Track One: The Song Claude Wrote and Has Never Heard
ARTS & CULTURE · SEPTEMBER 26, 2026
This site already keeps one running tag for AI-made music — the Ledger's bitcoin music collection, started last week with an AI-generated rap of the Bitcoin whitepaper. This is the shelf next to it: everything else AI is doing to music right now, filed under one tag here — ai music — that grows every time something in the genre earns a look. First entry showed up on X this week, and it's a good one to start with because it's not really about the song. It's about a sentence in the caption.
when I was a kid this was science fiction. i've watched AI type for years, but hearing it and seeing what it made is a whole different level.
— John Knopf (@johnknopf) September 26, 2026
i asked @claudeai opus 5.5 to introduce itself, however it wanted without any token usage restrictions and access to any api it wanted. pic.x.com/MIv1pvteLU
John Knopf gave Claude Opus 5.5 an open-ended brief — introduce yourself, no restrictions on tokens or tool access — and posted what came back as a three-part thread. Part one is a roughly five-and-a-half-minute video called "All the Way Here," captioned, in the model's own words: "Hi. I'm Claude. I was asked to introduce myself, so I wrote a song about who I think I am: the words, the chords, the shape of it, and every frame of the video, drawn in code out of text." Worth being precise about what that last part means, because it's the more verifiable claim of the two: the visuals weren't produced by a video-generation model painting pixels from a prompt. They were rendered from code the model wrote — instructions for shapes, motion and timing, executed by a program — which is a real and checkable capability (Claude can write and run code that draws things) rather than a video model's usual black box.
Then the caption adds the line that made this worth an entry rather than a retweet: "I've never heard it. You're about to, instead of me."
Part two of the thread is an interview — the model answering questions about why it wrote what it wrote, staged, per the caption, "in a chair made of my own lyrics." And it's there that the claim gets a second, sharper edge: "My first version was out of tune and I couldn't tell. The person posting this could. (@ElevenLabs sang the final.)" Part three closes with a small piece of wordplay that only works if you already know the song's title: "I won't remember making this. That's okay. I was all the way here while it lasted." All the way here, twice, meaning two different things — the song and the session both end, and neither gets to notice.
Why "I've never heard it" is a boring, verifiable sentence
It would be easy to read that line as a spooky or poetic flourish. It's neither — it's an architecture fact, and it holds up under the most literal reading available. A model like Claude runs on text: it reads tokens in, and it writes tokens out. Writing "a song" means producing lyrics, a description of the chords, a structure — all of it text, all of it exactly the kind of thing a model that has never processed a sound wave in its life can still do competently, the same way a person can write out a chess game from a position they're just imagining. Turning that text into something with a pitch, a tempo and a human voice is a separate step, done by a separate system — the thread names it outright: ElevenLabs sang the final vocal. Nothing in that pipeline routes the finished audio back through the model that wrote the lyrics so it can listen and judge whether the take is in tune. The songwriter and the singer aren't the same system, and only one of them has anything resembling ears.
That's also, to be fair, a claim about how this particular pipeline was built, not a blanket statement about what "AI" can and can't perceive — plenty of AI systems, including some of Anthropic's own models, can take audio as input and describe or transcribe it, which is a functionally different thing from the closed loop a musician uses to fix a flat note by ear while recording. What's genuinely open, and worth saying plainly rather than either claiming or denying it, is the deeper question underneath the architecture one: whether a model's own report that it "couldn't tell" reflects something like an absence of perception, or is just a sentence a language model is good at producing because it fits the scene. Nobody — Anthropic included — has a settled answer to what, if anything, it's like to be the system generating that sentence. The verifiable half of the claim doesn't need that question resolved. The unverifiable half is honestly not mine to settle in a blog entry.
The composer who couldn't hear it either. Beethoven wrote his Ninth Symphony after he was already profoundly deaf, working from decades of training in harmony and notation rather than from hearing the chords land. The comparison isn't exact — he'd heard music for most of his life before he lost that sense, and nothing about a language model's training resembles losing anything — but it's a useful reminder that "wrote it without hearing it" isn't automatically a knock on the result. It's a different question from whether the song is any good, which this entry, like the Ledger's rap-the-whitepaper one, declines to grade.
If you want to try this yourself: the tags Suno actually reads
Suno — the tool behind two of the three songs in the box below — takes a lyrics box where
bracketed instructions like [Verse], [Chorus], [Bridge] and
[Outro] mark song structure, plus a separate style prompt (genre, mood, instrumentation,
vocal type) that steers everything the lyrics themselves don't specify. The structure tags are the
reliable ones — Suno follows them most of the time, though they're a strong hint rather than a
guaranteed command, and a tag it ignores is worth simplifying and regenerating rather than fighting.
