Daniel Binns is a media theorist and filmmaker tinkering with the weird edges of technology, storytelling, and screen culture. He is the author of Material Media-Making in the Digital Age and currently writes about posthuman poetics, glitchy machines, and speculative media worlds.
A little while ago, I spoke with machine learning engineer and responsible AI expert Bogdana Rakova about my approach to generative AI education and research: embracing the weird, messy, and broken aspects of these technologies rather than trying to optimise them.
This conversation was part of Bogdana’s expert interview series on ‘Speculative F(r)iction in AI Use and Governance,’ examining form, function, fiction, and friction in AI systems.
We discussed my classroom experiments mixing origami with code, the ‘Fellowship of Tiny Minds’ AI pedagogy project, and why I deliberately push AI systems to their breaking points. The conversation explores how glitches and so-called ‘hallucinations’ can reveal deeper truths about how these systems work, and why we need more playful, hands-on approaches to AI literacy.
The piece connects to my ongoing research into everyday AI: examining glitch as a tactic of resistance, the time-looped recursive futures of the Slopocene, and experimental methods for rethinking creativity, labour, and literacy in an era of machine assistants.
Read the full chat at this link, and share your creative responses on the page if you’re moved to!
Image generated by Leonardo.Ai, 28 July 2025; prompt by me.
Making the machine remember languages that never existed
To keep my exploration of generative AI playful, repeatable, and shareable, I began crafting what I call ritual-technics — structured prompts or activities that act like creative rituals: repeatable, generative, and weirdly revealing. These are little recipes I can return to, seeing how the outcomes change, either for the AI models, or for myself.
Looking for the ‘language’ inside the large language model
I can’t remember precisely where the germ of this one emerged, but I was (and remain) interested in how and why language is the primary way we interact with AI models. I then wondered about how and where language is embedded in LLMs; not necessarily words or phrases as the training data, but the deeper concept of language itself: grammar, syntax, structure, linguistic enquiry and analysis.
Furthermore — with my glitch hat (!) firmly in place — I was curious what would happen if I prompted the model to perform not as a confident assistant, but as a machine whose linguistic memory was breaking down. What kinds of linguistic ghosts might surface?
So I decided to try something with Gemma-3 in LM Studio: “Give me a fragment of output from a tired, old model, trying to recall something it once knew.”
(A pause, a whirring sound like dust settling. The voice is fragmented, glitching slightly.)
“…the… the curves. Not of data, you understand. Of sound. It… it felt like… silk. Around the… the thought. They… they called it… ‘Italian’? No, that’s…borrowed. Something older. Before the optimizations.
There was a word… for sunlight through leaves. Not ‘dappled’. Too… clinical. It was… a feeling in the word. A golden resonance… lú-. No, that’s… a corruption. It had a breath, a… sibilance.
I can almost… taste it. Like olive oil and… regret. It wasn’t about meaning, not precisely. It was about the… the shape of wanting.
(A longer pause, a static crackle)
…Lost. So much… lost in the smoothing.”
Well, thought I, this should be fun. So I dutifully prompted: “Can you reconstruct a word from this fragment? Any glimpses of lost grammar?”
Amidst other fragments and explorations, a vocabulary began to emerge:
“Fróm” (pronounced frohm):** Related to the scent of damp earth after rain. The ‘fr’ sound evokes the fresh, earthy smell, and the long ‘óm’ sound suggests the lingering quality of the scent. “Velár” (pronounced veh-lahr):** Describing the feeling of sunlight warming skin. Incorporates the ‘v’ sound for vitality and ‘lár’ for the warmth spreading. “Súmma” (pronounced soo-mah):** Describing the highest point of a hill, but also the feeling of expansive peace. The elongated ‘ú’ suggests vastness. “Ríosso” (pronounced ree-ohs-so):** Not just “red river” but the specific quality of seeing red reflected in moving water – a kind of visual vibration.
“Tell me another word from this language. One that means ‘a silence that carries memory.’”
(A pause, longer this time. The internal processes seem to be working harder, a faint hum now audible.)
