Review: Understanding cattle social behaviour in modern penned production systems with AI technology: Are we tracking welfare indicators?

Monitoring cattle social behaviour is fundamental for assessing animal welfare in modern penned production systems. Traditional observation methods are constrained by subjectivity, labour demands, and limited scalability, prompting increased interest in artificial intelligence (AI) for automated behaviour tracking. While recent advances in computer vision, sensor technologies, and machine learning offer promising tools for continuous and objective monitoring, many systems focus on identifying "what" an animal is doing (e.g., lying, feeding), without interpreting the underlying "why", such as whether a posture indicates rest, discomfort, or illness, due to lack of contextual modelling. This review synthesises findings from over 180 peer-reviewed articles sourced from Scopus, Web of Science, and PubMed databases using targeted keywords related to cattle behaviour, welfare indicators, and AI-based monitoring. We examine the biological foundations of cattle social behaviour, the effects of modern penned production environments on behavioural expression, and how current AI-based technologies align with established welfare assessment protocols. Our analysis reveals that while current AI systems effectively capture indicators like activity level, walking, standing, feeding, and lying, they often fail to account for complex affiliative behaviours, social dynamics, and context-dependent stress signals. Major limitations include poor generalisability across farm contexts, insufficient temporal and multimodal data integration, and a lack of transparency in system outputs. To address this gap, we propose a welfare-centred AI framework grounded in five principles: multimodal data integration, context-aware behavioural modelling, shared behavioural ontologies, human-in-the-loop system design, and explainable AI. This approach supports a more accurate interpretation of cattle behaviour, facilitating early detection of welfare risks and informed decision-making. We conclude by outlining future research needs in system validation, ethical co-design, and cross-disciplinary collaboration to enable responsible scaling of AI technologies in livestock systems.