The Challenge of Detecting AI-Generated Text: False Positives and How to Address Them