The erasure of intensive livestock farming in text-to-image generative AI CCS: Computing methodologies Image representations CCS: Applied computing Agriculture CCS: Applied computing Sociology CCS: Social and professional topics Censoring filters

Kehan Sheng 1, Frank A.M. Tuyttens 2,3, Marina A.G. von Keyserlingk 1 · 2025

Read The erasure of intensive livestock farming in text-to-image generative AI CCS: Computing methodologies Image representations CCS: Applied computing Agriculture CCS: Applied computing Sociology CCS: Social and professional topics Censoring filters by Kehan Sheng 1, Frank A.M. Tuyttens 2,3, Marina A.G. von Keyserlingk 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-19771.txt.

Introduction

Kehan Sheng and co-authors show that an image generator romanticizes intensive livestock farming, in a paper from 2025. Most farmed animals in rich countries live indoors with little space. Public values, though, favor open pasture, so the gap between image and reality matters.

The team asked the image model inside ChatGPT for farm scenes. It returned dairy cows on pasture and pigs rooting in mud, even when asked for realistic views. The cause was the tool's automatic prompt rewriting, added to fight bias. When that rewriting was switched off, the images changed sharply. Cows stood indoors at metal feed barriers, and pigs stood behind rails on concrete floors. The authors note the irony: a bias guard created a new bias here.

Catalog record

CREATORKehan Sheng 1, Frank A.M. Tuyttens 2,3, Marina A.G. von Keyserlingk 1
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE80 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-19771

