Music generation models use two main approaches: audio-native models (generate raw audio waveforms using architectures similar to diffusion models or autoregressive Transformers) and MIDI-based models (generate symbolic music notation that's then rendered with synthesizers). Audio-native models (Suno, MusicGen) produce more realistic results but are computationally expensive. MIDI approaches are more controllable but less natural-sounding.
Music AI raises intense copyright questions. Models trained on copyrighted music may reproduce recognizable elements — a melody, a vocal style, a production technique. Some platforms have been sued by record labels. The legal status is evolving: generating "music in the style of" an artist may be legal (style isn't copyrightable), but generating something that sounds like a specific song isn't. Most commercial music AI services implement filters to prevent generating content too similar to known copyrighted works.
Beyond replacing musicians, AI music enables new creative workflows: generating demo tracks that producers then refine, creating adaptive game soundtracks that change based on gameplay, producing personalized music (a lullaby with your child's name), and enabling music production for people with ideas but no instrumental skills. The most interesting applications treat AI as a creative collaborator rather than a replacement.