Abstract: A background application running on a child computing device periodically takes screenshots which are samples of what the child is viewing. If the child is playing a video game, text can be extracted from the screenshot within areas of the screen that chat is known to be displayed for the particular game. A cloud computing platform provides services for machine learning (ML) to analyze the chat text to ascertain a likelihood that the chat includes bullying or sexual predation. If it appears likely, the game is disabled and a notification is sent to a parental monitoring application along with a copy of the offensive chat. The parent can override the determination, and in this case, the user application running on the child computing device is re-enabled. The parent's action either to override the determination or to let it stand is then fed back to the ML model so that the ML model continues to learn the semantics of in-game chat based on the parental feedback.