I don't understand what you mean. Don't you train the classifier on some input set? There's your user's confirmation bias.
I think what he means is that if you take the entire article as a bag-of-words model, the general topic of the article will be a stronger signal than the position the article takes with respect to the subject.
The latter is a lot harder to extract from a bag-of-words. In fact the only thing I can imagine is if an article uses a lot of euphemisms or negative synonyms of a topic.
So you get the factual reporting articles regardless of left/right bias. With the more polemic hyperbole using articles, it will filter a bit more of the angle you disagree with, which adds some bias, but it also good for your blood pressure and a polemic article you disagree with is not going to make your view more balanced either.
One of the nicest aspects of it is that it doesn't support a user's confirmation bias: your perspective isn't taken into account in filtering since it's just looking at keywords. That's probably not as important here, but especially on political news it's highly relevant. If I'm a Democrat, I don't want just left-leaning news to come through the grapevine, because that prevents me from seeing the other perspective.