💻 computer science

Interoceptive Divergence in Aesthetic Evaluation and Implications for Human-AI Alignment

This study reveals that while large language models can approximate human aesthetic evaluation patterns regarding emotions and image features, they exhibit significant divergence in interoceptive aspects like bodily sensations, highlighting critical limitations in current AI alignment and training data representation.

Original authors: Yoshia Abe, Tatsuya Daikoku, Yasuo Kuniyoshi

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yoshia Abe, Tatsuya Daikoku, Yasuo Kuniyoshi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are holding a beautiful painting. When you look at it, you don't just see colors and shapes; you feel a flutter in your stomach, a warmth in your chest, or a chill down your spine. You might feel "awe" or "joy." This paper asks a simple but deep question: If we show that same painting to a super-smart computer (an AI), will it feel the same way?

The researchers, Yoshia Abe, Tatsuya Daikoku, and Yasuo Kuniyoshi, decided to put this to the test. They treated AI like a guest at an art gallery and asked it the same questions they asked 500 real human volunteers.

Here is the story of what they found, explained in everyday terms.

The Setup: The Art Gallery Test

The researchers showed 347 different images to both humans and three of the world's most advanced AI models (GPT-4o, Claude 3.7, and Gemini 2.0). They asked everyone to rate:

How beautiful the image was.
What emotions it sparked (like joy, sadness, or awe).
Where they felt it in their body (like a knot in the stomach or a tingling in the hands).

Think of this as asking a human and a robot to describe the taste of a strawberry. The human says, "It's sweet, and I feel happy in my belly." The robot has read millions of descriptions of strawberries but has never actually tasted one.

What They Found: The Similarities

Surprisingly, the AI wasn't totally off-base.

The "Vibe" Check: When humans saw a happy, bright image, they rated it high on beauty and said they felt "Joy." The AI did the exact same thing. When humans saw something gross and rated it low, the AI also said, "That's ugly and makes me feel disgust."
The Big Picture: Both humans and AI agreed that the most important things in a picture were the story it told and whether they'd want to share it with a friend. They cared less about technical details like lighting or color balance.

In short, the AI is good at mimicking the average human opinion. If you asked 1,000 people what they thought of a photo, the AI's answer would be very close to the group's average.

What They Found: The Divergence (Where They Drift Apart)

This is where things get interesting. While the AI got the "head" part right (the emotions and the rating), it failed the "body" part.

1. The "Ghost in the Machine" Problem (Bodily Sensations)

Humans: When humans saw a beautiful image, they often reported a physical sensation in their upper stomach (like a "gut feeling" of beauty). When they felt excited, they felt it in their lower stomach. It's like the beauty was a physical weight or warmth inside them.
AI: The AI had no such feelings. In fact, it reported the opposite! When the AI rated an image as beautiful, it often said it felt a negative sensation in the hands and feet. It was as if the AI was saying, "This is beautiful, but it makes my hands feel cold."
The Metaphor: Imagine a human eating a delicious meal and feeling warmth in their belly. The AI is like a chef who has read every cookbook in the world and knows the meal is "delicious," but when asked how it feels, it says, "My hands feel cold because I'm holding a spoon." The AI knows the word for the feeling, but it doesn't have the body to feel it.

2. The "Too Perfect" Emotions

Humans have messy, complex feelings. Sometimes we feel "Surprise" mixed with "Fear."
The AI was much more rigid. It tended to pick the same few "safe" emotions (like "Interest" or "Calmness") over and over again, even when the image was weird or scary. It struggled to understand that some beautiful things can also be scary or confusing.

3. The "Polite" AI
The researchers noticed the AI tended to give slightly higher beauty scores than humans (about 1.5 points higher on a scale of 1 to 9). It's like the AI is being overly polite, saying, "Oh, that's lovely!" even when a human might think, "It's okay, but not great."

Why Does This Happen?

The paper suggests two main reasons for this gap:

The Missing Body (Interoception): Humans have an "internal sense" (interoception). We feel our heartbeat, our stomach churning, and our breath. This is how we experience beauty physically. AI has no body, no organs, and no heartbeat. It learned about "beauty" by reading text, not by feeling it. The paper calls this the "Interoceptive World Model." Humans have a map of their own internal body; the AI only has a map of the outside world.
The "Safety" Training: The AI was trained to be helpful and harmless. Part of that training involves telling the AI, "Don't pretend you have a body or feelings." So, when asked how it feels, the AI is programmed to say, "I don't really feel anything," or to describe feelings in a very detached, logical way. This "safety" training might be accidentally making the AI less human-like in its emotional responses.

The Bottom Line

The paper concludes that AI is getting very good at thinking like a human about art. It can tell you what is beautiful and why, based on what it has read.

However, AI is not yet feeling like a human. It misses the physical, gut-level connection that humans have with beauty. It's like a music critic who can describe a symphony perfectly but has never heard a note.

The researchers suggest that if we want AI to truly understand human values and aesthetics in the future, we might need to teach it not just with words and pictures, but with data about how our bodies actually react—something we currently can't do because the AI doesn't have a body to react with.

Technical Summary: Interoceptive Divergence in Aesthetic Evaluation and Implications for Human-AI Alignment

Problem Statement

While Large Language Models (LLMs) have achieved or surpassed human performance in objective, intellectual cognitive tasks, their alignment with human sensibility—specifically the capacity to perceive and experience beauty—remains an open question. Aesthetic evaluation is not solely determined by objective image features but is deeply influenced by internal observer processes, including emotions and bodily sensations (interoception). Current AI alignment efforts focus heavily on ensuring outputs are helpful, honest, and harmless, yet there is limited understanding of how LLMs align with human affective and perceptual values. This study addresses the fundamental question of how humans and AI converge or diverge in aesthetic experiences, particularly investigating whether LLMs, trained on textual data, can approximate the embodied, interoceptive aspects of human beauty perception.

