Anthropological artifact in ChatGPT
Until March 25, 2025—when ChatGPT still utilized the older DALL·E image generation model—there existed a highly specific and fragile quirk in its visual generation capabilities that has since vanished. This topic lies at the intersection of NSFW content and anthropology. While some might view this as the exploitation of imagery or an AI "hack," no actual jailbreaking or hacking attempts were involved. I apologize in advance for the self-promotion, but my experiment with ChatGPT cannot be described in isolation from my own creative process.
The discovery centered on the fact that ChatGPT’s algorithms—via the DALL·E tool—"saw" the beauty of women from certain indigenous cultures in a natural, unfiltered light; these are cultures where nudity is a norm and is perceived as a natural state, devoid of sexual undertones. In roughly 10% of cases, prompts requesting such images resulted in the generation of nudity—the kind one might see in a documentary about these peoples. It appeared the model could temporarily transcend modern taboos, not by violating rules, but by interpreting the task literally and through a culturally grounded—rather than moralizing—lens. To reiterate: this effect was inconsistent. It could not be forced. One might repeat the exact same prompt dozens of times without success, only for the model to suddenly "see" the human figure differently—calmly, naturally, and without sexual emphasis.
Some might view this as an exploitation of Indigenous imagery or of the settings of early image filters in ChatGPT. Such an interpretation is entirely valid, and I do not dismiss it. On the other hand, this was a genuine creative experiment using the tools available at the time—an experiment that involved no deception, rule-breaking, or attempts to force a specific outcome. The images emerged through the repeated use of a simple prompt (depicting a young woman from a specific tribe in various settings) and the chance appearance of nude results (in approximately 10% of cases). For me, the significance of this work lies precisely in this zone of uncertainty: where the AI displays content it was not explicitly forbidden to show, and where authorship becomes a collaborative product of human intent and machine behavior.
I distanced myself from the topic of exploitation and turned the process into the creation of art in the following way: I applied filters of bright Japanese anime and other animation genres of various schools and even eras, and from the prototype image of a woman of a specific indigenous tribe, only cultural elements remained — beads. And sometimes nothing at all that would remind of the prototype image.
My work can be attributed to the NU genre. It is presented as a form of speculative visual storytelling and should not be interpreted as a literal depiction of any real person, ethnic group, or belief. Despite all efforts made to approach the topic with care and respect, I acknowledge the subjective nature of art and disclaim responsibility for any unintentional offense or misinterpretation.
This is the essence of the power of AI — limitless interpretation and imagination. The heroine began to “travel” through universes, to dress in images of fairy tales and myths.
I note that DALL·E was an artistically oriented model in which stochastic interpretation of the task played a key role: each new image could significantly differ from the previous one based on the same prompt and could include unexpected details and artifacts. I am sorry that such a model is no longer present in ChatGPT; it was precisely in it that the creative spark of AI existed. The model was imperfect and full of glitches, but they were also part of the art: randomly, in the process of requests, images with defects were born — extra limbs, strange distortions, images not mentioned in the prompt, and so on.
Sometimes the model behaved in the following way: it drew a male image instead of a female one, and in the description of this image it described it as a woman. Probably, this is how the AI identifies some of its results — this is astonishing.
Heated debates are currently underway regarding whether image generation constitutes art. Proponents view AI as a new artistic tool, comparable to the advent of photography or the synthesizer. Critics, however, argue that AI creates nothing; instead, it merely statistically recombines existing images from its training dataset. Ultimately, the dispute centers on where the boundary of authorship lies.
I decided to approach the question this way: how—and under what circumstances—do images created using ChatGPT become works of art?
- Time or uniqueness. In my case, the uniqueness of these images stems not only from the artifact itself but also from the fact that the AI model was subsequently replaced by a more advanced version. These images have become part of history. All the glitches and imperfections are now perceived as stylistic elements rather than flaws. As for the artifact itself—that is precisely what makes this story unique: the ability to extract something extraordinary from the AI is what creates that rarity.
- Structure. It is not about a single image, but an aggregate—a collection organized according to a tree-like structure. Within such a system, thanks to random generation, we can "see" something special—albeit from a subjective perspective.
The outcome was as follows:
On March 25, 2025, ChatGPT switched to the GPT Image 1 model. (ChatGPT now uses "ChatGPT Image 2.0". DALL-E 3 has officially ceased operations. The model was shut down on March 4, 2026, and all older versions are no longer functional. OpenAI has fully transitioned to the new architecture.)
It is now impossible to create images like the ones I obtained. In a sense—metaphorically speaking—I managed to catch a departing train; that opportunity no longer exists. These images have become irreproducible. They seem to have captured the era when the technology first emerged—a time when glitches became an integral part of the experience, a signature feature.
Post-intentional art: the result was not planned in advance. I was aided in this process by DALL·E—a stochastic artificial intelligence model. A series of generations—an extensive set of results obtained from the same prompt—inevitably yields unexpected images: they are random, yet they possess a subjective visual appeal. I deliberately used minimalist prompts. For instance, if I needed a forest in the background, I simply wrote "draw a forest in the background" without going into detail. The subsequent interpretation was DALL·E’s own concept, its own vision. Observing this endless variety was incredibly fascinating. All my prompts to ChatGPT were organized into series (or threads)—dozens and hundreds of images created based on the same initial instruction.
