AI music news, week of 1 Sep 2026: fair-use training, collecting-society suits, and product changes

U.S. DOJ says training is generally fair use (1 September)

In The New York Times’s copyright case against OpenAI, the U.S. Department of Justice filed its first statement of interest, arguing that training LLMs on copyrighted text is not a copyright violation. It is not a binding judgment. The brief calls training “extraordinarily transformative,” says a training copy does not itself substitute for the original in the market, and treats training and outputs as separate questions.

The filing is about text; music is not named. Music cases against Suno, Udio, and Anthropic still turn on whether training is fair use, so the brief may be cited there. For makers, whether a model was trained on a work and whether you may commercially use an output are different layers. The latter still depends on each service’s terms and on the license you can actually grant.

Canada’s SOCAN sues Suno over outputs (2 September)

SOCAN, Canada’s collecting society, sued Suno in Federal Court. Where Europe’s GEMA and Koda have targeted training inputs, SOCAN challenges generated tracks and the act of streaming them in Canada. It lists 150 outputs it says are identical or substantially similar to works in its repertoire, seeks statutory damages of up to CAD 20,000 per song, and also claims CAD 10 million in punitive damages.

The fight is widening from “was it trained?” to “is the output too close to a hit?” and “may it be streamed?” When similarity of outputs is the issue, prompting toward famous songs or writers is riskier than before. For commercial BGM, games, and ads, avoiding artist imitation and using material with clear rights is the safer path.

Jason Isbell and others sue Suno over identity and voice (filed 31 August)

Musicians including Jason Isbell and David Lowery filed a proposed class action in Massachusetts federal court. It is not a copyright claim. It alleges commercial use of identity—name prompts that call up voice, style, and image—and the capture of voiceprints under Illinois’s BIPA. The theory is that a label license with an AI company does not, by itself, cover a performer’s right of publicity. Suno says the claims are without merit and intends to defend.

The practical takeaway is clear. Putting a track generated as “in the style of” or “in the voice of” a named artist straight onto monetized content carries a litigation risk apart from copyright clearance. Material with published commercial terms, such as Kominami, or a workflow where you hold the rights, is easier to explain for streams and client work.

Google opens Lyria 3.5 to every Gemini user (4 September)

DeepMind’s Lyria 3.5 music model is now available to all Gemini app and API users worldwide. Google says it improves vocal expression and arrangement, and adds genre control, vocal or instrumental, short or longer tracks, and templates. It is also available in Flow Music, AI Studio, and Google Vids.

Google’s music-generation documentation says generated audio includes a SynthID watermark. Big-tech music generation is moving from an experiment to an everyday tool. More watermarked output also makes it easier for platforms and charts to identify AI audio. Anyone using it still needs the product’s commercial terms and the rules of the destination (YouTube, a store, ads) checked together.

Suno applies download caps and a new Terms of Service (from 3 September)

From 3 September, Suno caps file downloads on paid plans (Pro 20 / Premier 60 a month). The free tier has seven lifetime trial downloads. Premier use inside Suno Studio is unlimited, and extra downloads can be purchased. A new ToS took effect the same day, including commercial-use language and updated arbitration terms. Suno also previewed new industry-partnered models and said older models will later retire, with existing songs remaining in the library.

This is the change that hits export-heavy workflows first. Anyone batching large monthly downloads needs to revisit that process, and commercial terms should be re-read in the ToS itself. Plan numbers and why Kominami still takes files from the library rather than Studio are in the operations note “Suno's September download caps, and why Kominami is not using Studio.” Catalog libraries and already-exported collections become relatively more useful each time a generator adds an export cap.

The one point for makers this week

The lawsuits will take time. Three things already matter in production:

• Do not prompt toward famous artists, voices, or songs (non-copyright claims are growing)
• Re-check each tool’s commercial terms and export caps before you publish
• Treat “was training lawful?” and “is my release safe?” as separate questions

Where Kominami sits

Kominami publishes AI music that may be used commercially within the site license. See the on-site license page for the actual terms.

This article summarizes public information. It is not legal advice. Get professional advice when a rights decision is required.