Typically fake news is written in plain language just as any other set of statements would be.



Typically fake news is written in plain language just as any other set of statements would be. The fact that AI can’t distinguish the two based on the text alone means that the qualities that make it “fake” are not encoded in the writing style. In my mind that could easily point to the source. It could point to intentional fake news. In fact that might be the line that AI cannot cross. How could it? It can’t determine the truth or falsity of a statement in plain declarative form. Because it has no way of knowing if the statement is true based on the the writing style alone.

Take two statements. The first one is

Ralph Nader is the leader of the Green Party.

While the second one is

Chuck Norris is the leader if the Green Party.

What’s the difference? Nothing really. No difference in writing style. We just changed a name. To know whether this statement is true or false, the network would have to be able to know something about Ralph Nader and Chuck Norris and how they relate to the Green Party. Which typically isn’t the case for networks trained to detect it.

When it comes to discerning real versus fake news in which the real news is written by someone who actually believes it and the fake news is written by someone who is deluded into actually believing it, these networks would do very well. Because the writing styles of undeluded and deluded are distinguishable. Their minds are different. One is possibly mentally ill while the other is not. It makes for different patterns and semantics in their written language.

But now imagine the real and fake news are both being written by someone who knows the true news is true and someone else who knows the fake news is fake. Then you’re comparing the semantics produced by like minds.

Source: Deep Learning on Medium