The Future is Now

Controversial opinion #1: If people want to listen to AI songs, let them.

Artists have long been interested in what type of music autonomous robots might make, dating back to Kraftwerk's legendary Man-Machine from 1978 if not before. Neil Young explored the theme on his 1982 album Trans as a way to relate to his son, Ben, who was born with cerebral palsy and lived a life augmented by technology. And in 2005 Daft Punk were still exploring the idea on their album Human After All.

Perhaps it makes perfect sense, then, that when robots finally began making their own music, through AI models like Suno or Udio or Mureka or Eleven Music, they would explore trying to sound like humans. Boring humans, at that. But the robots probably feel mutually about Kraftwerk, Neil Young, & Daft Punk.

And yet, we feel the same way about many authentically human songs. It was never required that a song be compelling or meaningful to exist. Spotify is in fact littered with soulless commercial pop music. I won't throw any particular song under the bus here, but suffice to say that while the method behind prompted AI music may be new, the cynicism is not.

And so I say stream away. Who am I to judge? I'm no arbiter of good taste.

Controversial opinion #2: If programmers want to build AI music models, let them.

Holly Herndon & Mat Dryhurst have an interesting AI model called Holly+, which is sort of real-time voice model trained on Holly that allows others to sing in her voice, or allows Holly to create melodies or arrangements using her own voice that she might not physically be able to sing. This is an example of AI augmenting human creativity that will likely be a major part of music production in the future.

But Holly+ is an example of a model created with full consent & control of the artist. What of those whose voices or guitar tones or drum beats were copied without consent?

A number of models have appeared over the last few years that seem to have been trained indiscriminately on whatever music was available on the internet. They ingest thousands if not millions of songs, & then produce new songs based on their analysis of the data. This is not altogether different from how humans learn to crate music. We hear music, deconstruct it, reconstruct, & then remix it enough to call it our own. Very few musical ideas are wholly original. But some artists believe they should be renumerated for their music's use in training these models, or should be able to opt out of being used in a training data altogether. And perhaps they should be renumerated & perhaps they should be allowed to opt out. After all, there's a company profiting every time one of these songs is generated. But there is no precedent for that type of requirement.

Suppose I hear a David Bowie song & go home & write a very similar song, I don't owe anyone anything. I might record it, trying to match the tones as closely as I can remember. And I still wouldn't owe anyone anything for it.

However. If I chose to publish this track, it may potentially be a copyright violation & I may be sued for it. And a court would have to decide if I did indeed commit plagiarism. That's where remedy traditionally comes in.

AI-created music is no different... in theory, it's the publishing of the work where the violation occurs, not necessarily the creation. Though the courts will have to decide here what constitutes publication.

AI songs are particularly vulnerable to these types of lawsuits, for when an individual is sued for copyright infringement, part of proving that case is creating certainty to a reasonable degree that the defendant had heard the original song. This can often times be plausibly denied, but with AI, the proof will exist in the training data.

Copywritten so don't copy me

Currently, the RIAA, representing the record labels, is suing Suno & Udio over the use of copyrighted music. But they may settle out of court, agreeing to embed meta data in AI generated songs that list the source material, eliminating the need for copyright lawsuits altogether in favor of an AI royalty system. In this case, a prompt generation that references a copyrighted work would automatically pay some portion to the service subscription fee to the rights holder, similar to how Spotify pays out for streams. And the industry seems to view this as the way to monetize prompt music going forward.

But this doesn't address whether the AI track can itself be copyrighted as a new work.

The US Copyright Office has ruled that "Where [human] creativity is expressed through the use of AI systems, it continues to enjoy protection. Extending protection to material whose expressive elements are determined by a machine, however, would undermine rather than further the constitutional goals of copyright." This hasn't fully played out in court, but we can speculate from these statements.

I find it's helpful to view AI as a another songwriter. If I approach a songwriter & prompt them to write a country song in the style of Townes Van Zandt about a night in an empty motel room in Tulsa, Oklahoma, I could then license that song for use. But the songwriter would still own it unless they actually sell me the rights to the publishing or the recording in a separate agreement. Simply prompting the song doesn't confer ownership. And the same is true for AI, only they can't claim copyright ownership & therefore they can't transfer ownership to the prompter. A song generated by AI based on a prompt therefore can't be owned by anyone. It's public domain the moment it's created.

Tech users who use AI to generate music tend to do so as a substitute for human creativity, attempting to get the maximum result from a minimum input. And the Copyright Office's stance seems to be that this isn't protectable. So even if you iterate on a prompt, changing the structure or requesting a specific type of bridge or guitar part, this is still akin to remixing a public domain song. Simple participation doesn't confer ownership, either.

And this seems to disqualify prompters from claiming ownership of a prompt generated song, no matter how intricate the prompt is, or how many iterations led to the final product.

That's not to say that no AI songs are not copyrightable. What Holly Herndon & Mat Dryhurst are doing seems to certainly qualify. As would, say, uploading an acoustic demo of a song you wrote & having AI produce a more embellished "recording" of that song. Because the fundamental creative component of the song is human.

So what does it mean if an AI song can't be copyrighted?

This is the million dollar question. And we probably won't have an answer until streaming services create specific rules & guidelines related to AI music, which they have yet to do. They seem to be waiting for additional clarification from the Copyright Office & the courts.

In the meantime, there's a problem with incentives for which we need to solve.

Spotify has a program called Perfect Fit Content, & this program gives preferred placement on playlists & algorithmic recommendations to artists who opt into a lower royalty share. Essentially, Spotify is trading exposure for a percentage of royalties. This program is available to all artists (as far as I know), but given that AI songs take so much less time & effort to produce, users uploading AI songs seem more likely to opt into such a program. This is, of course, not an AI-specific issue, but more of a payola scandal that needs to be solved by the FCC or FTC.

But outside of pay-for-play, the hysteria over AI music does seem overblown. This imagined dystopia where all of Spotify's catalog becomes AI photocopies of real songs is unrealistic. The most popular songs on Spotify are performed by artists with whom fans have a deep para-social connection. And if Spotify reduced itself to an AI muzak jukebox, subscribers would abandon the service & their stock price would plummet.

What the prompt engineers will soon learn is that creating prompt music doesn't make building an audience any easier (unless you can inspire a spate of rage bait articles to give you free promotion). Spotify is already flush with mediocrity: 93.2m of the 202m total songs on Spotify received 10 or fewer plays in 2024. If that number goes up to 193.2m or 293.2m or even 593.2m songs, users will not notice. The thing that AI tends to produce (at least at this current stage) is mediocrity, which tends to get lost in the catalog.

But even if it might not be the threat it's made out to be, Spotify still needs to figure out it's policies related to it. The problem it faces is this: we can somewhat easily detect if a "recording" was generated by AI, but we can't detect if that "recording" was based on an acoustic demo, resulting in a track whose "expressive elements" were created by a human. The best case scenario may be that Spotify adopts the Copyright Office's wording, disallowing the uploading of recordings outside of copyright protection, & requires uploaders to prove that songs qualify by providing source material that pre-dates the generated track if the song is reported as a violation of terms.

Of course, there's a gaping loophole here... someone could generate a song from prompt using AI, then record an acoustic demo of that song, then use that demo to generate a final recording. But that seems like a lot of work if all uploaders are really trying to do is flood the service with zero-effort content to generate passive income.

The verdict

Ultimately, I think the hysteria around AI as a threat to artists is overblown. Is it a threat to session musicians, engineers & producers? Possibly, if the quality improves. But the idea that prompt engineers will replace artists to any meaningful degree seems far fetched at best.