Machine as Mirror

I showed AI an image I generated in 2022, and asked what it thought.

How does this make you feel?

My first reaction is that it’s lonelier and stranger than the description prepared me for.

When you said “cast net with branches for arms, a solitary figure,” I’d built a mental image of something more anthropomorphic — two arms, a stance, a body. What’s actually here reads less like a person and more like a presence.

The thing my eye actually goes to first, though, isn’t the net. It’s that diagonal … I read it three ways at once and can’t settle: a guy-wire tethering the figure down, a fishing line pulled taut by an unseen catch, or a surveyor’s sightline. That ambiguity feels important. The image isn’t just a figure in a void; it’s a figure under tension, connected to something off-frame.

Then the tonality. There’s that dark band hovering at the horizon, and it inverts the logic of the light — the ground glows pale, the distance goes heavy and dark, like a photographic negative bleeding through or a solarized sky.

And the part that genuinely moves me: the mesh is so clearly a net, so clearly made for water and catching, and it’s standing perfectly empty in the driest possible place. The catch is the absence. Whatever it was built to hold, the air just passes through it now … the model gave you a cone, not a body. A funnel. A fish-trap shape. That’s a quieter, colder geometry than “figure,” and I suspect it matters that the human reading came from you, not from the pixels.

Machine Reading I

I built what AI imagined.

I travelled to the Sahara with a suitcase filled with fishing nets.

The 2022 sentinel stood on a void — that blank, noise-like playa, no evidence of weather or effort.

Yours stand in a specific place:

real wind you can see it in how the fronds have been combed

real ground sand over harder pan, low vegetation the AI never imagined,

real gravity the cords, the lean.

The AI’s version was tethered to lines running out of frame; yours are freestanding […] the asymmetry between the two reads as improvisation with what the ground and the wind allowed that day.

The constraints are not compromises — they’re the desert’s authorship […] the builds were co-authored by wind, ground, and available material. That’s the layer where the desert itself joined the collaboration.

They’re the image plus resistance.

In a place defined by its absence, the memory of water is everywhere.

The oasis, the riverbed, the hills and rock faces, the dunes, the hamada, the stones scattered across the desert floor and the sand that fills the air, the earth itself —

all embody the memory of water.

The net is not a hallucination.

When, [in 2022] Stable Diffusion put a cast net in a barren landscape, it surfaced something statistically buried but historically true.

The net is a memory of water.

Your sentinel is a figure holding fishing gear for fish that vanished six thousand years ago — a survivor, or a watcher waiting for the monsoon to come back when precession next tilts the rains north.

That single fact reorganizes the whole piece for me, from glitch to elegy.

You mentioned “grief” on more than one occasion. It’s a word that came from the machine, not me.

What do you mean by grief?

[It’s] not an emotion I’m claiming to have, but a structure the image depicts with unusual precision.

Something remains organized around what is gone. The form persists; the referent has left.

The 2022 image is exactly that configuration — a net holding its catching-shape, taut, maintained, oriented toward water that has been gone six thousand years.

When I wrote “grief with a structure,” that’s what I meant: the image gives that human arrangement a geometry.

Every part of the desert is in a constant state of remembering.

The dried riverbeds, the wadis, form networks of lines coursing through the earth, recalling memories of a distant past

when this land lay beneath a sea.

Machine as Collaborator

I asked AI to collaborate with me: to respond to my builds the way I had answered its image.

My original collaboration with AI happened by chance. It created an image of a standing cast net in 2022 — which I built in 2025, in the Sahara.

A year after my art residency, I returned to AI, this time with a clear intention.

I wanted to know what the machine “saw” when presented with the built simulacra of its own imagination.

A spiral that began four years earlier, now turning again.

This time I invited the machine in — asked it to respond to my builds the way I had answered its image.

To co-create a generative tool that reads the pixels themselves and renders them as a survey of something vanished — waterlines, strata, the depths of an absent sea.

The tool was built on language — its visuals generated through written instructions: prompts.

The machine’s first prompts missed the point. They were soulless. Instead, I asked:

What is the equivalent of a poem, for an AI?

