By Lina Irawan 
When we stand before a painting—a funeral procession by Lippi, a fractured landscape, or an ink-wash mountain range—our emotional response is rarely singular. It is a messy, deeply human confluence of personal memory, cultural background, and immediate sensory input. Yet, in the summer of 2025, an ambitious project titled "EmoArt" sought to categorize the human experience of art into a digital dataset of 132,664 images. To accomplish this, researchers bypassed the nuanced, subjective judgment of human psychologists, opting instead to let GPT-4, a Large Language Model (LLM), handle the heavy lifting.
The result was an overwhelming, systematic flattening of human emotion: 55.95% of all artworks in the dataset were labeled as simply "calm."
This revelation is not merely a technical curiosity or a footnote in the history of computer vision. It is a warning about the silent, pervasive cultural hegemony embedded within the black boxes of generative AI. As we move toward a future where algorithms curate our museum experiences and dictate the metadata of our cultural history, we must ask: what happens when we replace the kaleidoscope of human feeling with an algorithmic "calm"?
Chronology: The Rise of EmoArt and the Quest for Automated Subjectivity
The EmoArt dataset was born out of a noble, if reductive, ambition: to train AI models for the purpose of art therapy. In mid-2025, Taiwanese researchers set out to map the emotional resonance of over 130,000 artworks. Recognizing that having a team of psychologists annotate every piece would be logistically impossible, the researchers turned to the most advanced LLM available at the time, GPT-4.
The process was straightforward: the model was tasked with assigning a single emotional label to each image from a fixed list of categories. The researchers reported a 91.47% alignment with human judgment. However, as independent analysis—conducted by data scientist Nastassia Shaveika—later revealed, this "alignment" was a symptom of a profound systemic bias rather than an achievement in emotional intelligence.
By the time the dataset was released, it had established a digital canon where the vast majority of human creative output was reduced to a singular, low-arousal state. This set the stage for a broader inquiry: how do these models "see" art, and why is their vision so consistently, and perhaps dangerously, monotonous?
The Data: The Architecture of Bias
To understand the scale of this "calm" phenomenon, one must look at the statistical distribution of the labels. A chi-square test, which measures how much observed data deviates from a random distribution, returned a value of 449,027. In the context of 132,664 samples, this result is not merely statistically significant—it is an indictment of the model’s decision-making process.
The pattern becomes clearer when viewed through the lens of Core Affect Theory. This psychological framework maps emotions onto two primary dimensions: valence (the positivity or negativity of an emotion) and arousal (the energy level associated with it). The data reveals that 87.9% of the artworks in the EmoArt dataset were assigned a positive valence, and 76.4% were assigned low arousal.
"Calm" sits exactly at the intersection of these two: happy, but not too happy; present, but not too engaged. It is the emotional equivalent of "elevator music"—a safe, neutral, and ultimately sanitized default that the AI defaults to when it lacks the cultural context to identify more complex states like Schadenfreude (German for joy in others’ misfortune) or Amae (Japanese for pleasurable dependence).
The "Orientalism" of Artificial Intelligence
The bias is not just emotional; it is deeply cultural. When the dataset is split by origin, the imbalance is stark. The model has been trained on a massive surplus of Western art, leading it to apply Western emotional categories to Eastern traditions that operate on entirely different aesthetic and emotional logics.
In Western contexts, red often signifies passion or danger. In many East Asian cultures, it is the color of celebration and prosperity. Similarly, while Western tradition might view black as a color of mourning or mystery, Japanese ink-wash painting uses it to denote mastery, discipline, and profound presence. The AI, having learned from a dataset skewed toward Western visual patterns, consistently fails to interpret these symbols correctly.
When analyzed for "entropy"—the variety of emotional labels applied to a specific style—the divide is clear. Western movements like Social Realism and Expressionism scored high on entropy, indicating that the AI felt comfortable applying a wide range of labels. Chinese ink paintings, conversely, scored an entropy of 0.35, effectively meaning the model was doing almost no classification at all; it was simply assigning "calm" as a blanket label.
Supporting Experiments: The Oracle Fallacy
To test if this was a failure of the specific EmoArt training or a broader issue with LLMs, an experiment was conducted using 23 diverse artworks analyzed by GPT-4 and GPT-5.1 (OpenAI), Claude Sonnet 4.5 and Haiku 4.5 (Anthropic), and Gemini 2.5 Flash (Google).
The findings were revealing. In the case of Gérrard Schneider’s Opus 110 (1968), a work of intense, kinetic abstraction, the model described it as having "profound quietude." In another case, Thomas Hart Benton’s Midwest (1931)—a painting depicting the grit of the Great Depression—the AI struggled between "excited," "tired," and "aroused," failing entirely to capture the underlying sense of communal resilience.
Most tellingly, when prompted to avoid "calm" or "contentment," the models often faltered, providing superficial explanations that highlighted their inability to parse visual metaphors. They are not "seeing" the art; they are predicting the next token in a string of text based on the statistical likelihood of what a human might say about a picture. They have no "emotion module." They have only a pattern-matching engine that, when pushed to describe something as complex as human grief or artistic ecstasy, retreats into the safety of the "calm" average.
Implications: The Automation of Cultural Perception
Why does it matter if an AI misidentifies a 15th-century painting? The stakes are significantly higher than an incorrect label in a museum database.
First, we are seeing the direct integration of these flawed systems into public life. The Cleveland Museum of Art’s ArtLens "Express Yourself" interactive is a prime example. By using facial-recognition software to categorize visitors’ expressions into fixed, AI-dictated labels, museums are effectively teaching the public to view their own emotions through a rigid, binary, and potentially inaccurate lens.
Second, we are witnessing the solidification of a "cultural monoculture." As AI becomes the primary tool for indexing, tagging, and retrieving images on the internet, the bias toward Western, low-arousal interpretations will become the default lens through which the world experiences global art. If our search engines, social media feeds, and educational platforms are all powered by models that prioritize a Western, sanitized version of emotion, we risk losing the ability to describe—and therefore to value—the specific, nuanced, and high-arousal emotional states that are central to non-Western artistic traditions.
Finally, there is the "Oracle Fallacy." By treating these models as objective arbiters of reality, we relinquish our critical capacity. When an AI tells us a painting is "calm," we are tempted to accept it as a neutral fact, rather than an algorithmic inference derived from a biased dataset.
A Call for Transparency
The research into EmoArt and subsequent model testing suggests that the issue is not necessarily the models themselves, but the way we deploy them. AI can be a powerful tool for accessibility and cataloging, but it cannot be allowed to stand in for the human experience.
If we continue to use these systems without acknowledging their inherent cultural and emotional biases, we will automate a world that is less diverse, less sensitive, and less capable of understanding the depth of human experience. We must demand "algorithmic humility." This means acknowledging that AI is not a neutral observer, but a reflection of the data it consumes.
Transparency is the only path forward. We must be honest about where these models come from, how they were trained, and what they fundamentally cannot do. If we treat AI as an oracle, we will inevitably be led to a digital landscape that is, in every sense, "calm"—a quiet, sterile, and ultimately hollow version of the vibrant, contradictory, and deeply emotional reality of human art.
The challenge ahead is not to make the AI "better" at feeling, but to keep the human capacity for critical judgment, debate, and subjective appreciation at the forefront of our cultural engagement. We must hold the mirror up to our machines, not to see them, but to see ourselves, and to ensure that the biases we have programmed into them do not become the permanent, unexamined architecture of our future.
Muslim 
Sagoh 
Reynand Wu 