Two Shell will release their album Infinite Now on 16th October, via Young. In June, an EP appeared, Infinite Now+, a five-track Y2K Euro-club sugar rush. By early July it had grown into 10 songs. Right now, looking at Apple Music I see two Infinite Now+ releases, both with different versions of the songs, 20 songs in total. With an upcoming unreleased Infinite Now LP. So.. is the confusion of the releases part of their brand as a band? Like leaking their self titled album by scattering around USB flashes on the Boiler Room floor. Or their fake "MP3 player" at shell.tech that felt like you'd stumbled into their personal drive of rough demos. Or a 2022 Boiler Room set they may have sent body doubles to. Or using AI to generate their music? It begins to look more like trolling.

In their first ever video interview with residentadvisor in July, they spoke about using AI, something that everyone’s been speculating about. The interview itself is shit, and Tom Breihan’s opinion piece on Stereogum isn’t hiding his frustration about it. If you’re going controversial you can’t be sloppy, and the band saying “there's an irony with it where I think sometimes it feels like the emptiness of it almost helps you identify the soul in it” is a sloppy and uninspiring quote. What’s even sloppier is comparing the pushback against AI to how musicians’ unions in the 70s objected to synthesisers.

Is this performance art? In a way, their EP sounds fresh, and references exactly what the sound is going for–90s R&B, dance and powerful vocals with a trashy, AI-generated glittery cover, like the ones that were sitting on my dad’s CD shelf in the 2000s. On a recent, not-often-occurring visit to the mall I heard a song that was abysmal in every way–nonsensical lyrics, over a saturated melody with too-excited yet flat vocals. Obviously AI, and an insult to the people shopping there. /Also, bad ambiance vibes for the store/. But the thing with Infinite Now+ is, it is a decent EP. It sounds well produced, and I highly doubt they prompted Suno into an EP of songs with no iteration on their side. They are good producers, their Icons EP from 2022 is a testament to it. They’re proving you can create decent music without a human.

So where are they going with this? I don’t know, the interview doesn’t make it clear, it is mostly a scramble of musings and an undefined, uninspiring direction. The songs are fun, sure. But in no way “soulful”. If you’re making a point, make it well.

I’m thinking about this line by Brian Eno, which in the context of AI might be ironic, because he is the person who spent 40 years handing off control to generative systems, records designed to be ignored, and the creator of Oblique Strategies. If I can think of a person who would sign off on an album made by a GenAI, it would be him. Two Shell’s comparison about the unions against synthesisers–Brian Eno won it, the synthesisers prevailed.

The line is from a diary entry from 1995, where he says the ugly parts of a new medium become its signature. The distorted guitar is something too loud for what's carrying it. The blues singer's cracked voice is a cry too big for the throat. The grainy film, bleached-out black and white, events too momentous for the medium assigned to record them... He ends it by saying that when a medium fails conspicuously the listener starts believing something is happening past its limits.

Two Shell want to make the same profound point about AI. But nothing in their EP exceeds anything. There’s no throat, nothing too loud, nothing too momentous. It isn't a medium giving way under something too big for it, it's a medium doing its job with nothing on the other side.

There is though, Vegyn’s project Headache, which I find wonderful, no matter the AI in it. It is a project I could call “soulful”. It is a trip-hop/downtempo record, with a spoken word, synthetic voice over the songs. The description on Bandcamp reads: it's produced by Vegyn, with all lyrics written by Francis Hornsby Clark, and then performed by AI. When the debut The Head Hurts but the Heart Knows the Truth came out, fans speculated that not only was the voice AI-generated, but the lyrics as well. There was no trace of a Francis Hornsby Clark online. Vegyn had to publicly confirm the lyricist is a real person, a low-profile friend who doesn’t care about social media. And although the debut was a total hit, Vegyn himself finds it totally unprecious, a record made in two weeks and "actively a joke". The point is, the AI voice is not trying to pass as something more than what it is, it could be anyone’s and everyone’s. The narrative of the lyrics is a long monologue, with no refrain or repetition. My favourite lyrics come from Nineteen Sixty Five, the opening song of the second Headache album, Thank You For Almost Everything:

And if you're still here, remember, I'm not real
And I'm sorry to break it to you
But if that ever mattered to you at all
It was over for you from the start

Holly Herndon was the first to license and build an ownership structure around her own voice AI model, which she named Holly+. She launched it in July of 2021 with Never Before Heard Sounds. You upload an audio file and get it back sung in Holly’s voice. A free tool, and open to anyone to use. She gave away the governance of her AI voice to a Decentralized Autonomous Organization /DAO/, a closed community of friends, collaborators and NFT supporters, and now an artefact of the crypto empire. The collective had the rights to the voice model, and as such the ability to approve and license her vocals. Approved sales split 50% to the creator, 40% to the DAO, 10% to Herndon. In a post announcing the project, she wrote: “In the event of my death, I feel more comfortable with distributed ownership of the rights to my voice model among a DAO of stewards who are invested in maintaining the value and reputation of my voice than the rights being passed down to someone less familiar with the values and standards associated with my work.” In a time of AI, where you can bring back Michael Jackson’s vocals in songs, she is asking an important question–who speaks for this voice when you’re no longer here?

Grimes followed her a couple years later, the first musician from a major label to do so. When the fake Drake/Weeknd track Heart on My Sleeve went viral and got pulled in April 2023, her response was to open-license her own voice rather than fight copycats of it. She posted that she'd split revenue 50/50 with anyone making a successful AI song using her voice, and said she wanted to kill copyright. A week later came Elf.Tech–a free tool, where you can upload an a cappella and get back a stem with her vocals. Her framing is that art is a conversation predating us, that entangling it with ego is a modern invention, and that an industry defined by lawyers kills creativity. But unlike Two Shell where they have used a plethora of voices, with no clarity on what or who they're trained on, Grimes is selling her own voice, and with a condition. She said she'd take down anything promoting hate speech or extreme violence. Unlike Herndon, who handed the decision of who gets to license her voice to a DAO with a financial stake in it, Grimes gets half the revenue while keeping the veto to herself. When in reality, in both cases no matter the licensor, someone else did the writing and production. What happens when more famous voices get trained the same way, like a toll booth generating the cash, with no human to perform it live, a fictitious “soul” behind the vocals?

In both cases, Holly and Grimes are using their voices as an asset. Vegyn on the other hand uses nobody’s voice, more like a text-to-speech generator than anything else. Who are the voices behind Two Shell? Unspecified. A dubious software, generating music with questionable licence over what that model is being trained on. In an interview for SciAm in March of this year, Holly had another observation to share, which I agree with: “One criticism of products [like Suno] is that they’re very ‘mid’ sounding—trained on everything or the most average. My models sound unique because I’m making the training data myself.”

I can relate. We’re using a shit ton of AI at work, to ideate and design. And it is an uninspiring process, with uncanny results. And at the end, no matter how good your prompt is, it all kind of has the same muddled look. The ceiling is already set by the model you’ve chosen. No Ghost, Just a Shell. ¹