NLP Researchers Call for Longitudinal Studies of Human-AI Interactions

A new paper on arXiv (2608.02491) argues that language models, due to their human-like nature and rapid integration into daily life, pose longitudinal risks including cognitive, developmental, and socio-affective changes that may not appear in short-term interactions. The authors propose a shift from static evaluations to long-term measurements of behavioral changes, drawing on social science methodologies to understand diachronic human-model interactions. They advocate combining computational NLP methods with these measurements to address long-term safety risks and guide model development. The paper is a preprint and has not yet been peer-reviewed.