Model Collapse: Love you, bye

Curated by
Eva Vaslamatzi

Archaeological Museum of Eretria
26–31.08.2025

In the exhibition Model Collapse: Love you, bye, Maria Mavropoulou explores parallels between the human brain and artificial intelligence in terms of their processes of deterioration. The gradual loss of memory—such as that seen in cases of dementia or Alzheimer’s—is likened to the phenomenon of model collapse in AI, which occurs when models are trained on data they themselves have generated, leading to errors and degradation.

Through research involving patients in the early stages of dementia, the artist documents objects tied to their personal memories, while also engaging with artificial intelligence systems to examine how the collective conscious is reflected in AI models.

The exhibition is enriched by the inclusion of photographs from family albums, which are closely linked to the formation and preservation of human identity. The works enter into a dialogue with the archaeological artifacts on display at the Eretria Museum—a space that serves as a repository of collective memory par excellence.

Through both material and conceptual means, the artworks invite viewers to reflect on the complexity of human memory, its role in shaping both individual and collective identity, and the role of technology in co-constructing these processes.

This work would not have been possible without the invaluable support and collaboration of the psychogeriatric organization Nestor.

The production is part of the Ministry of Culture’s 2025 program
"All of Greece, One Culture"
www.allofgreeceoneculture.gr

 

Model Collapse has been a research based project resulting in many pages of notes, ideas and comments that resulted from my longstanding interest on how memory is mediated by AI. This podcast has been created by feeding NotebookLM AI with all my research material and then edited down by me.

Model Collapse -Podcast 00:22:40 s
 

Memory Excavation, 2025

Just as the earth is mined for precious metals and minerals (lithium, cobalt, rare earths, etc.) used in the production of advanced technological products and AI infrastructure, so too is human experience—our memories and data— are “mined” (the English term for this process is data mining) and commodified to satisfy the insatiable need of AI systems for human-generated data.

Under the looming threat of Model Collapse—which can occur when artificial intelligence is trained on synthetic data —the value and rarity of our personal data become comparable to that of precious minerals.

Another starting point for this work came from my research into how the brain of people with dementia function, particularly how they perceive time in profoundly different ways than a healty brain does. In their perception, memories lose chronological order and become entangled in a simultaneous state. For example, they may describe a present moment in which they are working, raising children, and living with their parents—even though none of that is actually happening—yet they experience it as their current reality.

Wanting to convey this sensation of memory loss and the concurrent coexistence of multiple states, I created this work.

 
Composite image from Memory Excavation layering family photographs and scanned objects to evoke memory loss and the simultaneous experience of time.
 
 
Installation view of Memory Excavation showing archival family photographs layered inside a wooden box, connecting personal memory with data mining.
Detail of Memory Excavation with overlapping family photographs, reflecting how memories lose chronological order and become entangled in dementia.
Close view of archival photographs intersecting inside Memory Excavation like fragmented personal memories existing at the same time.
Memory Excavation detail treating family photographs and personal data as rare materials mined and preserved by artificial intelligence.
Memory Excavation installation of 13 archival photographs mounted on PVC and enclosed in a custom wooden box, representing entangled memory.

Memory Excavation, 2025. 13 Archival photos printed and mounted on pvc, enclosed in custom wooden box. 120 cm x 80 cm x 65cm

View through layered archival photographs in Memory Excavation, where partial family images coexist without a clear chronology.
Installation view of Memory Excavation, placing a wooden archive of fragmented family photographs in dialogue with collective memory.
 

Love you, bye 2025

An echo of human communication as filtered through predictive models (autocomplete) that both facilitate and shape our everyday interactions. The work explores the notion of echolalia—a term describing the involuntary repetition of words or phrases by individuals with Alzheimer’s and dementia—highlighting its resemblance to the mechanical behavior of predictive algorithms.

Engraved in marble, the work establishes a dialogue with the museum exhibits and the numerous ancient steles onto which significant texts were carved and displayed in public spaces to preserve information “for eternity.” In contrast, the text here emphasizes the ephemeral and fragile nature of communication in the digital age.

While initially appearing familiar and believable, its absurd repetition reveals the true author: a machine trained to predict our next word (predictive text AI).

There are 25 different images, Eaach poster features a text starting from one of the 25 letters of the english alphabet.

Love you, bye installation with an iPhone predictive-text message engraved on a marble stele beside stacks of posters for visitors.

Love you, bye , 2025. Marble column (stele) inscribed with iphone’s autocomplete predictive text, 70 cm x 50 cm x 10 cm . On the right, a stack of 400 posters for the visitors to take with them. Each poster features a text starting from one of the 25 letters of the english alphabet.

