they might have a "binge-ness" score for every show. If they know people usually don't binge more than three episodes of Scandal then they can safely assume tha

they might have a "binge-ness" score for every show. If they know people usually don't binge more than three episodes of Scandal then they can safely assume that the TV has been left on when the fourth episode kicks in. They may even construct segmented binge patterns where they know episodes 3 to 6 of House Of Cards usually gets binged where as episodes 1 and 2 usually do not. Building and training these segments would appear random to the end user but after awhile should be good at guessing when the TV has been left on or the user is binge watching.