Probably just the statistical pattern of engagement. People who like communist content are probably likely to troll/argue with/debunk or otherwise engage with fascist posts by commenting on them, and the algorithm won't be able to tell the sentiment of the comment without costly and mostly useless sentiment analysis. These type of algorithm generally do not "know" anything about the meanings of specific tags, only the statistical relationships that emerge in how users who like them interact with other posts. Which is funny because it means that it can do rage bate simply because it works without any human ever intending it to (but of course, no human stopping it either).