ES

CARLOS SANDOVAL MENDOZA



Post-COVID Series, 98/98

Berlin, Cuernavaca, 2021–2025

Ink on paper, 15x15cm on A4 acid-free paper

Interrupted nested GAN-based projection,
collapsing arbitrary images onto a narrow portrait manifold in a few recursive steps

Full collection — all 98 works

2023–24, Chongqing International Culture Art Center, Chongqing, China
Group exhibition, curated by Pashmin Art Management


 

Post-COVID #61

CS: POST-COVID #61

 

Three things happened at the same time: the pandemic locked us all down, AI began to enter our everyday lives, and in the spring of 2021, I rode 1,600 kilometers by bicycle, alone, through four countries. That journey, in the middle of the lockdown, was a survival decision, not a “project” or a protest. It was a way not only to fill my lungs freely again, but to escape the restrictions and a dystopian urban landscape of faces that no longer let themselves be observed.

The journey was also an inner one. Every night I slept alone in a different forest. My only company was distance, mental noise, and the silence of the trees, imposing shadows barely visible at midnight. If the Monas and Tent Visions series were born out there, outside, the Post-COVID series is the inner record: what happened to me inside, when the interruption was total. But, as always, I don't believe stories, and mine even less. That is reflected in the procedure of this series of almost one hundred drawings.

One of the first AI portrait apps used a now-outdated technology. Judging by the traces it left in its images, it was probably a GAN: trained on classical European portraits, it transformed users' selfies into 18th-century-style portraits, all of white people, almost plagiarized from a European museum. With a selfie, the app returned four finished portraits, with few differences between them, for the happy user to choose from. I, on the other hand, forced it into an anti-pareidolia: searching for patterns where there were none. Its first four output images were neither objects nor faces anymore. I saved one and fed it back as the source image, and with each round it became easier for the app to find its patterns.

But I didn't believe this story either: I stopped the iterative process, chose the most interesting image, and then transduced it into ink, in lavish detail, in work that was at once obsession and introspection. My stroke entered exactly at the point where I aborted the algorithm's fetus-form. An anti-pareidolia that turned into anti-morphosis. Of course, I neither fully transformed nor reinvented myself, yet I am not the same either: linear autobiography breaks halfway, and it is halfway where I feel most firmly planted.

Mother cell 1: found objects arranged and photographed for the app Mother cell 2: found objects arranged and photographed for the app Mother cell 3: found objects arranged and photographed for the app Mother cell 4: found objects arranged and photographed for the app Mother cell 5: found objects arranged and photographed for the app Mother cell 6: found objects arranged and photographed for the app

CS: MOTHER CELLS, MADE TO FEED THE APP

While the technological mediation began with the app, photography entered the process almost involuntarily. With use, I learned to predict, if only partly, how the app (PortraitAI) would interpret my photos. I began photographing textures, composing and manipulating found objects that could serve as mother cells. Since the app influenced the way I photographed, the drawing actually began with the photo. But from the start, the app was just one more tool, like the objects themselves, the camera, the paper, and the ink: together they formed a non-hierarchical field.

Post-COVID #28

CS: POST-COVID #28, BASED ON THE THIRD MOTHER CELL

Defossilization?

After finishing the series, I kept playing with the app while waiting somewhere: I photographed whatever was in front of me to challenge it to find faces, until I discovered its "upgrade". And that upgrade came in 2026. The new version runs through several steps on its own, bringing the image closer to a face. If its detector cannot find one with enough certainty (although I see a "future face" in my photos, my "mother cells"), it no longer lets me choose among four half-formed attempts: it cuts the process and superimposes a generic face over the intermediate state. In this way, whether for ethical, market-oriented, or financial reasons, the company built in something my procedure had anticipated: abruptly cutting the metamorphosis. What it added on its own was replacing the Eurocentric portrait it used to arrive at with an out-of-context image.

The traces of the Post-COVID series (the photos, the interrupted app sequences, and the drawings themselves) became the only testimony of how that version of PortraitAI worked as a whole: some steps still survive, buried under the new version, but not its destination. And perhaps for a reason. In the app's reviews, a user wrote: "Apparently the AI is racist: my African-American friends all turn white. It's problematic, the message seems clear: 'Black is bad, white is good.' Only white is normal and desirable." The company itself acknowledged this on the same App Store page, apologizing because its AI had been trained mostly on portraits of people of European ethnicity. The debate is not limited to this case: AI has been accused not only of racism, but of neocolonialism, conformism, biased opinions, and half-truths. But that is another topic; here we will talk about defossilization.

I do not use the word "fossil" lightly: in paleontology, body fossils are the remains of an organism's body; trace fossils are traces of its activity; and molds and casts preserve its form in another material once the body is gone. In this series, the body, that version of PortraitAI, no longer exists: only a few scattered steps remain, buried under the new version, like fragments of a skeleton. The interrupted sequences I saved are its trace fossils: traces of how it moved, imprinted on my photos as on sediment. And the drawings are molds: they preserve its form in another material, ink on paper, more durable than any digital file.