Detail of predictive autocomplete text inscribed on white Thassos marble in Love you, bye, contrasting digital ephemerality with ancient preservation.

detail- Inscribed text on Thassos white marble , 70 cm x 50 xm x 10 cm

Stacks of Love you, bye posters visualizing predictive texts for the 25 letters of the English alphabet, available for visitors to take.

Four stacks of 400 posters with visualizations of predictive texts, available for the visitors to take with them. dimensions: 70 cm x 50 cm each. Each poster features a text starting from one of the 25 letters of the english alphabet.

 

Soft Errors 2025

The work Soft Errors is inspired by the discovery of an anonymous family photo archive, which forms a visual narrative of the lives of its "protagonists." The artist focuses on a female figure—likely the creator of the archive—suggesting her dual role as both the recorder/photographer and the curator of family memory.

Starting from her portrait, Mavropoulou intervenes in the image using an artificial intelligence system, inputting her photo. Each image generated becomes the input for the next, forming a continuous chain of self-feeding iterations.

As the process unfolds, the human face is gradually deconstructed, ultimately transforming into a neutral, formless surface.

The work serves as an allegory for the deterioration of memory—both biological and artificial—directly referencing the phenomenon of model collapse.

At the same time, it offers a poetic tribute to an unknown woman as a vessel of familial memory, whose image is now deconstructed and recycled as data by an impersonal intelligence, caught in an entropic loop.

text by Eva Vaslamatzi

Soft Errors installation of ten semi-transparent portrait prints showing an anonymous woman's image transformed through recursive AI generation.


10 prints on semitransparet fabric 150 cm x 150 cm each, metal structure 2m x 4m x 230cm (hight)

Side view of the Soft Errors fabric installation, where overlapping portraits trace the gradual deterioration of human and artificial memory.


10 prints on semitransparet fabric 150 cm x 150 cm each, metal structure 2m x 4m x 230cm (hight)

Soft Errors installation detail showing the unknown woman's face deconstructed across translucent layers by a self-feeding AI process.
Output 1 from Soft Errors, an early AI iteration that still retains the portrait and identity of the woman from the found family archive.

output 1

Output 34 from Soft Errors, where recursive AI processing shifts the archived woman's portrait toward a younger, idealized face.

output 34

Output 68 from Soft Errors, showing the portrait distorted as repeated AI generations erode the original woman's features and identity.

output 68

Output 10 from Soft Errors, an early recursive AI image in which the woman's face begins drifting from the archival photograph.

output 10

Output 45 from Soft Errors, where the AI increasingly beautifies and exaggerates the woman's facial features through repetition.

output45

Output 74 from Soft Errors, showing the human portrait becoming distorted and less recognizable within the model-collapse sequence.

output 74

Output 23 from Soft Errors, where recursive AI generation moves the archived portrait toward a younger, idealized appearance.

output 23

Output 57 from Soft Errors, with exaggerated facial features revealing the AI model's beautifying and pornographic visual biases.

output 57

Output 88 from Soft Errors, where the portrait loses human detail and approaches a neutral, formless image through recursive AI processing.

output 88

Output 100 from Soft Errors, a minimal color field left after the archived woman's portrait collapses through repeated AI generation.

output 100

In the above images selected one can see the common internalized biases of the AI used (Stable Diffusion). The image drifts to younger faces that the initial input, then it gradually beautifies it, exaggerating the facial feature to the point of   reminding pornographic images (output 57). Then the images are being distorted and monsterified, until gradually losing details , turning to a male figure, and at the end losing every form, concluding to minimal color fields reminiscent of Rothko-esque style. 

Installation view of Soft Errors, presenting the unknown woman's recursively transformed portrait as translucent layers of fading family memory.
Photographic view of Soft Errors in the museum, where AI-deconstructed family portraits form a monument to biological and artificial memory loss.
 

Feedback Loops 2025

A video composed of a series of 30 feedback loops depicting the gradual decay of images as they are processed by an image-to-image generating AI system through 100 iterations. Custom-coded software was used.

The work highlights the vulnerability of digital memory and the risk of Model Collapse, emphasizing the deterioration that occurs when artificial intelligence is fed with data it has generated itself.

These videos, which can be likened to a visual version of the children’s game “broken telephone,” reveal AI’s often hidden tendencies to simplify and beautify images—eventually, through multiple feedback stages, converging into minimalist outputs that tend to revert to the original condition of such generative processes: noise.

