Images of Inequality: AI Created Depictions of Ancient Social Stratification

Abstrakti

The study of historical inequality has gained increasing attention in recent years, particularly in archaeology. Our understanding of the past is shaped by contemporary perspectives, especially
when examining concepts such as inequality, which were perceived differently in antiquity than they are today. This article explores how artificial intelligence (AI) contributes to the visualization of ancient inequality by generating images based on textual prompts. Focusing on the ancient Roman world, it examines the biases embedded in AI-generated images and their implications for historical understanding.

AI image generators, such as DeepAI and Adobe Firefly, produce visuals by drawing on extensive training datasets of images and captions. As AI-generated images become more prevalent – often used alongside textual narratives and oral presentations – they have the
potential to shape public and scholarly perceptions of historical subjects, including ancient inequality. This article addresses four key questions: 1. What do AI image generators produce
when given prompts related to antiquity and inequality? 2. What can the generated images reveal about the training data used in these AI models? 3. How useful are AI-generated images for studying ancient inequality and its reception? 4. What broader insights can these images offer for the study of ancient social structures?

To investigate these questions, this study analyzes AI-generated images of inequality using three levels of prompts: general (“ancient”), more specific (“Roman”), and highly specific
(“Pompeii”), and with different types of other definitions, such as social, wealth, and health inequality. The results reveal that AI-generated images primarily depict outdoor scenes featuring people and architectural elements, often including columns and similar supporting structures.

A key finding is that the visual styles of these AI-generated images reflect the biases of their underlying datasets. DeepAI tends to generate images resembling early modern and modern
paintings, suggesting that its training data associates these artistic traditions with depictions of ancient inequality. Firefly, by contrast, produces images with a more video game-like aesthetic,
likely influenced by social media and commercial sites.

Although AI-generated images offer valuable insights into how ancient material has been received and reinterpreted over time, they do not explicitly highlight ancient social hierarchies or inequality. Recognizing their relevance to historical inequality requires extensive knowledge of both ancient history and later artistic traditions. Moreover, AI-generated images inherently reflect and amplify the biases of their source material. Since much of the surviving ancient literature represents elite perspectives, these biases are embedded in the AI-generated reconstructions. Additionally, the training material – often in the form of historical paintings/pictures – likely embodies the preconceptions of the artists at the time the originals were created.
These images serve as starting points, which the AI modifies using the full range of textual connotations associated with the concepts it is prompted to generate, further compounding the layers of historical and cultural bias.