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

The erasure of intensive livestock farming in text-to-image generative AI CCS: Computing methodologies Image representations CCS: Applied computing Agriculture CCS: Applied computing Sociology CCS: Social and professional topics Censoring filters Kehan Sheng 1, Frank A.M. Tuyttens 2,3, Marina A.G. von Keyserlingk 1 Abstract . Generative AI (e.g., ChatGPT) is increasingly integrated into people’s daily lives. While it is known that AI perpetuates biases against marginalized human groups, their impact on non-human animals remains understudied. We found that ChatGPT’s text-to-image model (DALL-E 3) introduces a strong bias toward romanticizing livestock farming as dairy cows on pasture and pigs rooting in mud. This bias remained when we requested realistic depictions and was only mitigated when the automatic prompt revision was inhibited. Most farmed animal in industrialized countries are reared indoors with limited space per animal, which fail to resonate with societal values. Inhibiting prompt revision resulted in images that more closely reflected modern farming practices; for example, cows housed indoors accessing feed through metal headlocks, and pigs behind metal railings on concrete floors in indoor facilities. While OpenAI introduced prompt revision to mitigate bias, in the case of farmed animal production systems, it paradoxically introduces a strong bias towards unrealistic farming practices. 1. Introduction Since ChatGPT’s launch in November 2022, generative artificial intelligence (AI) has seen unprecedented growth, with ChatGPT now having over 180 million monthly active users ( Mortensen 2024 ) . Generative AI refers to models that can create new text, images, and other media by learning patterns from existing data, typically guided by text prompts ( Oppenlaender et al. 2023 ) . Evidence suggests that given its ease of use and efficiency, the general public is increasingly relying on ChatGPT over traditional search engines ( Xu et al. 2023 ) . However, AI ethics research has shown that these AI models inherited human biases through the use of internet-scraped training data, thereby embedding stereotypes, dis- and misinformation into their outputs ( Quaye et al. 2024 ) . Given that “a picture is worth a thousand words” , AI-generated images are inherently positioned to shape biases that influence public perception. Visual information can strongly influence the psychological impact of an issue, with AI-generated images proving particularly persuasive in shaping public discourse ( Capraro et al. 2024 ; Haq et al. 2024 ) . AI generated images are far more likely to be shared than text on social media, and are expected to dominate online content in the near future ( Yang et al. 2023 ; Wan et al. 2024 ) . Previous research has revealed prevalent representation biases about gender, skin tone, and geo-culture in human subjects ( Wan et al. 2024 ; Qadri et al. 2023 ) . To mitigate these representation biases and ensure guideline compliance (e.g., remove public figures and branded items in the images), OpenAI employs automatic prompt revision (i.e., rewrite user prompts) in DALL-E 3 to enrich images with greater details, while acknowledging that this process comes with the risk of introducing new biases ( OpenAI 2023 ) . Despite extensive efforts made in mitigating human-related bias in AI, it’s impact on non-human animals, particularly farmed animals, remains largely unexplored ( Hagendorff et al. 2023 ; Singer and Tse 2023 ; Coghlan and Parker 2023 ) . Humans constitute only 0.01% of total biomass and 35.93% of mammalian biomass on earth, while farmed animals comprise 59.88% of mammalian biomass ( Bar-On et al. 2018 ) . To date, global AI regulations and guidelines focus almost exclusively on AI’s impact on humans ( Jobin et al. 2019 ) , with some minor exceptions, including the recent Montréal Declaration that specifically emphasized that AI should consider the well-being of all sentient beings ( de Montréal 2018 ) . The European Union’s ethical guidelines for trustworthy AI in 2019 included the consideration of sentient beings other than humans ( Commission 2019 ) , but then removed this phrase in their updated AI regulation document in 2024 ( Commission 2025 ) . No research has systematically asked the question: how does text-to-image generative AI represent livestock farming, a sector that affects billions of lives of farmed animals and is a key pillar in global food production? This question is highly relevant given that the societal concern regarding the lives led by farmed animals continues to gain traction in recent years ( Weary et al. 2015 ) . 1.1. Cows? Pigs? Why do they matter? Driven by a growing demand for abundant, low-cost food supply, farming practices shifted from extensive systems (e.g., cows grazing on pasture, pigs foraging outdoors) toward intensive systems emphasizing productivity after the Great Depression ( Rollin 2011 ) . Intensive livestock farming is characterized by housing large numbers of animals per unit area ( Tactacan et al. 2009 ) including indoor confinement in cages or in pens with concrete floors, and severely restricting movement ( Tactacan et al. 2009 ; von Keyserlingk et al. 2009 ; Tuyttens et al. 2014 ) . While the increases in intensification are often justified as necessary to feed a projected global human population of 9.8 billion by 2050 ( Godfray et al. 2010 ) , many practices have faced mounting public scrutiny ( Bolton et al. 2024 ) . Extensive scholarship has shown that intensive livestock farming contributes greatly to antimicrobial resistance ( Silbergeld et al. 2008 ; Trevisi et al. 2014 ) , increased spread of zoonotic diseases (pathogen transmissible between animals and humans, such as highly pathogenic avian influenza) ( Magouras et al. 2020 ) , biodiversity loss ( Kok et al. 2020 ) , climate change ( Steinfeld et al. 2006 ) , posing direct or indirect risks to human health ( Hu et al. 2017 ) . Public concern over farmed animal welfare emerged in the mid-to-late 20th century, highlighting that many common livestock farming practices failed to resonate with societal values, such as the permanent separation of dairy calves from the dam within hours of birth ( Sirovica et al. 2022 ) , early slaughter of male chicks and dairy calves ( Bolton et al. 2024 ) , lack of pasture access for dairy cows ( Smid et al. 2020 ) and housing systems that severely restrict animals’ movement (i.e., pig gestation stalls ( Ryan et al. 2015 ) ; tie-stall housing in dairy ( Beaver et al. 2020 ) ). It is increasingly argued that the long term sustainability of food production systems depends not only on economic viability and environmental sustainability but also on social sustainability ( Thompson 1997 ; von Keyserlingk et al. 2013 ) . More recently some have also argued that sustainability frameworks should include a fourth pillar - ‘animals’ - that would require recognition that animals used for food are sentient beings whose welfare matters independently of public perception ( Drury et al. 2023 ) . Given that images shape public opinion, images of farmed animals accessible by the public will play a key role in shaping public perception of the lives led by farmed animals ( Ryan et al. 2015 ; Tuyttens et al. 2011 ) . Most public image datasets commonly depict clean and healthy farmed animals roaming outdoors, but these pastoral scenes drastically deviate from the

Read the complete text