What happens when an AI mistake is worth keeping?

A fish grows a pair of strangely human-like legs. Another creature emerges as a hybrid that seems to belong to no existing species. For most users of generative AI, these would be failed outputs, something to discard before trying again.

For Ayaka Yamazaki, a fourth-year student in the Department of Aesthetics and Art History at Jissen Women’s University, they became the beginning of her graduation research.

Ayaka’s project, “Research on the Creation, Characterization, and 3D Development of ‘Unpredictable Forms’ by AI,” explores the visual discomfort, deformation, and unexpected shapes that emerge during image and 3D generation. Instead of treating those results as errors to be corrected, she asks whether they can be observed, selected, developed, and eventually valued as a form of creativity specific to AI.

A growing academic collaboration

Ayaka’s research is the latest development in an ongoing academic and creative relationship between Jissen Women’s University and Tripo AI, a relationship that has grown through workshops, student experimentation, cultural heritage projects, and academic research.

The connection between Tripo and Jissen Women’s University had already moved into formal research before Ayaka’s thesis. In a 2026 paper published in Jissen Women’s University’s Aesthetics and Art History journal, Professor Hajime Shimoyama documented an AR art workshop model in Soma, Fukushima, in which participant-made physical artworks were photographed, converted into 3D models with Tripo AI, exported as GLB files, and deployed through WebAR.

The paper describes how this workflow made it possible to move from a physical artwork to a 3D model and AR experience within the limited time of a public workshop. It also records a more unexpected observation: because a single photograph cannot show every side of an object, the AI has to infer what is missing, sometimes creating forms that do not exist in the original object. Rather than dismissing these “fluctuations” entirely as defects, the research considers whether they may reveal a creative value specific to the current stage of human-AI collaboration.

That question has become even more central in Ayaka’s work.

From Chōjū-giga to unpredictable forms

Ayaka began with Chōjū-giga, the famous Japanese picture scrolls known for their lively anthropomorphic animals. Her early experiments asked a deliberately strange question: what would happen if generative AI were asked to create creatures that do not belong in Chōjū-giga at all?

She tested animals and creatures ranging from toy poodles, Pallas’s cats, giraffes, and goldfish to dinosaurs, extinct species, marine life, insects, and mythical creatures. She also compared outputs from ChatGPT and Gemini, observing differences in style reproduction, composition, and narrative expression.

The most interesting results appeared when the models stopped following instructions cleanly.

Some generations introduced forms that Ayaka had never requested. As she continued generating images in sequence, she noticed that Gemini could be strongly influenced by the image immediately before it. Features gradually disappeared or mutated, and the visual identity of the original creature drifted over time. In her notes, she compares the process to a “game of telephone,” where information changes a little every time it is passed along.

This changed the direction of the research. AI was no longer only a tool for reproducing a style. It became something that interprets, transforms, and sometimes invents.

When the error becomes the method

Instead of trying to eliminate those strange outputs, Ayaka began collecting them.

From the large set of generated images, she selected creatures whose forms felt especially unpredictable or distinctive. These became the raw material for the next stage of the project.

She then returned the selected creatures to generative AI and asked it to turn them into characters. Through repeated experiments in ChatGPT and Gemini, the creatures began to acquire names, personalities, facial expressions, and three-view character sheets.

Here again, the process remained intentionally open to instability. Different conversation threads produced different interpretations. Previous generations could influence later ones. Some results followed instructions closely, while others introduced unexpected names, shapes, or character traits.

The goal was not to force all of these systems toward a single “correct” character. Ayaka was observing how the character continued to change each time another AI interpreted it.

From 2D interpretation to 3D interpretation

The next stage brings the research into three dimensions with Tripo AI.

This step is important for a reason that goes beyond simply making a 3D asset. A two-dimensional image only provides one visible interpretation of a creature. When Tripo turns that image into a three-dimensional form, it must infer volume, depth, structure, and parts of the body that were never shown in the source image.

For Ayaka, those inferences become another research opportunity.

If unexpected forms can emerge when an image-generation model interprets a prompt, what happens when a 3D-generation model is asked to interpret an already unpredictable 2D creature?

Her thesis positions this 2D-to-3D translation as a central part of the research. Unintended deformation during 3D generation can become another layer in the creature’s evolution, extending the same question that began with the first unexpected image.

The project is still in progress. Ayaka plans to continue developing the resulting characters through WebAR and 3D printing, allowing them to move from generated images into digital space and eventually into physical space.

“Raising” an AI creature

One phrase in Ayaka’s research materials captures the project especially well: “raising” the characters.

A creature begins as an accidental form in an AI-generated image. It is selected, interpreted again as a character, translated into three dimensions, and later placed into AR or materialized through 3D printing. At every stage, human decisions and machine interpretations shape what it becomes next.

The result is less like a conventional production pipeline and more like a process of cultivation.

This also connects Ayaka’s work to a broader line of inquiry emerging from the collaboration between Jissen Women’s University and Tripo AI. Across workshops, cultural heritage experiments, and student research, the same underlying question keeps returning: can the uncertainty of generative AI become something we study, design with, and even value, rather than something we always try to remove?

For Ayaka, the answer is still being developed through the work itself.

An ongoing student research project

Ayaka’s graduation research is currently ongoing. The Tripo 3D stage is in progress, with WebAR, 3D printing, and further presentation planned as the project develops.

That unfinished state is part of what makes the project worth following. Each new medium creates another opportunity for the character to change, and another opportunity to observe the relationship between human intention and AI interpretation.

As the academic collaboration between Jissen Women’s University and Tripo AI continues to grow, Ayaka’s project offers one example of how student-led research can move beyond learning how to use generative AI and begin asking deeper questions about what generative systems are actually doing when they create.

For now, the strange fish with legs is still growing.

Explore Tripo AI Education

Universities, faculty members, and student teams interested in using Tripo AI for courses, hackathons, research projects, or student creation can contact Tripo Education:

[email protected]