Princeton study finds data leakage in hundreds of AI papers

Kapoor and Narayanan reviewed over 300 papers across 17 fields, including medicine and economics, finding data leakage. This means models were trained on information unavailable in real predictions, inflating reported performance.