Human asymmetries in AI art: Syntax and writing direction effects on agent position in AI-generated images

The present study investigates positional patterns in visual representations generated by two artificial intelligence (AI) models in response to textual prompts describing interactions between two animate entities. The primary objective is to assess whether the syntactic structure of a given sentence influences the spatial positioning of the agent (i.e., the entity performing the action) within the generated image. The study follows research showing that in art produced by humans, positioning of agents on the picture depends on reading-writing direction: entities mentioned first are positioned on the left side by people from cultures with left-to-right writing script disproportionately more often than on the right side. We prompted FLUX and DALL⋅E 3 with 20 English sentences, 10 passive and 10 active ones, and generated 4,000 pictures in total. In active sentences, FLUX positioned the agent to the left side of the picture significantly more often than to the right side. In passive sentences, both models positioned the agent to the right significantly more often than to the left. In general, DALL⋅E 3 placed agents to the right more often than FLUX. The models partially copied the tendencies of humans in active sentences conditions, however, in passive sentences conditions, the models had a much stronger tendency to place agents to the right than did humans. Our study demonstrates that these AI models, primarily influenced by English language patterns, may be replicating and even amplifying Western (English-specific) spatial biases, potentially diminishing the diversity of visual representation influenced by other languages and cultures. This has consequences for the visual landscape around us: AI pictorial art is overflowing our visual space and the information that we have imprinted into pictures as intrinsically human is changing.