Silencing controversial realities through prompt revision: The erasure of intensive livestock farming in text-to-image generative AI

Generative AI like ChatGPT is increasingly integrated into people’s daily lives as a vehicle for information access and to aid decision-making. However, misinformation and biases are embedded in these models, raising ethical concerns. Previous research has revealed AI biases regarding human attributes like skin tone, gender, and geo-culture. To mitigate these biases in text-to-image generative AI models, companies like OpenAI introduced prompt revision, an implicit process of rewriting user prompts to add more details such as showing humans with different genders and skin tones; thereby, promoting diversity. However, rewriting user prompts risks introducing other biases and changing user intention without the user knowing. Using intensive livestock farming as an example, we provide the first evidence that OpenAI’s text-to-image model (DALL-E 3) erased the reality of controversial topics through prompt revision. DALL-E 3 introduces a strong bias toward romanticizing livestock farming, depicting dairy cows grazing on pasture and pigs rooting in mud. This bias remained when we requested realistic depictions. The generated images are in stark contrast to reality where many farmed animals in industrialized countries are reared indoors with limited space per animal, conditions that fail to resonate with societal values. Inhibiting prompt revision resulted in images that more closely reflected modern practices of housing farmed animals indoors; for example, cows were shown feeding through metal headlocks, and pigs were shown behind metal railings on concrete floors. Prompts requesting images of farms in North America, Europe and Oceania yielded mostly pastoral scenes regardless of region, but once prompt-revision was disabled, regional variations emerged. The proportion of outdoor versus indoor housing systems depicted in AI-generated images after prompt-revision was disabled partially reflects actual housing statistics from these regions. Romanticizing and concealing the reality of livestock farming through prompt revision risks reinforcing mis- and disinformation about how farm animals are raised. This disconnect could hinder meaningful dialogue in aligning farming practices with public values, harming the social sustainability of both AI development and livestock farming. Moreover, our study reveals how prompt revision could systematically promote misinformation while silencing the reality. We caution against the broad sweeping romanticization of AI-generated imagery.