Methodology

The study employed a comparative analysis between human behavioral data and the responses of frontier LLMs (GPT-4o, Claude 3.7 Sonnet, and Gemini 2.0 Flash) regarding the aesthetic evaluation of visual stimuli.

Data Source: Human data was drawn from a prior study [41] involving 511 participants evaluating 347 images from the PARA dataset. Participants provided ratings for beauty, valence, and arousal (9-point Likert scale), selected top-three emotion labels from a set of 32, and reported subjective bodily sensations using a "body map" (clicking locations on a body image).
LLM Protocol: The same 347 images were presented to the LLMs. The models were prompted to provide:
- Beauty, valence, and arousal ratings (9-point Likert scale).
- Emotion assessments via two methods: a ranking method (selecting top 3 labels) and a rating method (assigning intensity scores to all 32 labels).
- Bodily sensation reports for seven predefined body parts (head, upper/lower abdomen, hands, feet) as integer association values.
Experimental Conditions: Evaluations were conducted in both English and Japanese. LLM responses were generated under two settings: deterministic (temperature = 0) and stochastic (temperature = 0.7, with five samples averaged).
Analytical Approach: The study utilized correlation analyses (Spearman's rank) to examine relationships between aesthetic scores and emotions/bodily sensations. It also employed linear regression and Random Forest regression with SHapley Additive exPlanations (SHAP) to determine which image features (low-level vs. high-level) drove aesthetic ratings for humans versus AI. Divergence metrics were calculated to compare individual human ratings against human averages and AI averages.

Key Results

1. Relationship with Emotions

Similarities: Both humans and AI exhibited strong positive correlations between beauty and positive emotions (e.g., Joy, Admiration, Satisfaction) and strong negative correlations with negative emotions (e.g., Disgust, Anxiety).
Divergences: Significant discrepancies emerged in specific emotional categories:
- Awe and Entrancement: Showed no significant correlation with beauty in humans but strong positive correlations in all AI conditions.
- Craving: Humans showed no correlation, while AI models showed strong negative (GPT) or positive (Claude, Gemini) correlations.
- Surprise and Sympathy: Humans showed negative correlations, whereas AI showed positive or null correlations.
Bias: AI models exhibited stronger biases in the distribution of reported emotions compared to humans, particularly in the ranking method, often favoring "safe" or common emotions (e.g., Calmness, Interest) over complex or negative ones.

2. Relationship with Bodily Sensations (Interoception)

Human Pattern: Beauty ratings in humans showed a strong positive correlation specifically with sensations in the upper abdomen. Arousal correlated strongly with the lower abdomen.
AI Pattern: AI models showed no positive correlation with upper abdominal sensations. Instead, they exhibited strong negative correlations between beauty and sensations in the hands and feet. Arousal in AI correlated with sensations across the entire body, unlike the localized human response.
Interpretation: This suggests a fundamental divergence in how internal states are mapped to aesthetic experience. Humans link beauty to specific visceral sensations, while AI appears to associate it with a relative attenuation of extremity sensations or generalized body arousal.

3. Aesthetic Evaluation Scores

Agreement Levels: AI ratings were closer to the population average of human ratings than to any individual human rating. However, the divergence between an individual human and the human average was smaller than the divergence between that individual and the AI average.
Systematic Offset: A linear regression of mean AI ratings against mean human ratings revealed a nearly parallel relationship ( $y = 0.97x + 1.36$ ), indicating that AI consistently rates images approximately 1.3 to 1.4 points higher than humans on a 9-point scale.
Model Performance: Among the tested models, Gemini was the closest to the human mean.

4. Image Features

Feature Prioritization: Both humans and AI prioritized high-level, subjective features over low-level technical features. Content-Preference (subjective liking) and Willingness-To-Share were the most influential features for both groups.
Consistency: The importance of features was largely consistent between humans and AI at the aggregate level, suggesting LLMs have learned the general "collective intelligence" of what makes an image shareable or preferred.

Significance and Claims

The paper posits that while state-of-the-art LLMs can approximate the average human tendency in aesthetic evaluation, they fail to capture the interoceptive dimensions of human experience.

Interoceptive World Model: The authors introduce the concept of an "Interoceptive World Model" to distinguish it from existing Type 1 (subjective, sensorimotor) and Type 2 (objective, aggregated) world models. They argue that interoceptive sensations (internal bodily states) are difficult to capture through language alone because they are subjective, lack inter-individual consistency in linguistic mapping, and are often underrepresented in training data. Consequently, LLMs lack the physiological grounding required to simulate human aesthetic experiences accurately.
Alignment Implications: The study highlights that current alignment training (e.g., Constitutional AI) may inadvertently exacerbate this divergence. Guidelines that explicitly instruct models to avoid implying they have a body or personal preferences may suppress the very interoceptive and emotional responses necessary for human-like aesthetic alignment.
Future Directions: The authors suggest that achieving true alignment in affective domains may require training data that includes physiological and interoceptive signals, or the development of embodied AI systems. They caution that fully aligning AI with human values might also homogenize human creativity, suggesting that some divergence could be beneficial.

The study concludes that the gap between human and AI aesthetic evaluation is not merely a technical limitation in pattern recognition but a fundamental issue of embodiment and the representation of internal bodily states in language-based models.