general patterns of music, much as musicians learn by listening, and makes new songs rather than copies.

The other view

No, it needs a licence

Record labels argue that copying millions of recordings to build a commercial product needs permission and payment, especially when the outputs compete with the artists whose work was used.

Most likely: As of October 2026 there is no final court ruling. Universal and Warner have settled with Udio, and Warner with Suno in November 2025, in exchange for licences, while Universal and Sony are still suing Suno. The trend is towards licensed training, but the law is not settled.

Questions people ask

Is AI-generated music copyrighted?

It depends on the country and on how much a human contributed. In the US, the Copyright Office says prompts alone are not enough for authorship, so a song made purely by typing a prompt may have no copyright. Lyrics you wrote yourself remain yours. Check each app's terms too.

Can AI copy a famous singer's voice?

Technically it can imitate voices, which is why it is restricted. Spotify, for example, only allows vocal impersonation when the artist has authorised it. Releasing a fake song in someone's voice can break platform rules and, in many places, the law.

Why are record labels suing AI music companies?

In June 2024, major labels sued Suno and Udio, saying they trained on copyrighted recordings without permission. Some cases settled in late 2025 with licensing deals. In September 2026, Universal and Sony filed a second lawsuit against Suno over its newer models.

Can I use AI music in my YouTube videos?

Often yes, if the app's terms allow it, so read them first. Some laws, such as India's IT rules amended in February 2026, require AI-generated content to be labelled. Keep a record of how you made the track, and avoid prompts that imitate a specific artist or song.

Music generation