Preprint finds sensitive-topic leakage in AI model routing logs

Researchers report in a preprint that logs recording which AI model handles a request can expose sensitive-topic patterns even when content logging is disabled. Their studies used 1.7 million real requests and tested systems that choose models based on request content. An exploratory analysis of 20 RouteLLM decisions distinguished frequent medical askers with an AUC discrimination score of 0.71, below the pre-registered primary endpoint's 0.75 threshold. A domain router reached 0.92, described as an upper estimate. Equalizing routing rates separately for each category at matched prompt lengths removed the gap on real traffic with accurate labels. The defense cost at most 0.2 accuracy points on RouterBench in post hoc testing; pooled equalization failed.