Languages exhibit statistical regularities concerning the frequencies and co-occurrences of words. Language users learn from such patterns without being consciously aware of them. We investigated statistical properties of the language used on television news in discussing politicians. We compiled corpora consisting of language used on four networks (MSNBC, ABCNews, CNN, FOXNews) from 2007-2012. We analyzed the frequencies with which 500 affectively-valenced words co-occurred with politicians' names (Obama, McCain, Romney) during the run-ups to the 2008 and 2012 elections. We used these co-occurrences to derive a summary measure, their net positivity score. Positivity scores for candidates changed over time in ways that reflect real-world events. Positivity towards candidates differed across networks. Net positivity toward President Obama during his first term was strongly correlated with approval ratings. The results show that statistical aspects of language, of which people are not consciously aware, convey varying attitudes on network news.