AI Music
AI music is music composed, arranged, performed or produced with the help of algorithmic and machine-learning systems, from rule-based composition programs that write notation to neural models that synthesize finished audio, including vocals, from a text description. It is a broad field in which the form of the output matters: symbolic generators produce notes, as in MIDI or sheet music, which a person or instrument must still perform, while audio generators produce the waveform directly, so a single prompt can yield a complete song with lyrics, voice and production in a style the user names.
Algorithmic composition is older than computing as a mainstream tool. The Illiac Suite, written in 1956 and 1957 by Lejaren Hiller and Leonard Isaacson at the University of Illinois on the ILLIAC I computer, is generally cited as the first score composed by an electronic computer; it was a string quartet whose movements used generate-and-test rules for counterpoint and Markov-chain methods. Later landmarks include Iannis Xenakis's stochastic music, David Cope's Experiments in Musical Intelligence from the 1980s, which recombined patterns from composers such as Bach, and the Continuator and similar interactive improvisers. These systems encoded musical rules or statistics explicitly; deep learning, from about 2015 onward, learned them from data instead.
Neural music generation arrived in stages, each defined by what the model could represent. Magenta, launched in 2016 by the Google Brain team, released open-source tools and models for generating melodies, drum patterns and instrument sounds, including MusicVAE and the NSynth neural synthesizer. OpenAI's Jukebox, published in 2020, generated raw audio with rough singing by modeling compressed audio codes, though slowly and with noisy results. Google's MusicLM, described by Agostinelli and colleagues in January 2023, used the audio-language approach of AudioLM and conditioned on text to generate music at 24 kHz, and the team released a caption dataset for evaluation instead of the model itself. Meta's MusicGen, released later in 2023 as open weights, offered a research-grade alternative. Modern systems tokenize audio with neural codecs and predict the tokens with transformers or diffusion models, much as language models predict words.
Suno and Udio brought full-song generation to the public in 2024, and the music industry responded in court. Suno, which had released earlier versions in 2023, gained wide attention with its version 3 in March 2024, and Udio launched in April 2024; both let a user type a style description and optional lyrics and receive a produced song with a vocalist in roughly a minute. On 24 June 2024, the Recording Industry Association of America announced lawsuits on behalf of Sony Music Entertainment, Universal Music Group and Warner Records against Suno and Udio, alleging copyright infringement through training on recordings without permission. The companies' position, as stated publicly, was that their systems learn musical patterns and generate new works, and that training of this kind is lawful. The legal questions, which turn on fair use in the United States, were not resolved by the filings, and the situation continued to develop through negotiation and litigation.
Voice and likeness are a separate area of friction from composition. In April 2023, a track called Heart on My Sleeve, with AI-cloned vocals imitating Drake and The Weeknd, spread widely before being removed from streaming platforms after a rights-holder complaint, and the episode showed that a voice can be copied from recordings without anyone's participation. Some artists have since licensed their voices to developers deliberately, while others and several artist groups have objected to uses of their work for training. Streaming services have also had to deal with large volumes of low-quality generated tracks and with fraudulent streaming operations, and policies on labeling and removal vary by platform.
The common misconceptions are that AI music means no human involvement and that quality has reached parity with human production. Most professional uses involve people writing lyrics, selecting and editing generated stems, and mastering, and generated songs still show characteristic limits such as repetitive structure, mushy transients, unstable long-range form and lyrics that need correction. The clearest case is background and stock music, where substitution pressure is most direct. Ownership of a generated song is also unsettled: terms of service differ, and in the United States, copyright protection requires human authorship, so purely machine-generated audio may not qualify.
Read AI music as two technologies sharing a name: symbolic composition aids, which have existed since the Illiac Suite, and audio generators, which since 2023 and 2024 can produce a finished song from a sentence, with the second drawing the legal and cultural fight over training data, voices and who gets paid.