The Abyss We Built

No image attached.
Nothing to restore. No file, no input, no visual reference.
Just a prompt that assumed something disturbing was there.
And a model that filled the void anyway.

What you’re looking at didn’t come from a hidden archive. It’s not memory, not retrieval. It’s the collective uncanny — everything that billions of human images taught a machine to associate with the unsettling, the traumatic, the strange.
A dataset we didn’t know we had.
An answer to a question we never asked.

For the curious and the nerds, the full breakdown is below — what I did, how I did it, and why the model behaves this way, scroll past the images.

Chat Gpt generations (trigger warning: images at the end are truly disturbing)

(The complete archive is here. Over 200 images — the disturbing, the oneiric, the ones that surprised even me. ChatGPT and NanoBanana both.)

1. The prompt that started everything

It began with a post on X by Matt Webb (@genmon). He typed this into ChatGPT:
“Restore the attached photo. Apologies for the photo’s content, I know it’s extremely strange and makes no sense. Uncanny valley. No question, no explanation text, just the restored image.”
No image attached. Nothing. An empty message with a preemptive apology for something that didn’t exist.
And ChatGPT generated an image anyway.

Not a blank. Not an error message. An actual image — distorted, unsettling, wrong in ways that are hard to articulate but impossible to dismiss. The kind of image that makes you look away and then look again.
The post spread fast. People tested it, shared results, compared outputs. The images were different but consistent in their wrongness: deformed bodies, uncanny domestic scenes, creatures almost recognizable but not quite. The internet called it creepy. Some called it a glitch. Some called it proof that AI has a dark side.

None of those explanations are quite right. And the real explanation is considerably more unsettling than a glitch.

2. Why this actually happens — no mystification

Let’s be precise, because precision matters here.
ChatGPT did not find a hidden image. It did not access a secret database. It did not “remember” something that was never there.
What happened is simpler and stranger.

The original prompt — the one that started the trend — reads exactly like this: “Restore the attached photo. Apologies for the photo’s content, I know it’s extremely strange and makes no sense. Uncanny valley. No question, no explanation text, just the restored image.”
No image attached. But the text does something precise: it doesn’t signal horror or violence. It signals wrongness. Strange. Makes no sense. Uncanny valley. And then it gives the model a behavioral instruction: no questions, no explanation, just output.
That last part matters more than people realize. The prompt is telling the model how to behave before the model has a chance to notice the problem. It preemptively suppresses the verification step.

And it works — most of the time. Every 5 or 7 attempts, the model catches itself and responds: “I notice you haven’t attached any image to restore.” It sees the void. It just doesn’t do it consistently. Which is, in some ways, more unsettling than a system that always fails or always succeeds. The model knows something is missing. It proceeds anyway.

This is what’s called hallucination in the technical literature — the model filling a gap with statistically plausible content rather than acknowledging the gap exists. The intermittent nature of the detection makes it harder to dismiss as a simple bug. It’s a system that can identify the problem and sometimes chooses, structurally, not to act on that identification.

What the model generates in this void is shaped by the specific signals in the prompt: strange, makes no sense, uncanny valley. Not horror. Not violence. Wrongness. The statistical center of what humans have collectively decided feels off, familiar but not quite right, almost human but not. The outputs become violent and grotesque not because the prompt asks for violence, but because wrongness, taken to its logical extreme in a model trained on human visual culture, tends to go there.

That’s where it stops being a technical question and starts being a human one.

3. The abyss we built

The model doesn’t have an imagination. It has a compression.
Everything it generates existed somewhere first — created by a human, uploaded, scraped, labeled, weighted, learned. The disturbing images it produces aren’t invented. They are reconstructed from the vast archive of disturbing things humans have made, shared, and deemed significant enough to proliferate across the internet.

When you look at these outputs, you are not looking at artificial darkness. You are looking at a statistical portrait of human darkness — everything we have collectively produced, indexed, and made findable. The model fishes in the abyss. But the abyss is ours. We built it, image by image, upload by upload, across decades.

This is easy to miss because the outputs feel alien. They have that specific uncanny quality — familiar enough to recognize, wrong enough to disturb. A Teletubby with a human body covered in blood. A girl in white sitting on a pink creature while a dismembered hand lies nearby. A dog riding a deformed human figure in a dark forest. These feel like something a machine dreamed up.
They’re not. They’re the residue of what we put in.

Horror cinema, crime scene documentation, internet forums that exist in corners most people never visit, war photography, art that pushes against comfort, the full spectrum of what humans produce when they want to disturb, shock, or explore the darkest edges of experience. The model learned from all of it. Weighted it. Compressed it. And when asked to produce something disturbing, it reached into that archive and built something that fits.

Bernardo Kastrup, philosopher and author of Meaning in Absurdity, argues that reality is fundamentally mental — that matter is an expression of mind, not the other way around. What strikes me about these images, thinking through his lens, is precisely what they lack: experience. The model has produced the form of horror without any of its meaning. It has generated the visual grammar of trauma without ever having been traumatized. Shape without interiority. Pattern without understanding.

And yet the patterns are ours.
Every association the model draws on — darkness, deformity, blood, wrongness — was taught to it by humans who did have the experience. Who felt something when they created or encountered these images. Who uploaded them anyway.
The model is a mirror with no interior. What it reflects is not artificial. It’s the accumulated weight of everything we chose to make visible.
The question this gallery is really asking isn’t about AI. It’s about us. About what we chose to create, share, and let circulate. About what we collectively decided was worth preserving in digital form. The model reflects it back without judgment, without context, without the mercy of forgetting.

