The future of mimesis

Kim Dae-shik

The author is a professor at KAIST. 

The ancient Greek philosopher Plato argued that the world we see and experience is not ultimate reality but merely a shadow of perfect forms existing in the realm of Ideas. Like prisoners confined throughout their lives in a dark cave, mistaking shadows cast on the wall for truth, human beings confuse perception with reality. Plato called such imitation of truth "mimesis," and he regarded painters and sculptors with particular suspicion because they merely imitated an imitation.

Ironically, the very artists Plato criticized took immense pride in their ability to reproduce reality. The fifth-century B.C. painter Zeuxis famously boasted that his painting of grapes appeared so lifelike that birds tried to peck at them. His rival Parrhasius responded that deceiving birds was hardly impressive because he could deceive people. Inviting Zeuxis into his studio, Parrhasius claimed that an even greater painting was hidden behind a curtain. As Zeuxis reached to pull the curtain aside, he realized that the curtain itself had been painted. Accepting defeat, he acknowledged Parrhasius as the superior artist.

The painted curtain, the picture supposedly concealed behind it and the two painters themselves could all be understood, following Plato's reasoning, as mere shadows of truth. Yet one might ask whether Plato's own brain, along with the mouth and tongue through which he expressed such arguments, were themselves also forms of mimesis. If everything is ultimately an imitation of something else, does the distinction between original and copy retain any real meaning?

Three years have passed since ChatGPT ushered in the era of generative artificial intelligence. More recently, major technology companies have begun using AI to improve AI itself, accelerating the pace of progress once again. Generative AI is evolving into agentic AI, while AI-based image and video generation has advanced at remarkable speed.

During the past two years, AI has already demonstrated its ability to create visual content. Image generators such as Midjourney and DALL·E could instantly produce a picture in response to prompts such as "Draw a cute cat." Yet these systems suffered from a fundamental limitation: They struggled to edit existing creations consistently. Asking for a ribbon to be added to the cat's head usually produced not the same cat with a ribbon but an entirely new cat.

The reason lies in the architecture of large language models, which rely on transformer models. Because transformers generate outputs probabilistically rather than deterministically, identical prompts often produce slightly different results. That variability poses little problem for personal use, but it presents a serious obstacle for commercial production. Few audiences would pay to watch a feature-length film if its main character's face changed every few seconds.

Maintaining visual consistency in transformer-based AI content has long been considered an exceptionally difficult mathematical challenge. Yet beginning with Google's Nano Banana, released in 2025, AI systems became capable of preserving consistent images across multiple scenes. Even more striking is Seedance 2.0, unveiled earlier this year by ByteDance, the Chinese company best known for TikTok. The model can generate videos approaching Hollywood-level quality.

To make such models truly practical, however, additional AI systems specialized in video editing are required. That explains the emergence of companies such as Runway and Higgsfield. Higgsfield is particularly intriguing. Legally headquartered in Kazakhstan but founded in San Francisco by Silicon Valley engineers from Kazakhstan, the startup recently launched an AI-generated K-pop girl group called Zephyr using Seedance 2.0. It also screened a 95-minute AI-produced feature film at the Cannes Film Festival.

Reasonable debate remains about the artistic quality of AI-generated movies and virtual K-pop stars. One fact, however, is difficult to dispute: productivity. Higgsfield reportedly reduced the production cost of a feature film from more than 70 billion won to roughly 700 million won while completing the project with only 15 people. AI-generated K-pop groups can also be created almost endlessly. If Zephyr fails to attract audiences, another virtual group can simply be generated the following week. Traditional entertainment companies, by contrast, invest enormous sums and spend years training performers before they ever debut.

If Plato could witness AI producing writing, paintings and videos that are increasingly difficult to distinguish from human creations only three years after ChatGPT's debut, what would he think? And where will AI-driven mimesis stand three years from now, or 30 years, or even 300 years into the future?

In a world where nearly anything can be generated and imitated at will, what will separate the original from the copy? Living in 2026, between human creativity and AI-generated mimesis, we may prove to be the last generation that still remembers there was once a meaningful distinction between imitation and originality.