I suspect you're right. Part of the supervised "learning" is hard-coding answers to gotchas posted on Twitter.