It's very difficult and I don't see a practical way to stop the trend. But we could at least push for inference (execusion of train model) to user devices so that real-time usage data is not leaked. Will mentioned local language detection in Damus.