By Jia Lissa 
In the summer of 2025, a team of researchers in Taiwan unveiled a dataset that promised to revolutionize the intersection of technology and art therapy. Titled "EmoArt," the collection comprised 132,664 images, each tagged with a specific emotional label. The goal was noble: to train generative models to produce art capable of therapeutic healing. However, the methodology used to achieve this scale—outsourcing the emotional labeling to GPT-4—has instead revealed a profound and troubling limitation in how artificial intelligence perceives human experience.
Data analysis of the EmoArt collection reveals that an overwhelming 55.95% of all artworks are classified simply as "calm." This is not a mere statistical quirk; it is a systematic flattening of the human condition. While the researchers claimed a 91.47% alignment with human judgment, a deeper investigation suggests that AI models are not "detecting" emotions in art at all. Instead, they are engaging in a form of cultural projection, defaulting to a low-arousal, neutral state when faced with the complexity of visual expression.
The Illusion of Emotional Literacy
To understand why "calm" has become the default setting for AI-curated art, we must first discard the popular, Pixar-inspired notion that emotions are discrete, universal categories. In reality, human emotion is a constructive process. We do not experience "sadness" or "joy" like a sensor detecting a light level; we construct them through inference, personal history, and cultural context.
Concepts like the German Schadenfreude (joy derived from another’s misfortune) or the Japanese Amae (a sense of pleasurable dependence) have no direct English equivalents. Yet, when we force AI to map thousands of years of global art onto a rigid, Western-centric list of twelve emotions, we strip those works of their cultural soul.
The mathematical evidence for this bias is staggering. When applying a chi-square test to the EmoArt dataset, the resulting score of 449,027 confirms that this is not random distribution. The model is effectively mapping art onto a coordinate system of valence and arousal. Because the AI’s training data is heavily skewed toward Western aesthetics, 87.9% of the art is categorized as having "positive valence," and 76.4% is marked as "low arousal." The intersection of these two metrics is, by definition, "calm." It is a state that is happy, but never too happy; it is a sanitized, corporate version of human feeling.
A Chronology of Categorization
The path to this "calm" dominance can be traced through the evolution of vision-language models:
- 2024: The rise of massive multimodal training sets like LAION-5B brings unprecedented scale to AI vision, but researchers begin to note a distinct "Western bias" in the descriptors of images.
- Summer 2025: The EmoArt-130k project is released. Recognizing the infeasibility of employing hundreds of human psychologists, researchers utilize GPT-4 to handle the labeling, setting the stage for algorithmic bias to become institutionalized at scale.
- Late 2025: Initial user feedback suggests that AI-generated art for therapy feels "repetitive" and "uninspiring," a sentiment that now finds its mathematical justification in the 55% "calm" statistic.
- 2026: Independent audits—including recent experiments using multiple models like Claude 4.5, GPT-5.1, and Gemini 2.5—confirm that the "calm" default is not unique to a single model but is a systemic issue across the industry.
Data Analysis: The Weight of Cultural Imbalance
When examining the predictors of these labels, color and style emerge as critical variables, yet they remain tethered to Western interpretative frameworks. In Western tradition, red is often linked to alerts or passion. In Chinese aesthetics, however, red signifies celebration, weddings, and good fortune.
When researchers split the EmoArt dataset by origin, the results were jarring. The model exhibited seven times more color-emotion patterns for Western art than for non-Western works. This suggests that the AI has not "learned" the emotional language of the world; it has learned the emotional vocabulary of the West and is projecting that vocabulary onto the East.
The entropy levels—a measure of diversity in labeling—further illustrate this divide. Western art styles like Social Realism (1.89) and Expressionism (1.68) show a wide spread of emotional labels. In contrast, Chinese art and traditional ink styles languish with entropy scores as low as 0.35. For these works, the model is not classifying; it is essentially defaulting to a single, meaningless label.
The Experimental Verdict: A Failure to Connect
To test the robustness of this bias, a secondary experiment was conducted using 23 artworks, balanced for regional and stylistic diversity. The models were given two prompts: one asking for a primary emotion, and one explicitly forbidding the label "calm."
In many cases, the models struggled to move past their initial bias. When evaluating a piece of Chinese ink painting, most models defaulted to "calm" despite the work depicting complex landscapes or historical battle scenes. Even when pushed to explain their reasoning, the machines often resorted to "hallucinated" emotional states, such as "profound quietude," to justify their lack of genuine insight.
Perhaps most damning was the evaluation of a Thomas Hart Benton painting, Midwest (1931). While the EmoArt label tagged it as "excited," current models largely categorized it as "tired." Both labels miss the mark. The painting is a depiction of community during the Great Depression—a nuance of struggle, resilience, and solidarity that the AI failed to capture entirely.
Implications for the Future
The risks of this "calm" classification are not limited to art history databases. We are seeing these systems deployed in public spaces, such as the Cleveland Museum of Art’s "Express Yourself" interactive, which uses facial recognition to sort visitor expressions into rigid buckets.
If we allow these systems to become the arbiters of our emotional lives, we risk a "flattening" of the human experience. We are creating a feedback loop where AI models, trained on biased datasets, define what human emotion "looks like," and society, in turn, adapts to fit those narrow definitions.
This is a cultural harm of the highest order. When we treat AI as an oracle, we stop looking at the art—and each other—with the critical, soul-searching depth that true engagement requires. We are effectively delegating the interpretation of our cultural heritage to a machine that, by design, prefers the path of least resistance: a quiet, unbothered, and entirely empty "calm."
A Call for Transparency
The research into EmoArt is not a final verdict on the impossibility of AI in art, but it is a demand for radical transparency. As we integrate these models into museums, therapy clinics, and public spaces, we must ask: Are we building tools that help us understand our complexity, or are we building mirrors that reflect only the most convenient, algorithmic versions of ourselves?
If we continue to ignore the cultural biases inherent in our training data, we will not only lose the ability to appreciate the nuance of a masterpiece; we will lose the ability to recognize the depth of our own collective emotional landscape. The machine is telling us that the world is "calm." It is up to us to prove that it is, in fact, messy, vibrant, and profoundly complicated.
Muslim 
Sagoh 
Reynand Wu 