They claim machine learning, but I'd guess their workhorse is a "grep -f badwords tweets.json | makepdf | send boss@employer" kind of thingy.
They're definitely using some form of sentiment analysis, but IMO these are exactly the kind of results you'd get after setting a programmer, with little to no actual data science background, loose to train his nets on data with a limited or absent intuitive understanding of bias from training data. And what's worse is that in their business false flags probably aren't even considered. It's an extra wide supercharged…
AWS and Azure both have services for this now. Tweet sentiment analysis is like the first getting started tutorial you land on when you hit both of their docs.