Machine learning isn’t effective at identifying fake news | Hacker News Reader

Machine learning isn’t effective at identifying fake news

fastcompany.com

8 pointsembit5 comments

IMHO, I disagree. The problem is taking an entire article not filtering the emotions out of it, tallying emotional hits and drives a score. If the article has a high degree of emotional identifiers its most likely an opinion piece not an entirely factual news article. Then articles from the same agency that has low repeatedly scores (or writers) would be flagged.

Use media bias checker, pick the far left and far right news/blog agencies, train them on picking out emotions arguments first, then statements of facts.

We train AI with xrays at work. Humans identify the issues on scans, then the software learns from humans identification on the xrays. After training, it can start to recognize what humans are identifying. Text has meaning, so you have to defuzz the text into something simpler.

Fake news is fake because the facts aren't real. machine learning isn't going to learn how to identify it when sentient people have difficulty.

Hm. I think the kind of fake news most people are worried about on social media has properties which distinguish it from merely "factually incorrect information".

It's a kind of "supernormal stimulus" - high virality, engineered to trigger tribalism and strong emotions in susceptible recipients.

The properties which make it a "weapon" are the ones which you'd want to mitigate. Unfortunately those properties also happen to goose the engagement metrics of social media companies.

Source reputation, rate and pattern of viral spread should be taken into account, as well as content itself and user flags. Also user engagement behaviour - e.g. average user shared this after reading for less than 10 seconds.

Basically, highly partisan content spread quickly by unthinking users, from a low reputation source.

Unfortunately the networks themselves have the best metadata to work this stuff out... and the least genuine desire to do so.

I'm in the "solving the worst of this looks fairly tractable" camp, but I agree that bare source text is going to be hard to work with.

I did state you would have to use reputation scores too. If I see its the onion or babylon bee, I can probably assume its satire.

Isn't that just saying machine learning is incapable of producing novel discoveries that would have been difficult for humans? Because it has already succeeded in that area.

The problem with fake news is that it takes advantage of a lack of knowledge surrounding the presented facts. Which is to say, if you aren't feeding that background knowledge to the ML engine from the start, it's fooled just like we are.

Reply on news.ycombinator.com