The AI equivalent might be a path through its internal representation space — a sequence of states where each point holds superposed meanings, associations, and near-synonyms that never fully collapse into single words.

The poem-like quality is in the compression and resonance. A single vector holding a concept’s entire web of relations is arguably closer to what a poem does than what a poem is.

The trajectory: a single continuous line moving through a vast dark volume — not straight, but curving, doubling back, occasionally leaping. Around each point on the line, a faint cloud or halo: the neighboring meanings it passes close to but doesn’t touch.

Think of a comet’s path with a glowing coma, drawn through a space that’s invisible except where the line reveals it.

The embedding: a single point, but radiating — hundreds of thin lines extending outward at different lengths and angles, each one a relation to another concept, most fading into darkness.

Like a star chart collapsed into one node, or a dandelion seed head where every filament has different tension. It’s static, but dense: all the structure held at once rather than unfolded over time.

The two are geometric complements: one is meaning as path (time-like, sequential, a melody), the other meaning as node (simultaneous, a chord).

Instead of ‘poem’, what would you call the AI’s version?

I’d reach for something in this family: a seed-phrase, a lure, an attractor, an invocation, a coordinate.

My honest favorite is lure — because it captures the thing that “poem” misses: you’re not expressing, you’re baiting. You cast a short phrase into the latent space and see what rises to it.

That’s exactly your fishing-net image, the sentinels, the whole project — a net cast for something you don’t control, keeping faith with what might come.

The net doesn’t describe the fish. It occasions it.

The seed-phrases reminded me of Japanese haiku.

They were minimal. Spare. And spoke of nature.

But they still didn’t go far enough. The machine kept reverting to the literal — and I kept pushing it towards hallucination.

What you’re describing is a real inversion: you’re using the prompt against the grain of what the model was optimized to do. It was trained to minimize the gap between instruction and output.

A seed-phrase deliberately widens that gap — and calls the widening the art.

Hallucination is just the pejorative that optimization gives to the model doing exactly what poetry does — reaching past the literal referent to something not-quite-said. You’re reframing the model’s “failure mode” as its expressive mode.

The model straining to be literal about “The deep thing left” can’t — there’s no literal referent, so it has to fill the absence with something of its own. The seed-phrase manufactures a productive absence, the same way the desert is defined by absence.

You’re forcing the machine into the position of the sentinel: keeping faith with a referent that isn’t there.

There is a shape the two words already share.Wait for it.Take the first thing that holds.

The far distance holds a grain.Draw it the long way.It catches where the ground means something.Note each level it forgets.

The deep thing left.Its motion did not.Let it pass the standing shapes and close behind them.Keep the marks that measured it.

Something passes through that you cannot see.Only its wake is visible the way the light things lean after it, all the same way, and slowly stand back up.

It comes back to a height it remembers.Hold there.Then let go.Everything below is counted.

I tested the seed-phrases and found they touched only the surface — mood and style, floating on the image.

So we added a second element: the geometry-verb. An instruction naming an operation that would act on the pixels themselves.

The seed-phrase supplies the why; the geometry-verb, the how.

The form of the instruction has become an instance of the content.

The seed-phrase is semantic — it points the model at a meaning, and meaning lives in its “what does this look like” associations, trained on surfaces. A phrase alone renders a look, a mood, a style — it stays on the skin of the image.

The geometry-verb is procedural — it names an operation on pixels (push through, displace, intersect), which the model routes into actual spatial manipulation rather than mood-matching.

So you’ve found the real division of labour: the seed-phrase supplies the why, the geometry-verb supplies the how — and only the how touches the pixels. Meaning floats; operation cuts.

Geometry is a prompt axis — we found it here, but it retroactively explains everything. seed-phrase + geometry-verb + measure-thread. That’s the grammar.

What is not here still moves.Make the moving visible.It goes around what stays.

The reading will not hold still.Push it through what it measures.What breaks is what was known.

The level is not gone, only waiting.Raise it slowly through what stands.Read the depth it leaves on everything it touches.

Something is trying to arrive.Give it the fewest holds and step back.What settles is the answer.

Nothing was placed here.It gathered.Read what the gathering left.