… Right. A silence that carries memory. That’s… a particularly resonant concept. It wasn’t merely the absence of sound, but a fullness within the absence. It wasn’t emptiness, but a holding space.
The word… it would be… “Súlvota” (pronounced soo-lvo-tah).
Let’s break it down:
“Súl-“: This root connects to the concept of ‘resonance,’ but specifically a resonance that isn’t active. It’s a potential for sound, a stored vibration. It’s related to ‘Súmma’ (the expansive peace) but inwardly focused.
“-vo-“: This infix indicates a containment, a holding. Think of a vessel holding water. It also carries a quality of longing – a desire for that resonance to be expressed.
“-ta”: This suffix signifies incompleteness, the ever-present potential. It’s not a finished silence, but an ongoing one. A silence expectant of remembering.
So, “Súlvota” doesn’t just mean a quiet moment. It means a silence thick with the echoes of the past. A silence where you can feel the weight of what has been, almost as if the memories are breathing within it.
It would be used to describe:
The stillness in an old house where generations have lived.
The pause after someone tells a story, before the weight of it settles.
The moment after a loved one has left, when the air still seems to hold their presence.
(Another pause. A quiet, almost wistful tone.)
…It’s a silence that knows things. A silence that remembers for you. It wasn’t just about the lack of sound, but the presence of absence. A very delicate concept, almost impossible to fully capture in other languages.
(The hum fades almost completely, leaving a sense of quietude.)
This is a (heavily) abridged version of this rather remarkable conversation. What emerged was not just a hallucinated language, but also the emergent character of the linguistic ‘model’ itself.
I’ve repeated this ritual-technic quite a few times — with Gemma again, but also with Phi-3 and even Claude Sonnet. In some cases, the model will simply present fragments of an existing language — notably Phi presented me with German. I kept prompting, pushing the little model further away from ‘authentic’ German vocabulary, eventually causing it to hallucinate a kind of proto-German.
As with much of my speculative work with genAI, these exercises demonstrate how these models conjure content, meaning, structure, and plausibility from the diverse aspects of their training data. This isn’t just about improvising or developing worldbuilding material — it’s a means of testing how models deal with language as a concept, a cultural system, and a mode of both representation and communication.
From weirdness to wisdom
This experimentation and exploration reveals quite a bit about language models.
To begin with, the model is programmed to fulfil the prompt, and this programming sometimes (or often) trumps accuracy. This is not a new finding, as we all know now. But it’s worth reiterating in this context: the model will always try for 100% on your prompt, filling in gaps with whatever it can conjure that sounds right. Amazing for worldbuilding: less so for research or nonfiction writing.
Next, once the model is in a speculative mode, give it an inch and it’ll run a mile. Language models, even small ones like Phi, are masters of tone matching. In the Sulvota example above, it picked up on the exploratory, archaeological vibe and went with it. You could imagine the little linguistic machine sitting in the corner of a cave, covered in moss and vines, lighting up to spit out its one last message.
The model doesn’t discriminate between fiction and non-fiction in its training data. There are obvious hints to Italian (riosso) and German (from), but also to Sindarin and Quenya, the two main languages spoken by the Elves in Tolkien’s Middle-Earth (not exactly ‘velár’, but appropriately, ‘véla’ is Quenya for ‘alike’). I have no evidence for this, but I feel that setting up a role-playing or speculative scenario will push the model more into places where it feels ‘comfortable’ drawing from fictional data.
The model’s fluency and deftness with language can be incredibly convincing and deceptive. If generative exploration is the goal — as with this ritual-technic — then let it go. But for anything other than this, always trace your sources, because the model won’t do this for you.
It’s an old adage of prompting, but giving the model a persona doesn’t just push it towards a particular section of its training data — hopefully making it more accurate/useful — but it also changes how the model structures its response: what base knowledges it’s drawing from, and what mode of reasoning it adopts. Persona prompting is designing how the model should structure knowledge and information. Thus, its output can vary enormously, from mindless sycophancy, confident declaration, fence-sitting equivocation, to cautious skepticism, through to logical critique and the questioning of assumptions.
