AI prompt confidentiality and false citations worry researchers

This study of 15 academic researchers found that users routinely input unpublished research questions and proprietary domain knowledge into commercial generative AI tools, creating confidentiality risks due to opaque data retention and training practices; participants also faced significant output verification and provenance gaps (hallucinations, attribution displacement, synthetic blending) and adopted manual verification and limited-use workarounds. The authors recommend slower AI adoption, better verification pipelines, metadata exposure, and clearer vendor data governance to mitigate risks to institutions and early-career researchers.