NLP Researchers Call for Pivot From Static Text Evaluation to Longitudinal Measurement of Human Behavioral Change

arXiv 2608.02491·low signalXBlueskyLinkedInCopy link

This position paper argues that language models' 'human-ness' and daily integration introduce longitudinal risks — cognitive, developmental, and socio-affective changes — that do not surface in short-term evaluations of generated text. It proposes importing measurement instruments from social science fields built for emergent phenomena in longitudinal data and combining them with computational NLP methods. The practical claim is that modeling human behavioral shifts as a function of model interaction would enable online rather than post-hoc detection of problematic behavior and should be built into alignment frameworks.