In the video Feedback Loops, the starting point is a set of personal objects chosen by individuals with dementia during a meeting with the artist. They were asked to bring only one type of item: objects with strong emotional significance for them. A variety of keepsakes—photo albums, clothing and embroidery, electronic devices, and everyday tools, whose functions have partially or completely faded—were meticulously photographed. These images were then fed into an artificial intelligence program that, through a process of feedback, generates each image as the basis for the next, forming a continuous chain of reproduction. The process acts as a technological inquiry into the very concept of memory: what remains when the content of experience has been detached from its form? Just as natural memory fades, the objects are transformed into abstract, often unreadable images drawn from the collective "memory" of artificial intelligence. The work takes shape as a technological monument to decay and forgetting—a condition that questions whether emotional value can endure even as information erodes.

The work reflects on the notion of entropy, while also evoking timeless symbols such as the ouroboros serpent—a motif also present in the museum's collection.

text by Eva Vaslamatzi

Feedback Loops video installation showing 30 recursive image sequences that deteriorate through 100 iterations of image-to-image AI processing.

Video installation. Video duration 12 min.

List of emotionally significant objects brought by people with Alzheimer's for Feedback Loops, including keepsakes tied to personal memory.

A list of the objects Alzheimer’s patients brought to me

 
Grid of 100 Feedback Loops iterations showing an emotionally significant object gradually decay into abstract AI-generated imagery.

Video stills from every step from iteration 1 to 100.

Original photograph of an emotionally significant object chosen by a person with dementia, used as the first Feedback Loops input.

Input image

Final abstract AI image after 100 recursive Feedback Loops iterations, showing personal information eroded through model collapse.

Output image after 100 iterations

Second grid of Feedback Loops video stills, tracing an emotionally significant object through 100 stages of recursive AI transformation.

Video stills from every step from iteration 1 to 100.

 
Original photograph of the emotionally significant object used as the input for the second Feedback Loops sequence.

Input image

Final AI output from the second Feedback Loops sequence after 100 iterations, with the original memory object no longer recognizable.

Output image after 100 iterations

 
Third grid of Feedback Loops stills showing another personal memory object gradually simplified and abstracted by recursive AI generation.

Video stills from every step from iteration 1 to 100.

Original photograph of the emotionally significant object used as the input for the third Feedback Loops sequence.

Input image

Final AI output from the third Feedback Loops sequence after 100 iterations, reducing the personal object toward abstract visual noise.

Output image after 100 iterations

 

Similar outputs exist for all 50 objects brought to me by Alzeimer’s patients

 

Archive of Extraction 2025

The work Archive of Extraction serves as the core of the exhibition’s research process: an open workspace, a miniature artist’s studio, where three different temporalities coexist — human, technological, and geological.

At the center of the work lies the motif of the ouroboros serpent, inspired by an artifact from the museum, functioning as a symbol of cyclical repetition.

On the table, heterogeneous elements coexist: reference books, excerpts from technological and medical terminology, minerals used in the production of devices, time-worn photographs, and work tests.

The space simultaneously operates as an archive, a laboratory, and an exhibition.

The work invites the viewer into an open-ended reading of the process — a reflective observation of creation that offers no definitive answers but instead scatters fragments, awaiting connection.

text by Eva Vaslamatzi

Archive of Extraction table installation combining books, medical and technological terms, minerals, family photographs and artistic tests.

Table top installation, 140 cm x 90 cm

Ceramic Ouroboros sculpture from Archive of Extraction, a tail-eating serpent symbolizing cyclical repetition in memory and technology.

Ceramic sculpture of the Ouroboros Ofis (the tail eating snake)

Dimension: aproximately 30 cm diameter

Detail of Archive of Extraction, an open research workspace where archival photographs, minerals and technological references coexist.

Detail

Installation view of Archive of Extraction as an archive, laboratory and miniature studio joining human, technological and geological time.
Archive of Extraction detail with research fragments arranged for open-ended connections between memory, AI, medicine and material resources.
 
 

Latent Memory, 2025

dimensions: aproximately 40cm x 40cm x 35 cm

This work is a physical construction made from photographs belonging to a found family archive. Inspired by descriptions of the multidimensional latent space that forms the “mind” of an AI model, I assembled this fragile and delicate structure to metaphorically imagine how memories might be organised in a similar way, simultaneously interconnected, forming correlations, while also cutting into one another and cancelling the possibility of seeing the whole image at once, but rather a fragment of it.

Latent Memory sculpture made from found family photographs, interconnected like memories and correlations within an AI model's latent space.
Detail of Latent Memory showing overlapping family photographs that intersect, obscure one another and prevent a complete image from forming.
Installation view of the fragile Latent Memory structure, where fragments of a family archive form a multidimensional network of memories.