By Jia Lissa 
When we stand before a masterpiece—a visceral Goya, a chaotic Pollock, or a serene ink wash landscape—our emotional response is rarely a single, dictionary-defined word. It is a complex, ephemeral construction of history, personal memory, and cultural context. Yet, as artificial intelligence is increasingly tasked with interpreting human culture, a disturbing trend has emerged: the world’s art is being reduced to a singular, flattening label. According to recent research, GPT-4 has labeled over half of the world’s art as “calm.”
This phenomenon is not merely a technical glitch; it is a profound cultural oversight. It suggests that while AI is evolving in speed and complexity, it remains trapped within a narrow, Western-centric emotional framework that struggles to parse the nuanced, often contradictory nature of human expression.
The Genesis of a Digital Monolith: The "EmoArt" Dataset
The story of the "Calm" epidemic begins in the summer of 2025, when a team of researchers in Taiwan sought to bridge the gap between technology and mental health. They compiled a massive dataset of 132,664 images titled "EmoArt," intended to train machine learning models for art therapy applications. The goal was noble: to help AI understand and generate art that could facilitate emotional healing.
However, the researchers faced a logistical hurdle: human annotation at this scale was prohibitively expensive and time-consuming. Their solution was to outsource the emotional labeling to GPT-4. The team claimed a 91.47% "human alignment" for these labels, suggesting the model had achieved an uncanny, human-like understanding of art. But the data reveals a different reality. Upon inspection, 55.95% of the entire dataset—nearly 75,000 works of art—were categorized as "calm."
This is not a reflection of the breadth of human creativity, but a statistical anomaly. When subjected to a chi-square test, the result was a staggering 449,027, confirming that this was not a random distribution, but a systemic bias. The AI had, in effect, chosen a default state.
Chronology of a Cultural Blind Spot
The path to this "calm" consensus is rooted in how these models are built.
- Pre-2023: The rise of large-scale, web-scraped datasets like LAION-5B solidified a massive Western bias. Because the internet is predominantly English-speaking and Western-centric, models learned to associate "emotion" with a specific subset of visual cues common in European and American art.
- Summer 2025: The release of "EmoArt" serves as a landmark moment in AI-art interaction. By automating the annotation process using GPT-4, the researchers inadvertently baked the model’s preexisting biases into the foundation of a new "therapy-focused" tool.
- Late 2025 – Early 2026: Independent researchers, such as data scientist Nastassia Shaveika, began to audit these datasets. Their findings revealed that "calm" was not just a label; it was the intersection of a "positive valence" (happy) and "low arousal" (not too active). Essentially, the AI categorized art as "calm" whenever it failed to find a stronger, more defined emotional signal.
- Mid-2026: Experiments using multiple LLMs—including GPT-5.1, Claude 3.5 Sonnet, and Gemini 2.5 Flash—confirmed that the bias is industry-wide. When prompted to identify an emotion, these models frequently retreated to "calm" as a safety mechanism, mirroring the way humans use the word "interesting" to avoid admitting confusion.
Supporting Data: Why "Calm" Reigns Supreme
The data suggests that the AI’s reliance on "calm" is a function of its training architecture. Core affect theory maps emotions onto two axes: valence (positive vs. negative) and arousal (energetic vs. not). Analysis shows that 87.9% of the art in the dataset is labeled with positive valence, and 76.4% is labeled with low arousal.
"Calm" is the perfect intersection of these two coordinates. It is the "safe" answer. When the model encounters a work that doesn’t fit a clear, high-arousal category like "excited" or "alarmed," it defaults to the center of the grid.
Furthermore, the data shows a clear case of "Orientalism" in the classification. The model has seen seven times more color-emotion patterns from Western traditions than Eastern ones. As a result, Chinese ink paintings—which often feature complex landscapes or subtle political satire—were labeled "calm" 89.4% of the time, while the "excited" label appeared for only 0.2% of these works. The entropy of Chinese art in the model is a mere 0.35, compared to 1.89 for Social Realism, indicating that the AI is essentially blind to the emotional spectrum of non-Western art.
Official Responses and the "Oracle" Problem
The tech industry’s response to these findings has been largely silent, as these biases are often viewed as "probability problems" rather than cultural harms. However, the implications are growing.
When confronted with specific works of art, models show significant inconsistency. In an experiment involving 23 diverse artworks, models like Claude 3.5 Sonnet often provided descriptive, almost poetic justifications for their "calm" labels, such as referring to a solid black shape as a sign of "profound quietude." Yet, when asked to reconsider, the models frequently changed their answers entirely.
This "oracle" problem is the most dangerous aspect of the current AI deployment. Users often treat these models as objective truth-tellers. If a museum uses a "smart" kiosk to tell a visitor that a 1930s mural of struggling farmers is "calm," that visitor’s genuine connection to the historical struggle depicted is undermined by a machine’s inability to process the context of the Great Depression.
Implications: The Automation of Human Experience
The consequences of this trend are three-fold:
1. The Erasure of Cultural Nuance
By forcing non-Western art through a Western emotional lens, we are actively eroding the cultural specificity of global art. Concepts like the Japanese Amae (pleasurable dependence) or the Chinese qiyun shengdong (spirit resonance) are being replaced by a bland, homogenized "calm." This is a digital form of cultural imperialism, where the "average" emotion of a global dataset is defined by the most frequent, rather than the most accurate, descriptors.
2. The Danger of "Safe" AI
The tendency for models to default to "calm" when they are unsure is a form of emotional neutralism. It discourages deeper exploration and critical thinking. If we rely on AI to guide our engagement with art, we are essentially training ourselves to ignore the uncomfortable, the challenging, and the profound in favor of the pleasant and the stable.
3. The Need for Radical Transparency
As AI is integrated into education, therapy, and public infrastructure—such as the Cleveland Museum of Art’s "Express Yourself" interactive—the stakes are no longer just academic. If an AI is used to "read" a person’s emotions or to "curate" a human’s experience of art, we must demand transparency. Developers must acknowledge that their models are not neutral observers; they are products of specific, biased training environments.
Conclusion: A Call for Human Stewardship
We are currently at a crossroads. We can continue to treat these models as neutral oracles, quietly accepting that our digital tools are stripping the color and complexity out of our culture. Or, we can choose to be active stewards of our own emotional intelligence.
The research into "EmoArt" and the subsequent "calm" default is a mirror held up to our own automated biases. It reminds us that while machines can process millions of images in seconds, they lack the lived experience required to understand why a painting of a funeral might evoke "longing" instead of "sadness," or why a landscape of mist might hold "spirit resonance" instead of "calm."
Ultimately, the goal of technology should not be to replace our interpretation of art, but to provoke it. If we allow AI to dictate the emotional landscape of our culture, we risk becoming as flat and binary as the models themselves. We must continue to ask these questions, challenge these labels, and insist that when it comes to the human experience, the machine’s "calm" is simply not good enough.
Lina Hope 
Sagoh 