One of the unfortunate quirks of current ML algorithms is that you can sometimes accidentally train your model to recognize something correlated with the target

One of the unfortunate quirks of current ML algorithms is that you can sometimes accidentally train your model to recognize something correlated with the target label that is not actually the target label based on other tells (think of the tank image recognition story that was circulating again a few days ago). In this case, some style of writing or a particular set of authors, publications, timestamps, etc could be features useful in determining whether something will be fake news with high accuracy. If so you don't necessarily need to always analyze the factual contents to detect fake news with high accuracy.