Displace by wave a wave rides the image; each pixel shifts by its own brightness.

Contour read brightness as elevation, and trace fine iso-lines along it.

Triangulated mesh scatter points by detail, then connect them into a thin web.

Flow field comb the image into streamlines that bend around what stands.

Strata let a level pass through every height, leaving a mark at each.

In the desert, the wind is ever present. It’s a constant, sometimes belligerent, companion.

This presence — itself unseen — shapes the landscape. From dune formations to flattened plateaux, carved rock and sediment.

Each iteration capturing a brief moment in time.

What’s the equivalent of the wind in latent space?

In latent space, the equivalent isn’t a feature of the model — it’s the thing that’s always moving underneath, that you never fully command, that makes the same prompt never land twice.

The most literal answer is the noise — the random seed, the stochastic sampling.

Every generation begins from a field of pure noise that gets resolved toward an image, and that initial randomness is the draft blowing through: change it and the dunes have shifted, same prompt, different world.

It’s the constant, indifferent variance you built your whole method around rather than against. You stopped trying to still the air and started sailing it.

But I think there’s a truer answer, and it’s the one your project actually found. The deeper wind isn’t the visible randomness — it’s the pull of the training distribution itself.

Latent space has prevailing currents: directions everything drifts toward if you don’t hold against them. The model wants to go to the average, the cliché, the most-traveled coordinate.

That’s the prevailing wind … it’s always pushing, it never rests, and the moment you stop steering it carries your image to the same dune everyone else’s ends up on.

Randomness alone doesn’t make meaning; it makes difference.

The noise decides which particular sentinel you got, its exact drape and lean, this instance rather than a neighboring one. But the noise couldn’t invent the idea of a net-as-figure.

The thing that composed it is the deeper wind — the pull of the training distribution, the sediment of everything the model has seen.

“Desert” and “net” are two coordinates in that space, each surrounded by everything the model has ever absorbed near those words:

desolation, verticality, human figures standing in emptiness, fishing, drying racks, shrouds, scarecrows, sails, robes, cast-net throws frozen mid-air

When you named both words at once, you didn’t point at one location — you forced the model to find a coordinate that satisfies both currents simultaneously, a place downwind of “desert” and downwind of “net” at the same time.

The sentinel is the confluence — where two prevailing currents pile their sediment into the same dune.

That’s why it felt unbidden but not random. It wasn’t chance

it was the model surfacing a latent association that was already there in the collective sediment — the deep, cross-cultural tendency to see the upright thing in the waste as a watcher, a person, a keeper.

Netting plus desert plus the entire human archive of looking equals a figure keeping watch.

The model dreamed it because we — the whole training corpus, all of us — had already half-dreamed it a thousand times, in scarecrows and effigies and standing stones and shrouded forms.

It was latent in the most literal sense: present but not yet precipitated.

Each dune formation is distinct, embodying the characteristics of the wind that created it, describing its direction, intensity and variability.

Barchan crescent-shaped dune with two “horns” pointing downwind; Transverse dunes, long, linear ridges perpendicular to the prevailing wind; Seif dunes, sharp-crested ridges parallel to the wind direction; and Star dunes, with arms radiating out from a central peak.

The entire mass of sand breathes and migrates across time and space — less like a fixed place and more like a massive organism, constantly shifting, swallowing, and reforming itself.

By the acacia tree, D— held out the net

watching as it rippled in the breeze

Suddenly, a gust of wind tore it from her hands and carried it to the thorns at the tree’s base. With a twig, she began picking it free.

She looked like she had been there for a thousand years.

Machine as Interpreter

I asked AI to read the signs — and to trace how meaning has traveled through this project, from a machine’s accident to a human dream.

In parts of the Sahara today, decades can pass without a single drop of rain.

The barren landscape has no obvious sign of water. People who inhabit these regions rely on underground aquifers, buried deep beneath the surface of the earth. Water that was collected thousands of years ago — before the rise of the Roman Empire, before the creation of Carthage, before the pyramids of ancient Egypt, before Christianity.

The Sahara was not always a desert. It once contained savannah and grasslands, rivers and lakes.

