This study examines cultural imperialism in text-to-image (T2I) generative AI models by evaluating how these systems represent diverse cultural contexts. We analyze whether T2I models like DALL-E 3 reproduce biases by overrepresenting U.S. cultural norms while underrepresenting or distorting non-U.S. perspectives. Our methodology employs 280 prompts across 14 variable domains of life, translated into 15 languages, corresponding to 30 national/cultural contexts. To ensure cultural sensitivity, researchers from 12 different national backgrounds (including 7 from the Global South or East) collaboratively developed these prompts. We introduced country-specific references to test cultural adaptability, comparing generic prompts (e.g., "a living room") with location-specific ones (e.g., "a living room in Italy").

Our preliminary results using CLIP embeddings and cosine similarity measurements reveal that DALL-E 3 consistently defaults to U.S.-centric imagery when processing country-neutral prompts, regardless of the prompt language. Images generated from the non-location-specific prompts across all languages demonstrate stronger similarity to explicitly U.S.-referenced images than to any other national context. This finding supports the cultural imperialism hypothesis, indicating T2I models systematically encode and reproduce U.S.-American hegemonic influence in digital cultural representations. Our ongoing work includes human annotation to assess cultural accuracy and stereotyping across different contexts, with the hypothesis that representations of non-Western contexts will demonstrate lower cultural accuracy and higher stereotyping compared to Western contexts.

U.S. civil society is legally pushing back against the subsumption of data in mass for generative AI (genAI). This paper will explore the two avenues of legal redress (1) copyright law and (2) consumer privacy. In 2023 visual artists were among the first to file their copyright lawsuits against genAI companies, like StabilityAI, for copyright infringement in the form of image generator outputs. Later, prominent comedian Sarah Silverman filed a copyright lawsuit against OpenAI and Meta claiming that the companies trained their genAI products on her book without her permission or compensation while internet users in California were suing OpenAI for training their models on stolen private information including photographs. While copyrighted works are more clearly understood as the products of labor, I integrate consumer privacy to understand how the cultural fruits that Terranova (2000) argues are the products of free labor have been sidelined to create data hierarchies. While current literature has framed genAI around the “copyright dilemma” (Kuai, 2024), underlying all three cases is another critical question ripe for theoretical intervention: whose data is considered labor, and how is labor compensated for in an age of genAI? By integrating copyright lawsuits presented by visual artists and book authors with a class action lawsuit that relates to consumer privacy, this paper blends surveillance, copyright law, and labor to understand what is labor in an age of genAI. I argue these lawsuits strategically dissociate “data” from the user who created, posted, or circulated it.

A central question in social sciences today is how reliable complex predictive models like ChatGPT are. This study tests whether subtle changes in prompt structure affect sentiment analysis results by the GPT-4o mini model. Using 100,000 Spanish comments on four Latin American presidents, the model classified the comments as positive, negative, or neutral in 10 trials with slight variations in the prompts. The analysis revealed that even minor changes in prompt wording, syntax, or structure significantly impacted the classifications. In some cases, the model gave inconsistent responses or used languages other than Spanish. Statistical tests confirmed notable differences in most prompt comparisons, except when the structure was similar. These findings challenge the reliability of Large Language Models for classification tasks, revealing their sensitivity to prompt variations. The study also showed that unstructured prompts lead to more hallucinations. Trust in these models depends not just on their technical performance but also on the social and institutional contexts guiding their use.