Hacker News Reader

Comment by sigotirandolas | Hacker News Reader

I admit I haven't fully read the article, but isn't it obvious from the start that you can't detect "fake news" without looking at the outside world for facts?

Any approach that is trained to detect "fake news" that just relies on feeding it the article's content is either just going to classify based on incidental features such as style (e.g. sensationalism), "hot words", etc., or if it's really good, classify based on whether the article's content is consistent with the "reliable" or the "fake" narrative extracted from the previous training.

In fact if I'm reading this right, they find out that taking a "fake news" article and rewriting it in "reliable news" style fools their detector. Garbage in, garbage out.

View on HN

This is obviously true, a very simple example is the following piece of news: "the most recent vote of the US Senate was 50-48", which not only requires you to look up the actual number, but is also time-dependent.

However, the common usage of "fake news" is not only literally "news that is not true", but rather "news that is deliberately not true in a very specific way".

Many object to the usage of the term itself, and honestly I myself agree that it's not the best way to describe the phenomenon, but it's something that exists regardless of how you choose to call it. I can see the research being useful, I wouldn't say "garbage in, garbage out" really applies here.

I guess you could say that "fake news" is "not true" + "propaganda/flamebait/viral style".

In my opinion this article focuses on the later but not on the former, while using the word for the combination (oh the irony).

Though in that sense I agree that it's not completely useless, as it may be good to strip or standardize the propaganda/virality aspect so one is forced to actually look at the facts.

Well the detection algorithm for that would simply be a function that always returns `true`. There might be a small percentage of false positives, but pretty much negligable.

> However, the common usage of "fake news" is not only literally "news that is not true", but rather "news that is deliberately not true in a very specific way".

The word "fake" surely doesn't have a connotation of an honest mistake, does it? That's not just "common usage", it's the only one I'm aware of.

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.

> is just going to classify based on incidental features such as style (e.g. sensationalism), "hot words", etc.

"Just?" Can you really not read a newspaper article about a topic you don't have a lot of depth in, and tell just based on the writing style (vocabulary, techniques, embedded logic, etc) that it is "likely" to be at least inaccurate?

How well this can be implemented with current machine learning technology is another discussion.

Reply on news.ycombinator.com