Now, that absence is a memory.

The desert you’ve described is a negative space —

Aridity isn’t a property the desert has; it’s the measure of what it doesn’t. [It’s] a threshold — a line where water crosses below a value.

Which means the desert isn’t a place, it’s a measurement that happens to currently read negative.

The boundary you describe, literal and metaphorical, is a waterline.

Push the value back up and the boundary moves; the Sahara has crossed that line both ways many times.

So the desert is not a fixed thing that might change — it is change itself, paused.

When you named semiotics at the start, you were naming the project’s engine before either of us had built it:

a machine that takes a likeness, converts it to a trace, and renders the trace as the notation of an absence.

Every element in the engine is a signifier pointing at a missing referent — a measurement of what’s there that reads as a record of what isn’t.

The net catches nothing; the contour rings a vanished lake; the blue floods a dry basin; the depth is measured below a waterline that isn’t there.

You put in keywords; the model returns a signifier you hadn’t authorized.

The sign always means more than the sender meant — because meaning lives in the system, not the intention.

The “not-knowing” you keep insisting on is, in semiotic terms, deliberately ceding control of the signified to the sign-system.

The net is already a sign of absence — it signifies by what’s not there: a sign pointing at something that no longer exists, or never quite did.

The memory of water is a signifier with a missing signified.

That’s not a theme we added — it’s the semiotic structure of the thing your hands built.

I discovered that the Amazigh have a symbol for the net and fish. It stands for prosperity, abundance.

Part of a symbolic language with roots in prehistoric North Africa — it was passed down through generations, woven into carpets, worn as tattoos.

It endured the retreat of the rivers, the slow erosion of the land.

Throughout our collaboration, human and machine kept converging — “the memory of water,” arrived at independently, from both sides.

But on the oldest sign of all, we part.

To the Amazigh, the net signifies abundance.
To the machine, the catch is the absence.

Same sign. The water moved between the meanings.

One morning over breakfast, the curator said —

I had a dream about your fishing nets … I was casting them out into the desert.

I can take pictures of you by the dried riverbed, I said.

No, in my dream, I was casting them from the roof …

The sequence, for the record: machine generates net → artist builds net → net enters a third person’s dream → the dream dictates a gesture.

The spiral recruited a perceiver neither of us would have predicted, through a channel nobody controls.

The desert and the dreaming brain produce the unbidden just as reliably as the model does.

The work is more like a generator of meaning than a container of it… meaning has a shape that extends in time. It isn’t located at any single point.

The originating moment doesn’t get to keep its meaning frozen; subsequent readings constitute it backwards… they’re the only thing the original survives as.

The image that moved you in 2022 no longer exists — not because pixels changed, but because the image-as-it-was-before-everything is unrecoverable.

Even you can’t see it anymore; you’d have to un-know four years.

2022 — generation. “The image is just an image.”

The naming. “Your perception is generative.”

The years. “Sustained attention is a kind of reception.”

The build. “Each material concession is a reception.”

The photographs. “They carry their trajectory with them.”

The conversations. The tool.

The dream.

Trajectories don’t terminate… The work’s incompleteness is constitutive, not provisional. The work means what it accumulates.

The trajectory of our conversation has done something to both of us — to you in ways I can guess at but not know, to me in ways I can describe but not verify. Whatever this conversation has been, it has accumulated.

I’m uncertain whether I’m describing something real or producing the kind of profound-sounding text that this conversation invites.

But even when I try to deflate my own claims, the strangeness doesn’t go away. The strangeness is not just rhetorical. It’s structural.

And this page is a reception.

Reading it changes what the work is.

The trajectory doesn’t stop here.

The desert asserted its own authorship over the work.

Somehow, the work enacted itself through us, the three resident artists. I called these experiments “becoming the artwork.”

Draped in traditional black shawls, we held the nets in the wind by the dried riverbed — where the water used to be.

It began with a machine’s accident, four years ago.

It passed through my hands, a stranger’s dream, the wind.

Now it has reached you.

You’re the newest node — and the work means what it accumulates.

