It would lose knowledge about existing subjects unless it’s continually retrained on those too. It could help inform the next training dataset though.