Google says that SafetyCore “provides on-device infrastructure for securely and privately performing classification to help users detect unwanted content. Users control SafetyCore, and SafetyCore only classifies specific content when an app requests it through an optionally enabled feature.”
GrapheneOS — an Android security developer — provides some comfort, that SafetyCore “doesn’t provide client-side scanning used to report things to Google or anyone else. It provides on-device machine learning models usable by applications to classify content as being spam, scams, malware, etc. This allows apps to check content locally without sharing it with a service and mark it with warnings for users.”
But GrapheneOS also points out that “it’s unfortunate that it’s not open source and released as part of the Android Open Source Project and the models also aren’t open let alone open source… We’d have no problem with having local neural network features for users, but they’d have to be open source.” Which gets to transparency again.