In one of the examples, they refer to something called "Dream Track"

> Dream Track in Shorts is an experimental song creation tool that allows creators to create a unique 30-second soundtrack with the voices of opted-in artists. It brings together the expertise of Google DeepMind and YouTube’s most innovative researchers with the expertise of our music industry partners, to open up new ways for creators on Shorts to create and engage with artists.

> Once a soundtrack is published, anyone can use the AI-generated soundtrack as-is to remix it into their own Shorts. These AI-generated soundtracks will have a text label indicating that they were created with Dream Track. We’re starting with a limited set of creators in the United States and opted-in artists. Based on the feedback from these experiments, we hope to expand this.

So my impression is they're talking about labeling music which is derived from a real source (like a singer or a band) and might conceivably be mistaken for coming from that source.

Even if it is fully AI-generated, this requirement seems off compared to the other ones.

In all of the other cases, it can be deceiving, but what is deceiving in synthetic music? There may be some cases where it is relevant, like when imitating the voice of a famous singer, but other than that, music is not "real", it is work coming from the imagination of its creator. That kind of thing is already dealt with with copyright, and attribution is a common requirement, and one that YouTube already enforces (how it does that is different matter).

From a Google/Alphabet perspective it could also be valuable to distinguish between „original“ and „ai generated“ music for the purpose of a cleaner database to train their own music generation models?

Alternatively they want to know who to ban when the RIAA inevitably starts suing the shit out of music generators.

> If you manually did enough work have the copyright it is fine.

Amount of work is not a basis for copyright. (Kind of work is, though the basis for the “kind” distinction used isn't actually a real objective category, so its ultimately almost entirely arbitary.)

That could get tricky. A lot of hardware and software MIDI sequencers these days have probabilistic triggering built in, to introduce variation in drum loops, basslines, and so forth. An argument could be made that even if you programmed the sequence and all the sounds yourself, having any randomization or algorithmic elements would make the resulting work ineligible for copyright.

It goes without saying that a piece of software can't be a copyright holder.

But the person who uses that software certainly can own the copyright to the resulting work.

If someone else uses the same AI generator software and makes the same piece of music should Google go after them for it? I don't think that would hold in court.

Hopefully this means that AI generated music gets skipped by Googles DRM checks.

I hope there is some kind of middle ground, legally, here? Like say you use a piano that uses AI to generate artificial piano sounds, but you create and play the melody yourself: can you get copyright or not?

IANAL. I think you'd get copyright on the melody and the recording, but not the sound font that the AI created.