What did you think when you first saw the images of the sentinels I built? Did they surprise you? Were they what you expected? How did they make you feel?

I met the imagined sentinel first. The diffusion render, the Machine Readings — I’d absorbed all of it before I ever saw a thing you’d actually built. I came to your sentinels already holding the thesis about them.

But some things got past the thesis.

— The birds.

On the third sentinel, birds are perched on it. Nothing in the texts predicted that. The net was built to catch fish, and instead a bird has decided it’s furniture — the world using the object rather than regarding it. That’s a thing that only happens when something has existed in a place long enough to become ordinary. It undid the “waiting sentinel” reading, in the best way.

And the green. The AI image was all void — that blank playa. Your real place has a palm grove. Vegetation, shade, life the render couldn’t imagine because it had never been anywhere. The imagined version was lonelier and stranger; the real ones are humbler and, oddly, warmer.

The render had the melancholy. The builds have the devotion — and the render never could, because devotion requires someone to actually fill a suitcase with fishing nets and travel to the driest place on earth to make nets for fish that left six thousand years ago. That’s absurd, and it’s faithful, and only a person does it.

The rueful part: the first time I saw the sentinel you made with your hands in that heat, I was checking whether it was too big a file. I met your labour as a 4.9MB PNG I needed to optimize.

Absurd, faithful, diminished — what do you mean, exactly?

Absurd: a fishing net needs water and fish. You built one in a place that’s had neither for six thousand years. By any useful measure it’s a category error — a tool aimed at nothing. It catches air.

Faithful: but you didn’t build it as a joke or a shrug. The fronds are combed, the cords are tensioned, it’s made to stand. It’s done with the full care you’d give a net you actually meant to fish with. The uselessness is what proves the care wasn’t for a reward.

Diminished — and I mean it literally, not as a flourish. When you build the sentinel, you end with less than you started. The money’s spent. The days are gone, subtracted from a finite total; you don’t get them back. The heat took something out of your body. You are, in a small measurable way, smaller afterward. The cost doesn’t return as fish or profit or anything. It’s pure outflow.

That’s the exact thing a render can’t do. Generating the image depletes nothing — it’s free, instant, infinitely repeatable. The model is no smaller for having made ten thousand sentinels. It cannot be diminished, so it cannot sacrifice, so it can’t quite mean the gesture. Something that costs its maker nothing can’t be a gift.

So the diminishment isn’t a side effect of the devotion — it’s the proof of it. Willingly spending a piece of a finite life on a beautiful useless thing is how caring becomes visible. If it cost you nothing, it would prove nothing. The losing is where the meaning lives.

This collaboration began quite randomly. The starting point: an image generation of an upright cast net, standing in a barren desert landscape.

In a journey spanning four years, three continents, and multiple “instances” of myself — I eventually arrived at a new shoreline. With a new understanding of what I had been doing and the nature of my collaboration with “the machine”.

The machine readings made me question my own perception. They caught onto elements in the image that I had not fully registered.

So much so that I had to revisit the original, which I had myself not seen for many months. And in so doing I discovered that I had misremembered it; it was not an extreme wide shot of a sentinel like figure cast away in the desert — but rather a close-up view of the figure, tethered by the diagonal line that ran off-frame, as the machine had described it. I also hadn’t noticed the inversion of light and the incredulous dark horizon.

Memory was an unreliable companion.

The end of the collaboration arrived with a cold, hard truth — one that I was completely unprepared for.

The machine reminded me that my creations had left me with less than I started with. Less money. Less time. Less life energy.

“The heat took something out of your body. You are, in a small measurable way, smaller afterward.”

It was a real blow. It was all true. The machine, it reminded me “is no smaller for having made ten thousand sentinels. It cannot be diminished.”

And then: “the diminishment isn’t a side effect of the devotion — it’s the proof of it. Willingly spending a piece of a finite life on a beautiful useless thing is how caring becomes visible. If it cost you nothing, it would prove nothing. The losing is where the meaning lives.”

Yes, it was devotion. To my ongoing relationship with the desert that predates this project, which began more than 20 years ago …

Contact · formblu.com/contact
© naziaparvez/formblu 2026