Study finds DALL-E 3 sanitizes factory farming via prompt revision

A study published in Technology in Society has taken intensive livestock farming as its case and, for the first time, produced evidence that OpenAI's text-to-image model DALL-E 3 strips the real conditions of a contested issue from the images it generates. The problem is tied to prompt revision, a process that OpenAI and other companies introduced to reduce bias in text-to-image systems. Prompt revision works silently: it rewrites the prompt a user submits and adds further detail, for example asking that people of different genders and skin tones appear in the picture, in order to promote diversity in the output. The process can also introduce other biases and alter the user's original intent without their knowledge.

In the study, DALL-E 3 showed a clear tendency to romanticize animal husbandry. The images it produced typically showed cows grazing in open pastures and pigs rooting in mud, rather than the conditions of intensive livestock farming. Even when researchers explicitly requested realistic scenes, the model still rewrote the prompt and returned idealized depictions. As of November 10, 2026, the study remains the latest verifiable development in this matter, as reported by Solidot.