If I'm understanding the article, then an aspect of variance is how much a sensor deviates from the wisdom of the crowds. So the 6.8 is harmless if it happens to be close to the right answer, and downweighted if not.

Which, if correct, means that Kalman filters are useful to detect malicious input to crowdsourced data like surveys, reviews, etc.

The kalman filter depends on having an accurate estimate of the variance to work, if you incorrectly weight an input as being far more trustworthy than it actually is then it won't work very well.

If your only sensor is broken and tells you nothing, then your system is unobservable with such a sensor and the only thing you can do is propagate your initial uncertainty through the dynamics. The Kalman filter will do exactly that in this case.

not if you incorrectly estimate that the broken sensor has zero variance. In that case the kalman filter will discard all the other information and trust it completely. I think that's the point the poster is making (the article uses a somewhat naive estimate of variance)

Yeah, actually working out the variance of your input is often the tricky bit (in a lot of cases it's just estimated once and then hardcoded, but more sophisticated approaches to try to estimate it dynamically are possible)