How AI-Generated Prompts Shape Emotional Expression in Digital Drawing: Effects on Visual Complexity, Creative Engagement, and Arousal

Technical Summary: How AI-Generated Prompts Shape Emotional Expression in Digital Drawing

Problem Statement
While drawing is widely recognized as a medium for emotional externalization, existing literature lacks quantitative and computational approaches to explain how specific emotional states are translated into visual forms such as color, stroke, density, and spatial organization. Furthermore, the integration of Artificial Intelligence (AI) into creative activities raises questions about how AI-assisted prompting influences the generative process of emotional expression. Specifically, it remains unclear whether AI prompts alter the visual complexity and behavioral engagement of digital drawings, and whether the resulting features can serve as reliable indicators for emotion prediction.

Methodology
The study employed a mixed-methods design involving 31 undergraduate participants from Macau SAR, China, who completed 12 emotion-based digital drawing tasks (covering Joy, Anger, Sadness, Fear, Excited, Relaxed, Anxious, Depressed, Awe, Nostalgia, Confusion, and Contentment). Participants were randomly assigned to either an AI-assisted prompting condition or a control condition.

• AI Condition: Participants received AI-generated emotional descriptions, visual suggestions, compositional ideas, and stylistic references before each task.
• Control Condition: Participants received only the emotion label.

Data Collection and Feature Extraction:
A custom web-based platform recorded process-based behavioral traces, including color selection (HSL), brush size, stroke trajectories, drawing duration, and spatial coverage. From 369 initial records, 363 valid drawings were retained.

• Features Extracted: The study quantified color features (hue, saturation, lightness, diversity), stroke features (count, length, speed, points), and spatial features (area, centroid, dispersion).
• Complexity Metric: A composite "drawing complexity score" was constructed by standardizing and averaging stroke count, trajectory points, path length, drawing area, and color usage.

Statistical and Machine Learning Analysis:

1. Linear Mixed-Effects Models (LMMs): Used to examine the effects of AI assistance and emotion prompts on drawing complexity and behavioral engagement, controlling for participant-level random intercepts and repeated measures.
2. Machine Learning Classification: Supervised models (Logistic Regression, Random Forest, Extra Trees, Support Vector Machine) were trained to classify emotions. Performance was evaluated using Top-1/Top-3 accuracy and Macro/Weighted-F1 scores for both 12 fine-grained emotion categories and broader higher-order affective categories.
3. Prototype Development: An interactive web-based tool was developed to demonstrate real-time emotion prediction based on extracted drawing features.

Key Results

• Impact of AI Assistance: Participants in the AI-assisted condition produced drawings with significantly higher visual complexity and behavioral engagement compared to the control group. Specifically, the AI group exhibited:
  • Longer drawing durations (~63% increase).
  • Higher stroke counts (~71% increase).
  • More trajectory points (~80% increase).
  • Larger drawing areas (~54% increase).
  • Greater color diversity (~22% increase).
• Effect of Emotion Prompts: Emotion prompts significantly influenced subjective arousal ratings and specific trajectory movement features (path length and point density), particularly for "Anxious" prompts which elicited longer and denser trajectories. However, emotion prompts did not significantly alter the overall composite drawing complexity score.
• Machine Learning Performance:
  • Fine-Grained Classification: Performance for distinguishing the 12 specific emotions was limited (Top-1 accuracy ranged from 15.4% to 22.0%), indicating that visual features alone struggle to map precisely to discrete emotional categories.
  • Higher-Order Classification: Performance improved significantly when emotions were grouped into broader affective categories (e.g., positive/negative, high/low arousal). Random Forest achieved a Top-1 accuracy of 45.1% and a Top-2 accuracy of 72.5% in this context.

Significance and Claims
The paper claims that digital drawing features reflect general affective tendencies rather than serving as fixed, diagnostic symbols for specific emotions. The study positions the developed prediction tool as an exploratory research instrument rather than a clinical diagnostic system.

Key contributions include:

1. AI as Creative Scaffolding: The findings suggest AI-assisted prompting functions not merely as a content generator but as a scaffold that expands the visual and behavioral resources available during emotional externalization, encouraging more exploratory and complex drawing behaviors.
2. Process vs. Product: The research highlights that emotional expression in digital drawing is embedded in the behavioral process (stroke dynamics, duration, density) as much as in the final visual product.
3. Computational Affective Modeling: The study provides preliminary evidence that while fine-grained emotion prediction from drawing features is challenging due to individual differences and category overlaps, these features are more stable and effective for identifying broader affective states.

The authors emphasize that these findings are based on a specific sample and experimental design, and the prediction models should be interpreted as demonstrating the potential for data-oriented perspectives in art therapy and creativity research, rather than as definitive diagnostic tools.