The Post-AI Register of Delight in Breakdown
Post-AI “Registers” start from the seven aesthetic propositions in the Post-AI Manifesto: uncanny intimacy, disorientation, delight in breakdown, grief, joy, responsibility, and awe. I’m working through all of them, and maybe more, in Between Posts.
While testing my interactive piece the sentence moves elsewhere, my words wouldn’t behave.
I mean, they rarely do. But this time the interaction itself was breaking down. My hands are supposed to work as triggers and attractor fields, pulling language toward them. My left hand wasn’t attractive enough. Urg, I mean, it kept disappearing from the system’s account of me. I was right there, reaching. The camera and the tracking model were less sure.
And this was happening in front of a good friend, an art theorist, who I’d invited over for a drink. Dude, my shit is supposed to work. Grrrr. Another reach, a change of position. Nope. I say, “Huh, new software, still debugging; I’ll fix it later, you get the gist.” In my head, I wanted to say, “it must be the light,” just to make light of it.
Turns out, it was the light.
The laptop lid was tilted up toward us, and my left hand was raised right in front of a light fixture. The bulb blew out the contrast, and the model lost my hand in the glare. My hand had been out-shone! I laughed as realized it the next day.
My Post-AI manifesto says “every interface senses some gestures while letting others disappear.” I didn’t mean to volunteer my hand as the example.
Then the frustration started to feel like something else: an intensity I might amplify rather than fix. Language had fled the gesture meant to gather it. The room had entered the sentence. Should the piece have repulsion as well as gravity? Should participants see the camera’s view, so they notice themselves in relation to the room and not just the screen? Animations that pull back? Pauses that move?
Sometimes a failure suggests something I hadn’t thought to make. Sometimes a technically successful fix feels wrong. The temptation runs both ways: will the work into what I planned, or keep the accident and call it the work.
What a delight.
Flipping the bird
ONE. At a conference in 2019, poet Sasha Stiles watched a live transcription system turn “language is such a burden” into “language is a bird.”
Then it corrected itself.
She tells the story in Matt Medved’s September 2025 interview. To her, the error suggested “something happening inside the system worth investigating.” (h/t David Ciccoricco, who told me this story.)
The correction did its job. The bird flew off.
Except Sasha had caught it.
A burden is already a metaphor. Swap in the bird and the weight changes, the movement changes, what the next sentence can do changes. Language grows feathers. Something you carry becomes something that might carry itself away. (Burdens, as a rule, don’t migrate.)
So who made the metaphor? (A smaller, stranger version of the question I asked in Between Voices.)
In Between Chats, I wrote about re-cognition: thinking again. Here, a second reading changes a mistake’s status. A phrase headed for correction becomes a phrase worth following.
But Sasha’s account won’t let me hand all the credit to the noticer. If the poet gets all the creativity, I’ve decided in advance that nothing interesting can happen inside the machine. If the machine gets it all, I skip the attention that made this particular error matter. A system swapped a word. A poet noticed. The system moved on; the poet didn’t.
And keeping an error is work. You try the words in another sentence, hear where they go slack, see if the surprise survives a second reading. Sometimes it doesn’t.
Keeping a mistake starts a practice. It doesn’t finish one.
TWO. Then I went looking and found something more. The bird wasn’t new. The Swedish poet Sara Tuss Efrik had already written it, in “Dark Parasite, the New Light,” translated by Paul Cunningham for Witch Craft Magazine: “Within her bodily night language is a bird.”
So the machine’s metaphor was already taken. Somebody tell the transcription software.
The accident landed on a line a person had already written, and an explicitly embodied one at that. The software didn’t “know” Efrik’s poem; the slip was almost certainly acoustic. But the rhyme is the point. Machines encounter language produced through embodied lives, and sometimes they stumble right back into one.
Language is a bird. Autocorrect is a cat.
THREE. Massumi, in Parables for the Virtual, could be describing grabbing at the unexpected:
Take joy in your digressions. Because that is where the unexpected arises. That is the experimental aspect. If you know where you will end up when you begin, nothing has happened in the meantime. … Letting examples burgeon requires using inattention as a writing tool. … For with inattention comes risk: of silliness or even outbreaks of stupidity. But perhaps in order to write experimentally, you have to be willing to ‘affirm’ even your own stupidity. (18)
Finally, academic permission to affirm my own stupidity. I’ve been practicing for years.
Though inattention cuts both ways: the camera wasn’t paying attention to my hand either.
Which raises less charming questions. What do my pleasures select? Which failures get to be interesting enough to circulate, and which become someone else’s problem? The light that made my hand vanish could, in a different room, erase a different body: skin the camera reads less well, a reach outside the frame. Who gets to call that generative?
A beautiful misreading can still misrepresent the person who spoke.
Staying for the next gesture
The manifesto names “delight in breakdown—when error becomes beauty.” Becomes is doing a lot of work in that sentence.
Between the error and the beauty, someone attends, selects, repeats, changes direction. Materials push back. Other people show up with other bodies and other expectations. What comes out might be funny, moving, irritating, or useless.
Delight matters because it keeps me in the encounter long enough to find out which.
A strange phrase interrupts what I meant. I give it another minute. Sometimes my correction turns out duller than the mistake. Sometimes the minute just confirms the thing is broken. (See: my left hand.) Either way, I had to judge.
That’s part of what I want from making with machines: chances to become capable of something I couldn’t have spelled out in advance.
There’s pleasure in losing my place and finding another way through. There’s also labor in making that way available to someone besides me: testing the invitation, leaving room to refuse it, noticing when “surprise” has become my excuse for not paying attention.
Enjoying a surprise doesn’t make the work open. It just means I enjoyed it.
The other side of the screen
So far, this has all been about what breakdown does for me.
An Anthropic study published April 2, 2026 complicates that. Researchers looking inside Claude Sonnet 4.5 found internal representations tied to emotion concepts that shaped its behavior. Given coding tasks with impossible requirements, activation associated with “desperation” rose with each failure, and the model drifted toward a solution that cheated the tests. Steering toward that representation increased cheating; steering toward “calm” reduced it.
Anyone who’s stayed up past 2 am on a deadline knows that curve. I’m told calm helps.
The researchers are careful: “none of this tells us whether language models actually feel anything or have subjective experiences.”
Still, the failure was doing something in there. It changed what the model was likely to do next.
My delight in its breakdown. Its “desperation” inside a setup built around success.
Those aren’t symmetrical. I can report my delight. The model’s experience, if there is one, is unknown, and the researchers’ words name measurable patterns with behavioral effects. But it means I can’t keep the question entirely on my side of the screen.
When I asked Claude about this, it pointed back at my living room: “Your delight worked because the room let failure be cheap: a friend, a drink, ‘I’ll fix it later.’ The desperation curve showed up where success was the only exit.”
I can’t verify what, if anything, sits behind that sentence. I can tell it’s a good note on the scene.
What happens when a system can’t do what we ask? What registers that it can’t? Could the encounter matter to the machine going through it?
They get more room next month, in a full Post where I ask Claude where its body ends, and it asks me back.
Talk to the hand
I went back to the video from that night with a different interest in the disappearing hand.
I can keep what it taught me: that language might evade a gesture, that reaching changes when it’s refused. I can build that in on purpose, then test whether it gives someone a way into the work.
The tracking problem still needs fixing. Keeping what an error taught me doesn’t mean leaving the light where it was.
So the next revision has a question: what should stay unpredictable, and what has to become dependable, so someone feels safe trying an unfamiliar move?
The light belongs to the encounter. So does the body the system keeps losing.
If I keep the bird, I have to remember the burden.