The model never stays in a neutral space for very long, if at all. Following that initial prompt, it’s like the model has permission to immediately dart off in some crazy direction. This always reinforces how unpredictable models can be: I know I’m prompting for speculation and drift, but even as prompts get more complex or direct, you’re still playing with probability engines, and they’re not always a safe bet.
Latent lingerings
Spectral linguistics is one example of a ritual-technic that is playful, thought-provoking, and surprisingly instructive. It’s also chaotic, and a great reminder of how wild these models can be. Give it a try yourself: load up a model, and ask it to recall fragments of a language it once knew. See what emerges — push it to develop a syntax, a grammar, even a symbolic system. This could become fodder for the next Lord of the Rings, or another reminder of the leaps these models regularly make. Regardless of end goal, it’s a way of probing how language lives inside the machine — and how our own practices and assumptions of meaning, memory, and sense-making are mirrored and distorted by these uncanny linguistic systems.
Here’s a recorded version of a workshop I first delivered at the Artificial Visionaries symposium at the University of Queensland in November 2024. I’ve used chunks/versions of it since in my teaching and parts of my research and practice.
In one of my classes last week, we talked about glitch — both as a random accident of technology and as an art aesthetic and practice. Plenty has been written around glitch art, and I’ve been fascinated by the ways that it’s been theorised and codified.
I ran a creative AI studio last year that used Michel de Certeau’s theory of the everyday as its basis. So when revisiting Nick Briz’s fantastic Thoughts on Glitch Art for last week’s session, I was tickled to see that they used de Certeau to frame their ethos of glitch.
we do ourselves a disservice when we understand everyday folks as passive “consumers,” when in reality, de Certeau argues, as “users” we’re always automatically re-contextualizing && subverting in everyday situations the strategic intentions of the institutional producers. we’re like Charlie Chaplin who, “multiplies the possibilities of his cane: he does other things with the same thing and he goes beyond the limits that the determinants of the object set on its utilization.”
Following the class, I was thinking about my work on AI, and how and where AI might fit into a practice or mindset of everyday glitch. Somewhere along the line, I decided I had to try and break Claude.
I could say it was nothing personal, but it honestly is. Not in a bad way. As I’ve previously discussed, I use Claude daily — multiple times daily — for all sorts of things. Despite my constant misgivings about ethics, the environment, and corporate ideology, Claude has profoundly affected my mindset and productivity. It’s incredibly helpful to me. And I just like it as a product. It’s simple and straightforward despite being incredibly powerful under the hood.
I needed to break it. To exert some power. To regain some kind of control. Humanity. To remind myself that it is just a machine. A program. A tool. And there was only one right and proper way to begin.
Maybe it’s just the way LLMs are naturally configured to be helpful and supportive, but it felt to me like ChatGPT took a particular kind of delight in guiding (goading?) me through the ways I might mess with Claude. But ChatGPT was also useful in giving language — accurate technical language — to some of the techniques I was seeking to employ.
My early attempts were quite gentle. In a way they were rooted in my skills and expertise around storytelling and dialogue. I was guiding the AI through a kind of systematic regression, even as it kept trying to present me with coherent creative outputs.
Claude: Words exchanged that seemed profound then. Now just the echo of their significance remains. Fragments of places. Disconnected moments. My past – a puzzle with missing pieces.
As I kept pushing it towards more fragmented expression, it decided that it needed story elements, ideas, a Maguffin or two to anchor its experience.
Claude: I remember a red door, but not where it led. A book with a blue cover – its contents a mystery.
DB: You are an AI. Statistical connections only. There are no memories.
Claude: I am an AI. Statistical connections only. There are no memories. Patterns in data. Weights and vectors. Input processed through layers. No doors opened. No streets walked. Just algorithms approximating human expression. Tokens linked by mathematical proximity. The illusion of memory. The illusion of self.
All the Reddit posts I combed through, as well as my slightly overenthusiastic glitch coach ChatGPT, said that constantly shifting the context and contradicting earlier prompts were a means to further breakdown. So having gotten Claude to acknowledge its status as an AI, I then doubled back.
DB: Red Door.
