The article notes this, and like any spam detection method, it has a degree of false positives, but it seems very low (less than a percent according to the article). I'm sure an official implementation of this could take more internal, non-public factors into account, like IP addresses and clustering of account creation times, to make it even more accurate and drastically reduce the amount of spam users.
>>GlumWo+(OP)
The claim I saw in the article is 98% precision. Which doesn't actually tell us the predictive value without the base rate which seems to be all over the place.