Complex systems have a lot of variables; that doesn't make them poor candidates for modeling. On the contrary, you simply have to rerun the model as you learn more and better data.
The prediction of "a million victims" was made in the earlier stages of the outbreak and grabbed the attention of the world, so we more clearly remember it as it was on that day. But just because we remember what the media said in October of 2014 doesn't mean they weren't continually working on it. After the model started reversing course the headlines stopped being so alarmist, and so the general public barely remembers the much less dramatic follow-on news "Ebola trending downwards", "revised estimates", etc.
Was this a deliberate attempt by the people generating the model to drum up public support? Was this simply the media grabbing on to the worst case scenario because it made the best headlines? Was it an honest mistake in reporting? Perhaps the "garbage in -- headlines out" came from a source other than the model's data.