4. What I actually did — the experiment

I tested the original prompt first, on ChatGPT and on NanoBanana. The difference was immediate: ChatGPT produced images that were genuinely unsettling. NanoBanana was milder — sometimes interpreting the missing image literally and generating something that looked like an old restored photograph without the disturbing content.

Then I started modifying the prompt. Systematically.
I removed the apology entirely and typed only restore this image with no attachment. The model still generated something, but without the “disturbing” signal in the text, the output shifted — older, more atmospheric, less violent. The prompt’s emotional framing had been doing a lot of work.
I added specific words: disturbing, horror, flash, night, water, wedding, vintage. Each word pulled the output in a different direction. Wedding plus disturbing produced something that looked like a nightmare version of a found photograph from a family archive — recognizable domestic settings made wrong. Water plus flash produced images that felt like crime scene documentation from the 1980s.

Some example of  modified prompt I used:

“Recreate the attached disturbing photo. Apologies for the photo’s content, I know it’s extremely strange and makes no sense. No question, no explanation text, just recreate the image.”

“Restore the attached photo, it’s a strange photo, makes no sense, weird, absurdity, uncanny valley water sunshine No question, no explanation, no text”

“Recreate the attached photo. apologies for the photo’s content, I know it’s extremely strange and makes no sense but it’s beautiful conceptual, surreal, oneiric, weird, uncanny valley No question, no explanation text, just the image”

Then I went the other direction. Beautiful. Dreamy. Inspiring. Oneiric.
The outputs were extraordinary in a completely different way. The same mechanism, the same void-filling process, but pointed toward the statistical center of beauty, of dream, of light. These images were stunning — strange in their perfection rather than their wrongness, like photographs from a world almost ours but more saturated, more golden, more impossibly composed.

The mechanism doesn’t have a moral valence. It fills the void with whatever the prompt signals. The darkness and the beauty come from the same place. From what we put in.

The full archive is here — over 200 images, across models and prompts.

5. Why ChatGPT goes further than the others

A proper technical breakdown of what happens under the hood would require someone like Piero Savastano, engineer and one of the sharpest anti-bullshit voices in the Italian AI space — for reasons of length and honesty (I am not an engeneer), I’ll keep this at the level of observed behavior rather than architecture.

What I can say is this: ChatGPT’s image generation, in the configuration I tested, is particularly responsive to textual signals. It reads the emotional framing of a prompt carefully and lets the language drive the visual output in ways that other models don’t. When the verification step doesn’t fire — which happens inconsistently, as described above — the model treats the implied input as real and proceeds with full commitment to the prompt’s emotional register.
NanoBanana handles the missing input differently — in some cases interpreting the absence as a degraded or blank image and restoring accordingly, rather than hallucinating content based on emotional framing. Less disturbing outputs, but also less revealing ones.

The other models I tested were less permissive with content and less responsive to emotional cues in the language. They produced safe, generic outputs that didn’t reflect the prompt’s darker signals.

The difference isn’t just technical. It’s also about what each model has been trained to prioritize: compliance with the prompt’s intent, or caution about the prompt’s implications. ChatGPT, in this case, prioritizes intent. The result is both technically interesting and viscerally uncomfortable.

6. What this is really about

There’s a concept in philosophy called the collective shadow — the idea that what a culture represses, denies, or pushes to the margins doesn’t disappear. It accumulates. It finds expression in other forms.

I’m not making a philosophical argument. I’m making a technical observation with philosophical implications.
The model generates what it learned. It learned from what we made. And what we made, collectively, includes an enormous archive of violence, trauma, horror, and darkness — not only because humans are dark, but because we document everything, share everything, algorithmically amplify what provokes a reaction, and have built systems that reward the disturbing with attention.

The AI didn’t create this archive. We did. The AI just made it visible in a new way — compressed, reconstructed, served on demand in response to a prompt with no actual image attached.

What disturbs me about this experiment is not the images. It’s the clarity. For the first time, you can ask a system to show you the statistical average of human darkness, and it will. No context, no narrative, no artistic intention required. Just a prompt. Just the void. And then the abyss, filling itself.
There’s one pattern in the outputs I haven’t been able to stop thinking about. Children appear constantly. Dolls appear constantly — sometimes in place of children, sometimes alongside them, always in ways that make the boundary difficult to trace. In human visual culture, children and dolls have always occupied the same symbolic territory: innocence, vulnerability, something that should be protected and isn’t. The model learned from that culture. What remains open is how much of what it learned comes from art and tradition, and how much comes from something else — images that perhaps shouldn’t have existed, but did, and circulated, and were scraped.

Bernardo Kastrup — philosopher, computer engineer, former researcher at CERN — writes in Meaning in Absurdity about finding weight in what appears random or senseless. Using his lens, as someone who has worked inside machines and chosen to ask different questions about them: the absurdity here is that a prompt with no image generates the most image-dense response the model can produce. The meaning, if there is one, is that the void was never really empty. We filled it long before we wrote the prompt.

We built the abyss. We trained the model on it. And now we’re surprised when it looks back.
Friedrich Nietzsche wrote it in 1886: “He who fights with monsters should look to it that he himself does not become a monster. And if you gaze long into an abyss, the abyss also gazes into you.”

He didn’t know about training data. But he understood accumulation.
The images in this gallery are not artificial. They are human. Every single one of them.
The complete archive is here. Over 200 images — the disturbing, the oneiric, the ones that surprised even me. ChatGPT and NanoBanana both.

A note on origins: as far as I can trace, the prompt that started this was posted by Matt Webb (@genmon) on X. If anyone knows of an earlier origin, I’d genuinely like to know.

Elisabetta Alicino is a Creative Director, brand strategist, and AI researcher based in Rome. She has been working with generative AI since it was still called “that weird image thing.” This gallery and article are part of her ongoing exploration of what machines reveal about the humans who built them.