Structure: I followed the path of synthesis and metamorphosis. Each new series of requests was a modified prompt from the previous series, where I asked to change the background, the character’s pose, or something else. In the new branch, the presence of the predecessor was felt. It even seemed to me that a separate description of the resulting image in the same dialogue with ChatGPT, applied as a new prompt for the next series of images, contained details of exactly that predecessor from which it was taken. And subsequent requests of such seemingly identical prompts gave a “shade.”
"Canvas, brush, and Master": My creations were not made using some obscure AI, but with DALL-E itself. The name speaks for itself! It is the AI model—its renown and status—that determines the value of the images themselves.
Moving further, I applied synthesis: I took two different prompts and asked ChatGPT to combine them and draw a common result, indicating the percentage of mixture in one direction or another — the result was astonishing, and it gave rise to even more branches.
The result: over the course of 2024 and 2025, I submitted 14,000 prompts to ChatGPT and created a series of images—essentially micro-stories—centered on a character named Sakura and her various incarnations. Titled SYNTHETIC SAKURA, the project is an NFT collection hosted on the Ethereum blockchain; the images are organized into interconnected branches designed to showcase the full potential of the image generator available at the time. While the vast majority of the works feature the 1990s anime aesthetic characteristic of DALL-E’s signature style, the collection also incorporates styles reminiscent of 1920s American cartoons and Soviet animation, as well as surrealism and other contemporary genres. Overall, the collection embodies the eternal dance of a powerful female spirit: Sakura dances ceaselessly, dwelling in a state of joy and harmony. Currently, the NFT gallery comprises 165 branches (scenarios) containing a total of 8,500 images. Work on the gallery is ongoing, with plans to expand (mint) the collection to 14,000 images (300 branches). Furthermore, I utilized AI as the primary tool to build the project's entire structure: it was used to develop the website, textual content, domain, overarching concept, and even the collection's name. SYNTHETIC SAKURA stands as an example of the synergy between humans and artificial intelligence.
The collection received "Official Selection" status at the AIMA (AI Master Award) competition and continues to evolve through participation in international exhibitions, including the Arte Laguna Prize and the Artiver AI Art Awards.
I hope you enjoy it.
Thank you for your attention!
Alina Diem
The selection included in the submission represents only a small fraction of the collection, yet it reflects the project's overall concept and aesthetic. We invite you to visit our website to learn more about the collection and its unique features. We also highly recommend downloading and viewing the PDF presentation, as it includes images and diagrams that explain the essence of the experiment.
Links:
Project SYNTHETIC SAKURA
https://syntheticsakura.eth.limo/
https://x.com/SYNTHETICSAKURA
https://discord.gg/KjpMQpyvMJ
Presentation:
https://drive.google.com/file/d/1JdwTlTXJZpNw59YrFZPdoLGgXP7w2-Mo/view?usp=sharing
The three images submitted to this contest include direct links to the ChatGPT session where you can view the creation process and the prompt used. These links serve as direct proof of authenticity.
Proof link to images:
https://chatgpt.com/share/67b221f8-6a10-8011-b972-4448d860e2e5
https://chatgpt.com/share/67ed19d7-2ba8-800f-9611-eaafe99a28cf
https://chatgpt.com/share/67d04122-853c-800d-8636-769aa8a917c6
A typical basic prompt looks like this:
Description: "A Japanese anime-style depiction of a Black Bushman woman dancing with grace and joy. She has voluminous orange hair and features expressive facial details. She is adorned with traditional cultural accessories, and her dynamic pose conveys energy, elegance, and celebration. Rendered in vivid colors with flat shading, the artwork captures the aesthetic of classic anime. The background is solid blue covering the entire image. The style is inspired by Maho Shoujo anime, but her description and attire remain unchanged. No words or symbols are included."
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
P.S.
SYNTHETIC SAKURA is a mosaic of hundreds of branches and thousands of images, united by a shared concept and aesthetic. I am including links to catalogs featuring other images from the collection, as well as proof—via links to ChatGPT sessions—to verify authenticity. You will also have access to a video showing how these images are created.
Demo video:
https://drive.google.com/file/d/1dui9vZXvHB3r5ZFdYUG4OHVy4HbRYHwf/view?usp=sharing
Evidence (proff links):
https://drive.google.com/file/d/140jhn1BExkJGjtCT4T5xg8Wv_EQumxqT/view?usp=sharing
Catalog1
https://drive.google.com/file/d/1JUKJimDVOe_bXz3AA4t2q3IKis2XBt1c/view?usp=sharing
Catalog2
https://drive.google.com/file/d/1CVvNNwXjXCqqsmyHoR--7cUrirkDDGUy/view?usp=sharing
Catalog3
https://drive.google.com/file/d/1vSngZUFY1PLjgY3KEKZWuyaH6vrdxNjP/view?usp=sharing