Photo taken for the app, 2025 PortraitAI, early version: step 1 of the nested sequence PortraitAI, early version: step 2 of the nested sequence PortraitAI, early version: step 3 of the nested sequence Post-COVID #97, one of the last drawings of the series

CS: 2025: PHOTO TAKEN FOR THE APP, THREE NESTED STEPS IN PORTRAITAI (EARLY VERSION), AND POST-COVID #97, ONE OF THE LAST DRAWINGS OF THE SERIES

PortraitAI in 2026: step 1 of the nested sequence PortraitAI in 2026: step 2 of the nested sequence PortraitAI in 2026: step 3 of the nested sequence PortraitAI in 2026: step 4 of the nested sequence The final image produced by PortraitAI in 2026

CS: 2026: THE SAME PHOTO (ABOVE, LEFT) IN TODAY'S PORTRAITAI: FOUR SIMILAR STEPS AND THE FINAL IMAGE

Meanwhile, millions of images from the first AI apps lie scattered across servers, with no author or context. Unlike the Anthropocene, whose plastics and ashes will mark the rocks for millions of years, this "AI-ocene" may leave no legible layer. Its material layer will remain: chips, minerals, electronic waste. What is lost is the inscription. A single image is already distributed across the sectors of a disk, held together only by pointers that indicate where each fragment lies and which one follows. If the pointers are lost, the fragments remain like bones without an anatomy, and unlike bones, no fragment means anything on its own. When the disk degrades, the medium survives as an empty fossil: the object remains, the image does not.

Excavated bison bonebed at Hudson-Meng, Nebraska: hundreds of bones mixed without anatomical order

HUDSON-MENG BISON BONEBED, NEBRASKA, ABOUT 10,000 YEARS OLD: SCATTERED BONES, WITHOUT ANATOMY. PHOTO: SKYBIRDFOREVER, CC BY-SA 3.0

The paradox is that AI needs the strata it destroys. Fed on its own output, it collapses: the rare disappears and everything converges toward the generic. That is why a market has already emerged for books published before 2022, offered to AI companies as guaranteed free of contamination. Anthropic, for example, according to court documents, bought millions of printed books, cut them apart, scanned them in bulk, and destroyed the originals. The company does not deny the destructive scanning; it only claims that none of those books were rare or antiquarian. But according to internal documents cited by Snopes, it bought with an emphasis on "less common" titles, and one of those documents describes the project as an effort "to destructively scan all the books in the world." The paper fossil is shredded, leaving a digital copy, fragile and private.

Apparently trivial, this phenomenon of cultural "defossilization" means that, unless there is a repository of AI versions, it will be impossible to reliably reconstruct what this technology was: a supposed quantum leap of twenty-first-century techno-culture that is inevitably disappearing. Hence these terms, which you can consult in my Lexicon: Archaeology of AI and Paleontology of AI.

References

Gabbott, Sarah and Jan Zalasiewicz. Discarded: How Technofossils Will Be Our Ultimate Legacy. 2025.

Maiberg, Emanuel. "AI Companies Are Buying Tons of Old Books Because They're Free of AI Slop". 404 Media, July 21, 2026.

Parikka, Jussi. A Geology of Media. University of Minnesota Press, 2015.

Shumailov, Ilia et al. "AI models collapse when trained on recursively generated data". Nature 631, 2024.

Zalasiewicz, Jan et al. "The technofossil record of humans". The Anthropocene Review 1, 2014.

"Are AI companies scanning and destroying millions of books, including rare titles?". Snopes, 2026.

"This Dutch bookseller thought a request for 3,000 copies was 'spam or phishing.' Instead, AI companies are scanning and destroying books to train AI". Fortune, July 31, 2026.

Bartz v. Anthropic. Ruling by Judge William Alsup, U.S. District Court for the Northern District of California, June 2025.

 

Post-COVID Series, interrupted GAN self-portrait 31 of 98

 

Post-COVID Series, interrupted GAN self-portrait 36 of 98

 

Post-COVID Series, interrupted GAN self-portrait 38 of 98

 

Post-COVID Series, interrupted GAN self-portrait 45 of 98

 

Post-COVID Series, interrupted GAN self-portrait 46 of 98

 

Post-COVID Series, interrupted GAN self-portrait 48 of 98

 

Post-COVID Series, interrupted GAN self-portrait 49 of 98

 

Post-COVID Series, interrupted GAN self-portrait 57 of 98

 

Post-COVID Series, interrupted GAN self-portrait 59 of 98

 

Post-COVID Series, interrupted GAN self-portrait 60 of 98

 

Post-COVID Series, interrupted GAN self-portrait 61 of 98

 

Post-COVID Series, interrupted GAN self-portrait 62 of 98

 



 

 

Enlarged image