A Dance Floor Built on Borrowed Notes
When dancers at a recent performance moved to an AI-generated track, they likely had no idea the music was shaped by a song like "You Get What You Give" by the New Radicals — and millions of other recordings fed into the model without the original artists' knowledge or compensation. That is the uncomfortable reality at the center of generative AI music: the tools sound impressive precisely because they absorbed so much human creative work first.
Platforms including Suno, Google's music AI products, and Udio can generate full vocal tracks, instrumentals, and genre-specific compositions in seconds. The results are often indistinguishable from lo-fi indie or stock production music — which is exactly the problem critics and rights holders are raising.
What These Models Actually Do
AI music generation works by training large neural networks on audio recordings and, in many systems, paired metadata like genre tags, lyrics, and tempo markings. The model learns the statistical patterns of melody, harmony, rhythm, and timbre across enormous datasets. At inference time, a text prompt — "upbeat 90s pop with female vocals" — steers the output toward a region of that learned space.
The challenge is that training datasets are rarely disclosed. According to reporting by The Atlantic, the models were likely trained on recordings spanning decades of commercially released music — meaning hit songs, indie albums, and everything in between, ingested at scale. Whether that constitutes fair use under copyright law is now being tested in U.S. courts.
Why It Matters Beyond the Legal Fight
The copyright question gets the headlines, but the downstream effects on working musicians are more immediate. Session musicians, composers who license tracks for sync deals, and independent artists who sell background music to content creators are already reporting lost work. Unlike visual AI — where the end product is an image — AI music is directly substitutable for the product human artists sell.
There is also a quality trajectory to consider. Early AI music tools produced obvious artifacts: unnatural chord transitions, clipped vocals, rhythmic drift. Current systems from the major players are substantially better, and analysts expect continued rapid improvement as multimodal training — combining audio, lyrics, and video — becomes more common. The gap between AI-generated and human-produced music in mid-tier commercial categories is closing faster than most in the industry anticipated.
The Cultural Side: Festivals Embrace the Ambiguity
Not everyone in the creative world is approaching AI music defensively. The Altera Festival, which focuses on AI, extended reality, and spatial computing in film, is actively showcasing how these technologies can reshape creative pipelines — including the audio layer of filmmaking. For independent filmmakers, AI-generated scores represent a meaningful cost reduction; a custom soundtrack that once required a composer now costs a prompt.
This tension — liberation for some creators, displacement for others — is likely to define the next phase of AI's integration into creative industries. The same tool that frees a solo filmmaker from licensing fees is the one that undercuts the session musician who used to score that film.
A Concrete Example: "You Get What You Give"
The New Radicals' 1998 hit is a useful case study. It is a distinctive, well-known recording with recognizable production choices — a specific snare sound, a particular key, an era-specific mix. If an AI model trained on it can reproduce the feel of that song without reproducing the literal audio, current copyright frameworks may offer no remedy. Copyright in sound recordings protects the specific fixed recording, not the style, tempo, or emotional character it embodies. That distinction is at the heart of every pending lawsuit against AI music companies.
Suno and Udio have both faced legal challenges from major record labels, according to earlier reporting, with labels arguing the training process itself constitutes infringement regardless of whether the output reproduces specific recordings verbatim.
What To Watch
- Court rulings on whether ingesting copyrighted audio for AI training constitutes fair use will set binding precedent — a decision expected in at least one major case before end of 2026 could reshape how all generative audio companies operate.
- Licensing frameworks are being negotiated quietly: some labels are reportedly in talks with AI music platforms over revenue-share arrangements that would retroactively legitimize training data and create a royalty stream for rights holders.
- The emergence of provenance tools — watermarking and content-credential standards that flag AI-generated audio — will determine whether streaming platforms, sync licensors, and broadcasters can even tell what they are distributing.