You may laugh, but it seems scarily plausible to develop a machine algorithm that generates 12 articles, have a human go through to pick the 3 least-ridiculous ones and do some light editing, and push it out to the news.
The underlying data are available, the vocabulary is distinctly limited, and the reporting was already highly pro-forma. Mad-libbing a few variants of language and selecting the best of the lot does work fairly well.
That actually raises the underlying question: why publish narrative copy at all in cases such as this, rather than data tables or charts? News and journalism are curiously allergic to data or presenting it in a usable fashion -- I've seen tables essentially written out over several paragraphs that could have been expressed in a few grid squares.