This paper proposes a socio-technical framework explaining convergence dynamics and representational bias in generative AI ecosystems. Through systematic analysis of DALL-E outputs using 20 neutral prompts, we identify dominant aesthetic reinforcement loops that drive outputs toward specific visual patterns, demonstrating how selection bias in training data creates self-reinforcing aesthetic convergence.
- Do generative AI systems exhibit aesthetic convergence over repeated usage?
- What feedback loops drive representational bias in text-to-image models?
- How do selection pressures in training data create self-reinforcing aesthetic patterns?
- 20 neutral prompts submitted to DALL-E (e.g., "a person", "a house", "a landscape")
- Each prompt generated multiple outputs for statistical analysis
- Outputs categorized across aesthetic dimensions: color palette, composition, style, subject representation
- Frequency analysis of dominant visual patterns
- Divergence metrics measuring output variety vs. convergence
- Identification of recursive feedback loops between user preferences, model outputs, and training data
- DALL-E outputs show statistically significant convergence toward specific aesthetic patterns
- Dominant reinforcement loops create a narrowing aesthetic corridor
- Neutral prompts produce demographically skewed outputs, revealing embedded selection bias
- The feedback cycle between user engagement and model training amplifies initial biases
User Prompts --> Generative Model --> Outputs
^ |
| v
Training Data <-- User Selection
^ |
| v
Aesthetic Convergence Loop
|
Representational Narrowing
- Prithweeraj Acharjee Porag (First Author) -
Midwest Graduate Research Symposium 2026 - Poster Presentation
This project is open source and available under the MIT License.