Assessing the Geographic Diversity of AI's Platial Representations in Image Generation

A new study finds that AI image generators from OpenAI, including GPT and DALL-E models, can produce stereotypical and homogeneous representations of places, with older models sometimes showing more geographic diversity than newer ones. The research, submitted to arXiv on 28 Apr 2026, assessed how these models depict locations, revealing that the process of revising a text prompt yields greater geographic diversity than the final image generation step itself [1][2]. The authors, who drew inspiration from species diversity measures in ecology, incorporated similarity weighting into their analysis to quantify how varied the outputs truly are [2]. They selected GPT and DALL-E as state-of-the-art examples, noting that the AI ecosystem is becoming rapidly multimodal, making an examination of geographic diversity across image modalities critical [2]. DALL-E, first announced by OpenAI in January 2021, and its successors have been widely used for text-to-image generation, with DALL-E 3 later integrated into ChatGPT before being replaced by GPT Image in March 2025 [6][5]. GPT Image, a text-to-image variant of the GPT family, is native to ChatGPT and also available through Microsoft Copilot and Apple Intelligence [5]. The underlying GPT-4o model, released in May 2024, can process and generate text, images, and audio, and was retired from ChatGPT on February 13, 2026 [7]. The study found that despite producing lower-quality images, older models can exhibit greater geographic diversity, a counterintuitive result that challenges assumptions that newer systems are universally more capable [1][2]. The researchers also observed explicit model homogeneity, with the systems consistently depicting the same prototypical geo-specific feature or similar features for a given place [2]. This homogeneity is concerning, the paper states, because it risks producing stereotypical representations of places [2]. The findings connect to broader warnings about foundation models, which are trained on broad data and adaptable to many tasks, but whose defects can be inherited by all adapted downstream applications [4]. The study fills a research gap by extending geographic diversity evaluation, previously applied to AI chatbot outputs, into the image generation domain [2].