What are the privacy concerns with using AI this way, and what are best practices to mitigate the risks?