A new preprint on arXiv, titled "From One Query to Many: A Critical Survey of Privacy Accounting in RAG," takes stock of how privacy is measured in retrieval-augmented generation. The authors note that RAG lets large language models pull in external knowledge, but this creates risks when the underlying corpus holds sensitive information.
The paper positions differential privacy as a formal tool for addressing those risks. However, the survey's critical angle suggests that current accounting methods may not fully capture privacy loss when a system answers many queries over time. The abstract highlights the need to scrutinize how guarantees hold up beyond a single interaction.
Because the source is a single preprint, the article reflects only its stated scope and claims. The full methodology and specific findings are not detailed in the abstract, so the survey's conclusions remain to be examined in the complete paper.