In a way, it's (probably unintentionally) training its human to use a narrower range of voice patterns (which is handy when collecting data to train a network with), similar to how we're more likely to write in sans-serif than cursive now that computers are ubiquitous.