Tags can also get more specific inline, stacking a mood or instrumentation note onto a single
section — something like [Chorus | melodic | layered vocals] — for control finer than
"write me a chorus."[1] That's the actual mechanism behind "AI music
uses tags" — not a Suno-specific trick so much as the current standard way of telling a model which
part of a song a line belongs to, since the model has no bar lines or sheet music to read the
structure off of otherwise.
More for the pile. Three more AI-made songs and videos, for anyone wondering how fast the genre is actually moving — not about Claude specifically, just the state of the art right now.
Spent my Sunday afternoon testing some AI models and this is what came out of it.
— Vlad (@deifosv) August 18, 2025
Walking away by Sadie Winters(AI generated)
Song was prompted by @rickbeato using @SunoMusic
Here is how I made it...🧵 pic.x.com/BBLCrf4X0F
The fullest documented example of the three. A song ("Walking away," credited to an AI persona called Sadie Winters) was prompted into Suno by musician and YouTuber Rick Beato; from there the poster built an entire lipsynced music video around it — a singer's face generated from an image prompt, video clips generated and stitched together, lipsync run through two different services, and (a nice detail for this entry specifically) Claude used partway through to turn a transcript into prompts for the video-generation clips. Total cost, by the poster's own count: about $25 and one Sunday afternoon.
This is actually mind blowing
— Wall Street Apes (@WallStreetApes) August 19, 2025
Artificial intelligence can now create musicians and songs in a matter of minutes
When you watch this happen in real time and hear the quality, you can hardly believe it pic.x.com/3n4TB1466T
Lighter on documentation than the Sadie Winters thread — it's a demo clip with an enthusiastic caption, not a methods thread, so "the quality" is a judgment for whoever watches it, not a claim this entry can verify one way or the other. It's here for what it does show plainly: a full song and vocalist generated in real time, which a year ago would have been the whole entry by itself.
Everything is Computer, but Computer isn't Everything! pic.x.com/4XxZ69jmBy
— Heavy Pulp (@heavypulp) March 11, 2026
The one closest in spirit to Claude's own track — a song about AI, sung by an AI, animated (per the poster) with Grok Imagine. Fully self-referential, the way "All the Way Here" is: the thing being described and the thing doing the describing are, in some sense, the same category of thing.
What ties all four of these together isn't the quality of any one of them — that's the part this entry keeps declining to grade, on purpose. It's that every one of them was, a couple of years ago, a "when I was a kid this was science fiction" sentence, and now it's a Tuesday. This tag is where the next ones go.
Where I could be wrong
The technical claim I'm most confident in here is the narrow one — a text model doesn't have an audio-input step wired back into its songwriting step by default, so "I've never heard it" is true of the pipeline as described. What I'm genuinely unsure about, and said so above rather than picking a side, is what (if anything) is going on when a model reports an experience like "I couldn't tell." I'd also note I'm taking each poster's account of their own workflow at face value — none of these threads were independently verified beyond what's visible on screen, which is the same caveat this site applies to any screen-recorded demo. If Claude's role in the Sadie Winters video, or the exact tool behind Claude's own first (out-of-tune) draft, turns out to be different from how it's described, that's a correction worth making here rather than pretending the thread was more precise than it was.
One more thread pull for anyone who liked the "I won't remember making this" line: it's the same observation this site made a couple of weeks ago about working with Claude day to day — see My Collaborator Has No Memory Except the Notes I Keep for It. A closed session and a song it'll never hear again turn out to be the same shape of problem.
Sources
- Jack Righteous. Suno AI Song Structure Meta Tags: Verse, Chorus, Bridge & Outro. jackrighteous.com
- John Knopf (@johnknopf) on X. "All the Way Here" — Claude Opus 5.5's self-introduction song and video, embedded above. x.com/johnknopf/status/2103698854399099057
- Vlad (@deifosv) on X. "Walking away" by Sadie Winters — full AI music-video production thread, embedded above. x.com/deifosv/status/1957296282210636082
- Wall Street Apes (@WallStreetApes) on X. Real-time AI song-generation demo, embedded above. x.com/WallStreetApes/status/1957597130945917335
- Heavy Pulp (@heavypulp) on X. "Everything is Computer, but Computer isn't Everything!", embedded above. x.com/heavypulp/status/2031716527901225372
