AI Music Digest

Google announced on Friday, September 4 that Lyria 3.5, its music generation model, is now available in the Gemini app to all users globally on web and mobile, and to developers through the Gemini API and Google AI Studio. It also reaches Google Vids, the company’s Workspace video tool. The model itself is not new — Google launched it inside Flow Music, its dedicated music product, on July 29 — but until now it lived in a place you had to go looking for. In its announcement, the company calls Lyria 3.5 “our best-sounding music generation model,” citing more expressive vocals and richer arrangements, and describes the Gemini experience as guided creation: pick or describe a genre, choose vocal or instrumental, start from a template, and choose a short or longer track.

The developer surface is more specific. Per PPC Land and Unite.AI, both of which read the API documentation, there are two variants: Lyria 3 Clip, which returns a fixed 30 seconds, and Lyria 3.5, which returns multi-minute structured songs with verses, choruses and bridges. Output is 44.1 kHz stereo, prompts can include up to ten images alongside text, and lyrics can be supplied with section tags. Every generated track carries SynthID, Google DeepMind’s inaudible watermark. Safety filters block requests for specific artist voices and for copyrighted lyrics.

Two caveats belong next to the announcement, neither of them Google’s framing. The first is legal: Google says Lyria is trained on material it and YouTube have the right to use, but that position is being tested in court. Independent artists sued Google in the Northern District of Illinois in March over Lyria’s training data, and in June Google moved to dismiss on the theory that uploading to YouTube already granted it a “worldwide, non-exclusive, royalty-free, sub-licensable and transferable licence” to prepare derivative works — an argument Complete Music Update summarised as YouTube uploads being fair game with nothing musicians can do about it. That motion has not been decided. The second is regulatory: PPC Land argues the rollout opens a gap under the EU AI Act’s Article 50, in force since August 2, which requires providers to mark synthetic output machine-readably but requires deployers to disclose it visibly. A brand jingle generated in Gemini arrives watermarked; whether the business that publishes it knows it must also label the thing is another matter.

Why It Matters: Suno and Udio require a decision to go and use them. Gemini does not. Putting complete song generation into a mainstream assistant changes the size of the funnel rather than the state of the art — and does it while the lawfulness of the training data is still before a judge.

Timeless, released August 28, collects ten previously unheard Prince recordings made between 1977 and 2016, ten years after his death. Within days, fans reading the credits found something that did not line up on “Calabama,” a track recorded in 2003: the horn players listed on it were not thought to have worked with Prince until roughly 2012. The conclusion a good part of the fanbase reached was AI.

On Thursday, September 4, NPG Records and the Prince management team answered on Instagram. Their account, as reported by NME and MusicRadar: at the bridge of “Calabama,” Prince’s vocal calls out twice for “horns,” but the archived multitrack has a synthesizer filling that space. Horn players who had worked with Prince were assembled to overdub the synth part with the intended line — trombone from Greg Boyer, saxophone from Adrian Crutchfield, trumpet from Lynn Grissett, all three credited in the album packaging. “All of Prince’s recorded parts remain intact,” the estate said, and no AI was used on any Prince music, on this track or elsewhere; no other song on the album received a posthumous overdub. The team said it plans to “provide even more details regarding the release of the music” going forward, defended releasing unfinished vault material at all, and asked fans to raise concerns with “kindness and respect.”

The sequence is the interesting part. MusicRadar reports that Boyer had already said on Facebook that the horns were “recorded a couple of months ago in my basement studio,” and that estate representative Charles Spicer Jr. initially told fans nothing had been “added, nothing was edited” before correcting himself on X to say that assertion “was not accurate.” So the substance of what fans objected to — a posthumous addition, disclosed only after they found it — turned out to be real. It just was not AI.

Why It Matters: AI has become the default explanation for any anomaly on a posthumous release, which means estates now face an evidentiary burden that did not exist five years ago. Here the suspicion was wrong on the mechanism and right on the instinct: something was added, and nobody said so first.