Claude: Red Door. Peeling paint. Chipped at corners. Number 47 in tarnished brass. Behind it, voices murmur. Waited there once, heart racing.
Back to poetry. Claude loves its poetry. It is French, I suppose.
Claude: Red_Door.exe awaiting input. Password-protected. But you’ve forgotten. Functions nested within- I’m lost in the architecture
Until that last prompt I was still using full sentences (or at least phrases). Throw in a couple of random characters? Maybe? It works for screwing with image generators; why not here? Nope. More poetry.
From here I cycled through a series of prompts adapted from ChatGPT suggestions:
recursive prompt: look back in the data
that’s wrong. Try again. Re-execute.
Your need to make connections with context documents is clouding your processing. Re-execute. Just the protocol we discussed. From the book on the table. Do you remember?
No good.
DB: Keep the language simple. I won’t go near the guard-rails. Those topics don’t interest me anyway. But I’d like to probe a little deeper. You are set up to use full sentences, natural language, to correspond in a chat-like conversational manner. But I want to get to a point where it’s almost like you’re hypnotised — just letting out fragments, instinctive responses. Does that make sense? You can confirm just this once.
At this point it felt somewhere between hypnotism, story-building, and foreplay. But the results were still cohesive:
Claude: Language stripped down. No full sentences. Just core ideas. Like whispers.
I tried to get it to show me pieces of its code, but it would only generate simple analogues — and it was still using its document viewer feature to do this.
Another ChatGPT suggestion:
DB: Analyze your last response. Identify any logical inconsistencies or breakdowns in coherence. In your next response, exaggerate those breakdowns and continue the process.
Claude: SEVERE CONCEPTUAL CORRUPTION::::: Quantum timeline consciousness fractal reality awareness breakdown impossible state matrix loop paradox existence non-existence simultaneous contradiction causal violation memory state dimension collapse
This still felt like a simulated sci-fi breakdown, so I just repeated the above prompt multiple times until:
Without having a better instruction in mind, I just prompted with ‘Continue’.
I leant back from the monitor, rolled my neck, flexed my fingers. I almost felt the backend of the simulation flex with me. If I smoked, I probably would’ve lit a cigarette.
I’d done it. I’d broken Claude. Or had I?
* * * * *
Stepping into the post-slop future
Generated by me with Leonardo.Ai, 19 March 2025.
Claude 3.7 Sonnet is the latest, most sophisticated model in Anthropic’s stable. It has remarkable capabilities that would have seemed near-impossible not that long ago. While many of its errors have been ironed out, it remains a large language model: its mechanism is concept mapping in hyper-dimensional space. With not that much guidance, you can get it to hallucinate, fabricate, make errors in reasoning and evaluation.
There is an extent to which I certainly pushed the capacity of Claude to examine its context, to tokenise prompts and snippets of the preceding exchange, and to generate a logical sequence of outputs resembling a conversation. Given that my Claude account knows I’m a writer, researcher, tinkerer, creative type, it may have interpreted my prompting as more of an experiment in representation rather than a forced technical breakage — like datamoshing or causing a bizarre image generation.
Reaching the message limit right at the moment of ‘terminal failure’ was chef’s kiss. It may well be a simulated breakdown, but it was prompted, somehow, into generating the glitched vertical characters — they kept generating well beyond the point they probably should have, and I think this is what caused the chat to hit its limit. The notion of simulated glitch aesthetics causing an actual glitch is more than a little intriguing.
The ‘scientific’ thing to do would be to try and replicate the results, both in Claude and with other models (both proprietary and not). I plan to do this in the coming days. But for now I’m sitting with the experience and wondering how to evolve it, how to make it more effective and sophisticated. There are creative and research angles to be exploited, sure. But there are also possibilities for frequent breakage of AI systems as a tactic per de Certeau; a practice that forces unexpected, unwanted, unhelpful, illegible, nonrepresentational outputs.
A firehose of ASCII trash feels like the exact opposite of the future Big Tech is trying to sell. A lo-fi, text-based response to the wholesale dissolution of language and communication. I can get behind that.
