What are we teaching the machine? : toward responsible AI and rigorous data science for farm animal welfare
Technology is increasingly integrated into dairy cattle management to monitor health and productivity. The rise of artificial intelligence (AI) has led to a plethora of claims that AI could improve animal welfare, yet much of this discussion lacks critical reflection. The overall aim of my thesis was to critically evaluate how data are collected, analysed and interpreted to understand animal welfare and how are farm animals portrayed in general-purpose AI. I addressed four overarching questions: 1) Are we interpreting existing data streams in biologically meaningful ways? 2) Are the methods used to analyse these data reproducible and reliable? 3) Do we truly understand what “ground truth” means when training machines to detect health problems? 4) As generative AI becomes mainstream, how is it shaping public perception of livestock farming? In Chapter 2, I challenged the common practice of using agonistic interactions recorded right after fresh feed delivery to calculate dominance hierarchies in indoor-housed dairy cows, showing that these interactions likely reflect motivation to access fresh feed rather than dominance per se. In Chapter 3, I developed moo4feed, an open-source R package for reliably extracting biologically meaningful behavioural variables, including agonistic interactions, non-nutritive visits, and feeding strategies, from electronic feeder and drinker data, revealing feeding strategies among different individuals. In Chapters 4 and 5, I challenged the reliability of traditional lameness scoring and proposed a new approach: asking untrained observers to judge which cow is ‘more lame’ while watching two cows walk side by side. Even untrained observers could do this reliably, and these pairwise comparisons were used to rank cows from healthy (not lame) to most lame. In Chapter 6, I found that, despite being aware of the realities of modern livestock farming, generative AI uses an internal mechanism called ‘prompt revision’ to silence that reality, reflecting the pastoral ideal that dairy cows graze on pasture and pigs root happily in mud. Together, these findings demonstrate the importance of critically examining how animal data are collected, analysed, and interpreted, and reveal how generative AI encodes implicit assumptions about how animals are raised and understood.