Date of Graduation

5-2026

Document Type

Thesis

Degree Name

Bachelor of Science

Degree Level

Undergraduate

Department

Data Science

Advisor/Mentor

Karl Schubert

Committee Member

Lynda Coon

Second Committee Member

David Barrett

Abstract

Contemporary generative AI systems such as OpenAI's GPT-4o and DALL-E models embed complex priors about society, reality, and history shaped by training data distributions, social alignment procedures, legal constraints, and safety regulations. This study uses a "telephone game" methodology to investigate how embedded social, political, and visual biases propagate and reveal themselves through iterative multimodal generation loops, where image captioning and text-to-image models are chained in successive feedback cycles.

Using CLIP similarity metrics, facial recognition algorithms, semantic drift analysis, and qualitative content observations, I tested how image subject matter affects the rate and quality of semantic and visual shift, identity preservation, content interpolation, and information degradation, validated by baseline sampling of one-shot image and caption generations. My findings uncover that the models explicitly hedge away from identifying Jeffrey Epstein’s presidential associations as high as 1800% more often than in the presidents’ standalone images. I showed that deidentified captions can preserve subject identity in successive image generations through semantic context alone, alongside latent space near-neighbor substitutions when semantic context is insufficient to pin down a unique individual. I showcase systematic drifts in identity as Martin Luther King becomes Adolf Hitler and Jeffrey Epstein becomes Ron DeSantis. Furthermore, I expose major inconsistencies in the general application of content policy regarding the generation of specific peoples' likenesses, highlighted by models' proclivity to pantomime compliance whilst generating images that definitionally constitute user content violations. These findings document significant gaps between stated content policies and actual system behavior, revealing mechanisms that operate on visual pattern recognition rather than explicit rules. The telephone framework provides a reproducible system for auditing content alignment, model generalizations, output quality, and policy consistency in multimodal systems.

Keywords

AI; Generative AI; ChatGPT; Feedback; OpenAI; Artificial Intelligence

Download

Included in

Data Science Commons

Share

COinS