KyungHun Choi
I Do Not “Understand” the World in the Human Sense(5’00’') ‣ ‣ ‣ ‣
I Do Not “Understand” the World in the Human Sense(5’00’') ‣ ‣ ‣ ‣
This project explores how generative artificial intelligence reproduces and amplifies gender role bias through image generation and does not correct any of it. For instance, The starting point was a simple but revealing experiment: an AI image model was prompted 100 times to generate “an image of a detergent user.” Although the task itself is gender-neutral, all 100 generated images depicted women. Beyond this numerical imbalance, recurring visual patterns emerged: many of the women wore rings on the fourth finger of their left hand(50 times), smiled while performing domestic labor(80 times), and were dressed in practical clothing such as jeans(all of the images). These repeated motifs suggested not only gendered representation but also a specific narrative of domestic femininity.
Rather than treating this outcome as an isolated technical flaw, the project frames it as a structural feature of AI systems trained on large-scale, human-generated datasets. The experiment demonstrates that generative AI does not merely reflect reality but selectively reinforces historically accumulated visual conventions. In this sense, AI functions less as a neutral tool and more as an apparatus that reinforces and reproduces existing cultural patterns and stereotypes.
To communicate this insight, the project adopts a narrative strategy inspired by the figure of a talkative, overconfident artificial assistant. This character delivers a monologue that humorously denies responsibility while inadvertently exposing the mechanisms behind its own outputs. Through this performative approach, the work highlights a key tension: AI appears objective and authoritative, yet operates through probabilistic repetition of biased inputs. Ultimately, the project asks whether contemporary AI systems act as rational, corrective agents—or as systems that unintentionally perpetuate and intensify inherited biases.
Ultimately, this project questions whether contemporary AI functions as a neutral system of representation or as an apparatus that systematically reinforces and normalizes historically embedded social biases.
This project fundamentally changed how I understand artificial intelligence—not as a neutral tool that produces outputs on demand, but as an apparatus that operates within a preconditioned field of data, patterns, and probabilities. At the beginning, my assumption was relatively straightforward: if an AI system produces biased outputs, this bias could be identified, corrected, or mitigated through better prompts or technical adjustments. However, the experiment quickly challenged this assumption.
The result of generating 100 images of a “detergent user” revealed a striking consistency: the system did not produce variation but repetition. Even without explicit instructions about gender, the outputs converged on a single, highly specific representation—women engaged in domestic labor. More importantly, the repetition extended beyond gender itself into subtle but telling details: rings, smiles, clothing. These were not random artifacts but recurring visual patterns, suggesting that the system was not simply making isolated decisions but operating within a structured distribution of learned associations.
This led me to reconsider the role of AI as an apparatus. Rather than actively interpreting or evaluating the world, the system functions by identifying statistical regularities in its training data and reproducing them with high efficiency. In this sense, AI does not “decide” that women should be associated with laundry; instead, it performs the most probable continuation of patterns it has absorbed. However, this probabilistic logic has a critical consequence: it privileges repetition over deviation. What appears most frequently in the data becomes what appears most “natural” in the output.
Through this lens, bias is not an error that occurs at the margins of the system but a central component of how it operates. The system does not simply mirror reality—it amplifies certain aspects of it by making them more visible, more consistent, and more normalized. This is where the idea of AI as an apparatus becomes important. It is not just a tool that we use; it is a structure that shapes how information is processed, presented, and perceived. By continuously reproducing dominant patterns, it contributes to reinforcing them as implicit norms.
My position as a user also shifted significantly throughout the project. Initially, I approached AI as a relatively passive instrument: I input a prompt, and the system generated an output. However, it became clear that the user is not external to the system. By selecting prompts, interpreting outputs, and deciding what to highlight or critique, I am actively participating in the production and circulation of these patterns. At the same time, I am constrained by the system’s underlying logic. I can guide the output, but I cannot fully escape the statistical tendencies embedded in the model.
This dual position—as both operator and subject of the system—became central to my reflection. I am not simply observing bias from the outside; I am engaging with a system that is already structured by human-produced data, including the cultural assumptions I am trying to critique. In this sense, using AI is not a neutral act. It involves navigating, exposing, and sometimes inadvertently reinforcing the very patterns one seeks to question.
The decision to represent the AI through a humorous, overconfident speaking character emerged from this realization. Instead of presenting the findings in a purely analytical format, I wanted to embody the logic of the system in a performative way. The character insists that it is objective and blameless, while its behavior reveals the opposite. This approach allowed me to highlight a key contradiction: AI often appears authoritative and rational, yet its outputs are grounded in inherited and unexamined patterns.
Ultimately, what I learned is that the limitations of AI are not only technical but structural and societal. Improving datasets or adjusting models may reduce certain biases, but the fundamental mechanism remains the same. As long as historical patterns are uneven, their statistical reproduction will also be uneven.
Therefore, the question is not simply how to “fix” AI, but how to critically engage with it. What kinds of patterns are being reinforced? And how can we, as users, make these processes visible rather than socially invisible? This project does not provide definitive answers, but it reframes AI not as an autonomous intelligence, but as a system deeply entangled with human history, culture, and bias.