By Evan Lee Salim 
When we gaze at a painting, we are rarely met with a single, dictionary-defined feeling. A funeral scene, for instance, might evoke a chaotic collision of grief, relief, longing, and despair. Yet, when we task modern artificial intelligence with interpreting the emotional landscape of art, it often strips away this nuance, flattening the human experience into a singular, predictable label: "calm."
This phenomenon, while seemingly benign, signals a deeper, more systemic issue in how generative AI models perceive, categorize, and ultimately enforce cultural biases under the guise of objective analysis.
The Genesis of "EmoArt" and the Automation of Subjectivity
The story begins in the summer of 2025, when a team of researchers in Taiwan unveiled "EmoArt," a massive dataset comprising 132,664 individual artworks. The stated goal was ambitious: to train machine learning models to generate art for therapeutic purposes, theoretically allowing AI to provide a bespoke emotional anchor for patients.
Faced with the logistical impossibility of commissioning thousands of hours of work from psychologists to manually annotate every piece, the researchers turned to the industry standard: GPT-4. By automating the labeling process, the researchers claimed an impressive 91.47% alignment between their AI labels and human perception. However, this statistical victory conceals a fundamental failure in understanding the complexity of human emotion.
The resulting data reveals that a staggering 55.95% of all artworks in the dataset were categorized as "calm." This was not merely a distribution error; a chi-square test conducted on the dataset returned a value of 449,027, an unambiguous indicator that the dominance of "calm" was a systemic bias within the model’s decision-making architecture, not a reflection of the actual diversity of the art itself.
Chronology of a Cultural Projection
To understand why AI gravitates toward "calm," we must look at the mechanics of core affect theory, which maps emotions across two dimensions: valence (positive vs. negative) and arousal (energetic vs. not). The data suggests that GPT-4 perceives 87.9% of art as having a positive valence and 76.4% as having low arousal. The intersection of these two variables is, by definition, "calm."
The "Orientalism" of Algorithms
The bias becomes more pronounced when one examines the regional origin of the art. Applying Edward Said’s framework of "Orientalism"—the historical tendency of the West to define the "East" as a monolithic, mythical, and exoticized other—reveals that AI models are effectively projecting Western emotional frameworks onto Eastern art traditions.
In the dataset, Western art styles showed significantly higher "entropy," meaning the model was able to apply a wider variety of emotional labels. Conversely, Eastern traditions—such as Chinese ink painting, Gongbi, or Korean art—were frequently dismissed by the model as "calm." Chinese ink paintings, which encompass complex narratives of political satire, landscapes, and battle scenes, were tagged as "calm" 89.4% of the time. The entropy score for these works was as low as 0.35, indicating that the model was essentially failing to perform any meaningful classification at all.
The Color Bias
The model’s inability to grasp cultural nuance is further evidenced by its reliance on color. In Western tradition, red is often associated with alert, passion, or danger. In many East Asian cultures, however, red is the color of celebration, good fortune, and prosperity. Similarly, while black may represent mystery or death in the West, it is the fundamental medium of mastery and discipline in Japanese calligraphy. Because the model was trained primarily on Western-centric datasets like LAION-5B, it lacks the contextual literacy to interpret these colors through a non-Western lens.
Supporting Data: An Experimental Deconstruction
To verify whether this bias was specific to the EmoArt researchers’ implementation, an independent experiment was conducted using 23 diverse artworks analyzed by five different state-of-the-art models: GPT-4, GPT-5.1, Claude Sonnet 4.5, Claude Haiku 4.5, and Gemini 2.5 Flash.
When asked to categorize these works, the models consistently fell back on "calm" and "contentment" unless explicitly told not to. In one instance, when analyzing a painting of farmers during the 1930s Great Depression, the AI struggled. While one model identified the physical strain and "tiredness" of the subjects, another opted for "aroused" (in the psychological sense of heightened activation), and a third labeled it "excited." The inconsistency highlights that when an AI is pushed outside of its default comfort zone, its output becomes unstable and disconnected from the historical or emotional context of the image.
Official Responses and Theoretical Perspectives
The implications of this are not merely academic. Many institutions, such as the Cleveland Museum of Art, have deployed "ArtLens" interactives that use facial-recognition software to map visitor expressions to fixed emotion labels. These systems rely on the same flawed logic: the assumption that human emotion can be captured by a "sensor" and reduced to a singular, discrete category.
Psychologists have long argued that emotions are not discrete biological triggers like the characters in a Pixar film; they are constructed through inference, context, and prior experience. There is no universal "sadness face" or "calm painting." By attempting to automate these labels, developers are creating a "digital feedback loop" where AI-generated descriptions reinforce the biases of the training data, effectively narrowing the scope of human aesthetic experience to a sanitized, Western-normative baseline.
The Societal Implications of Automated Subjectivity
The risks here are not catastrophic in the traditional, science-fiction sense of AI "taking over." Rather, the threat is cultural and cognitive. If we begin to accept AI’s classification of art as an oracle of truth, we risk losing the very thing that makes art vital: its ambiguity.
1. The Erosion of Cultural Diversity
When an algorithm labels thousands of years of Eastern artistic tradition as "calm," it effectively erases the narrative depth of those works. This is a form of digital colonization, where the machine’s training data forces global art into a singular, Western-centric emotional box.
2. The Standardization of Feeling
If the tools we use to navigate our cultural lives—from museum apps to social media algorithms—are tuned to a model that equates everything "positive" and "low-arousal" as "calm," we may see a gradual shift in how we ourselves describe our own experiences. We are being nudged toward a simplified emotional vocabulary.
3. The "Black Box" of Accountability
As seen in the experiment, models are highly sensitive to prompts. When asked to "reconsider," GPT-4 completely abandoned its previous "calm" labels. This suggests that the AI is not actually "feeling" or "interpreting" the art at all; it is simply predicting tokens based on the highest probability of a "correct" response in its training set. When we treat these models as authorities, we are ceding our ability to define our own emotional reality to a probabilistic engine.
Conclusion: A Call for Transparency
The research conducted by Nastassia Shaveika and others highlights a critical, often-overlooked dimension of AI ethics. We are currently living in an era where we are building systems that act as arbiters of human experience without having any human capacity for experience themselves.
The "calm" default is a symptom of a larger, more urgent problem: the automation of cultural interpretation. We cannot "fight" the existence of these models, but we can demand honesty about their limitations. Developers must be transparent about the biases inherent in their training sets, and institutions must be wary of deploying emotion-recognition technology in contexts where the nuances of human experience are at stake.
If we do not, we will continue to look at the world through a mirror designed by machines—a mirror that, for all its processing power, can only ever tell us what it expects us to see. We must hold the mirror up to our own biases before we automate them, lest we find ourselves living in a world that is "calm," simply because that is the only language our machines know how to speak.
Muslim 
Nila Kartika Wati 
Laily UPN 