It's not buggy: it detects patterns like the majority of posts being retweets instead of new posts; any new post consisting of an image and accompanying text instead of a message or thread; replies being short comments on top responses to popular posts; and a high degree of connectivity to other accounts that are likely bots.

False positives are generally curated accounts, corporate accounts, and government accounts (which are really bots with their software in managerial documents); porn accounts; and people who reblog the news but don't really "use" twitter.

And if it's anything like some other naïve bot detection algorithms I've seen, it might be viciously biased against non-English-speakers. To the extent of, say, labelling anyone primarily using Arabic or Cyrillic alphabets as a bot.

Or that writes a lot in Dutch.

Just the other day a prominent dutch account got auto-blocked on twitter for saying 'Die Rutte toch.', where 'Die' (a pretty common dutch word, which in this case translates as 'That') was mistaken for a threat to our prime minister.

If it detects my account as a bot I would say it's buggy, because I've done none of that stuff.

Or you just post a lot of misinformation.

When I was in the spam blocking business we joked about the problem of the Russian urologist. Somewhere, there must be someone who really is trying to buy Viagra from a Russian e-commerce site. But the other 99.99999% of the time it's spam. Perhaps this person's Twitter suffers a similar fate: all the classification features align with robot traffic, but they aren't a robot.

That would nullify the study entirely since the criteria in the test directly overlaps the criteria of the results.

There are multiple people in this thread that have been incorrectly tagged as bots, I’d like to see proper tests done with the software identifying the level of accuracy. It shouldn’t be too much to ask since their entire study is hinged on the accuracy of this software.

Right, yes, that must be it.