Two companies, Suno and Udio, are at the center of every argument about AI music. Both raised nine-figure rounds in 2024. Both have generated more tracks in the last 18 months than the entire recorded history of rock, jazz, and hip-hop combined. Both are being sued by the RIAA. Both are, in different ways, right about what they have built and wrong about what it means.
Here is what is actually true about AI music in 2026, without the marketing, without the moral panic, and without the dismissal.
What the models can do
Suno v4 and Udio v1.5 can produce a coherent 3-minute song, vocals included, in a specific style, in under 30 seconds. The output is indistinguishable from human-produced music in blind A/B tests more often than chance. The “uncanny valley” of AI vocals is closing fast, especially on pop, electronic, and ambient styles where the human reference is already heavily processed.
What they cannot do, or cannot do reliably:
- Sustain a coherent lyrical narrative across a full song. Choruses drift. Verses repeat the same rhyme scheme with no progression.
- Match a specific singer’s timbre at a high level. Voice cloning is a different (and earlier-stage) problem.
- Produce music that is meaningfully new. The output is recombinations of patterns the model has seen, not departures from them.
- Compose, in the sense that a working musician uses the word. There is no draft, no revision, no rejection of an idea. There is generation, selection, and publication.
The lawsuit, in plain terms
The RIAA filed suit against Suno and Udio in June 2024, alleging that the models were trained on copyrighted sound recordings without license. The labels want statutory damages, injunctions, and a per-play royalty regime. The AI companies argue fair use, transformative work, and that the training data was lawfully obtained.
The case has not gone to trial. Suno’s internal communications, revealed through discovery, include engineer messages acknowledging the training data was “just all of the songs” they could find. That is not a smoking gun. It is also not nothing.
What is likely to happen, based on the existing case law and the political environment: a settlement that includes some form of licensing, possibly a compulsory license for training data, and most likely a structured opt-out for rights holders. The labels will get paid. The training will continue. The specific mechanism is the open question.
What it is replacing, and what it is not
The honest version, with sources:
What it is replacing. Stock music for video projects, background music for podcasts, royalty-free libraries, jingles, sound effects, ambient pads, demo recordings, “songwriting scratch tracks” for evaluation. These are areas where the existing supply is fungible, the per-track budget is small, and the human alternative was already heavily produced and unremarkable. The displacement here is real and accelerating.
What it is not replacing. Top-line pop, country, hip-hop, or rock. Stadium tours. Catalog sales. The reason is not that AI cannot produce comparable audio. It can. The reason is that the human artist is the marketing, and the marketing is the product. People buy the song because they buy the singer. AI can clone the voice. It cannot clone the meaning the voice has accrued.
This is also why the artist-vs-AI framing is mostly wrong. The question is not whether AI can do what Beyoncé does. The question is whether the next person who would have become the regional touring act now just generates their own music and skips the touring. The displacement is at the bottom of the pyramid, not the top.
The labor question
The most honest answer is also the most uncomfortable.
Session musicians, jingle writers, and soundtrack composers are the categories most at risk. The work is skilled, the supply is large, the per-project budget is fixed, and the output is judged primarily on quality. None of those attributes favor humans in a generative model world.
What is not at risk: songwriters, especially songwriters with distinctive voices, because the lyrics and the structure are where the new thing comes from. Producers who are also artists. Sound engineers who mix live recordings. Tour musicians, because the tour is a live event and live events are physical.
The legal answer to “who gets paid” is unsettled. The economic answer is: anyone whose contribution is judged by reference to their identity will do fine. Anyone whose contribution is judged by reference to a generic output will not.
What the actual next 18 months look like
Three near-certain developments.
Licensing deals. The major labels will close licensing arrangements with at least one of the major AI music companies by the end of 2026. The terms will be the model. The first deal sets the template.
Watermarking standards. C2PA-style provenance metadata, plus audible watermarks for generated content, will be standard on platform distribution by 2027. The watermarks will not stop infringement. They will make enforcement easier.
Indie tooling explodes. The interesting products are not the consumer-facing Suno-and-Udio clones. They are the producer-facing tools. Stem extraction from generated tracks. Arrangement suggestions. Reference matching. Mixing assistance. The next billion-dollar music company is more likely to be a tool for working producers than a Suno competitor.
The bottom line
AI music is real, capable, and legally contested. It is not killing music. It is killing the parts of music that were already fungible. The work that is left, the work that requires a person with a point of view, is harder to replace, and more valuable, than it was a year ago.
If you are a working musician, the threat is to your day job, not your career. If you are an investor, the threat is to your assumption that the major labels’ catalog moat is durable. It is not. The moat is the brand, the relationship, the tour. The catalog is a depreciating asset, the same way newspaper classifieds were a depreciating asset.
Whether the industry handles the transition well or badly is the open question. History suggests badly, on average. But the new things that get built in the gap are usually better than the old things, eventually.