Hacker News Reader

Comment by the_arun | Hacker News Reader

If we are building ML models on news, how MLs can filter disinformation? Will we be fact checking every bit of news they receive?

View on HN

MLs would also need tell the difference between disinformation and real information. This is incredibly hard to do, since disinformation looks like real information unless your understanding of the world is broad and specific. And then eventually you end up with the same problem as humans, which is that if the world context leaves out trustworthy information needed to discriminate real information from faulty information, then the ML will also fall prey to the hack. And of course, for truly nefarious misinformation, it is just consistent enough with trustworthy information to appear trustworthy.

This is no easy slice.

Absolutely. There is all sorts of methods that can be employed. Facts can be obfuscated, I suppose a natural language model might be able to detect if this is unusual for that source but it doesn’t mean fake news. Another is simply lying with statistics, this would be much harder for NLP as now your talking getting the model to understand statistically information, understanding the background and why certain ways of presenting stats might be disingenuous for that subject. You are also just relying on a model that only its owner could verify.

Ultimately though, this is fighting a losing battle because the source of the issue isn’t in propaganda which has long existed but in the channels through which it spreads. Something that has been known for a long time is flaming passions is more effective than fact. So too for social media is the spread of information faster when false[https://www.smithsonianmag.com/smart-news/fake-news-spreads-...]. Which makes it difficult to counteract if the false information spreads faster than the rebuttal.

The models would need a live, updated data source. What could they have possibly relied on to accurately label e.g. the Wuhan lab leak theory?

Reply on news.ycombinator.com