In my opinion fake news detection is an ill-defined problem, since the definition of "fake news" itself is not clear.

You might be able to spot blatant inaccuracies like "Mount Everest is located in France"; but then do you really need an NLP tool to tell you that? Verifying anything slightly more complex than that is an AI-complete task, if even humans have a hard time solving it.

Besides, you can manipulate people without writing factually incorrect statements, just by lying by omission and using careful word selection. And this is basically impossible to detect. Other times you have news that just can't be verified, even by humans, or whose verification depends on what source you consider "reliable" (which then introduces a whole lot of bias).

I've worked on the problem personally, and I think the only workable solution would be a tool that would in some way highlights possibly problematic parts of an article they are reading, maybe enrich it with possible related references where the user can read more about the topic (these would be suggested automatically).

Unfortunately, while this kind of solution could be helpful for a "trained" reader, most people don't want to put any effort into consuming news carefully; It would end up helping only those who already have a habit of verifying, which would then be useless to solve the "fake news problem".

You are absolutely right. But when someone from New Zealand is tweeting as an American or someone from Manchester tweeting as a Russian journalist, a subset of fake news can be flagged. (eta: I guess what I am trying to say is that we first have to identify the "unreliable narrator" archetype.) For example, we have AP stylebook in the United states that is related to associated press. Every publication has a recommended style book. Now...can it be mimicked? Yes, of course. But we have at least have one level of authentication. It takes years to master stylebook rules for a human. A fake news generator might be able to pass the Turing test, but I don't think its been done yet.

The downvotes were from the phrasing "Isn’t fake news detection a solved problem with NLP?" is a leading construction. It suggests that everyone already knows that it is solved, whereas that's not true at all.