It was only 60%-70% accurate, but that is much better than 50/50 random guessing. Everything like this (there's other attempts at the same thing, with similar

It was only 60%-70% accurate, but that is much better than 50/50 random guessing. Everything like this (there's other attempts at the same thing, with similar results) is more than useless for most practical purposes. More so, it's a great example of machine learning used all too wrong. You can blindly apply machine learning to almost any corpus and "predict" almost anything with 60% accuracy without gaining any real understanding of the corpus or the correlations you're trying to tease out. Often the only reason you can do better than guessing at all is because your corpus isn't perfectly unbiased itself. And with 60% accuracy, good luck using the results for anything! This is doubly the case here, where they're actually just using word frequency, not a real analysis of sentence structure. It's the simplest, dumbest form of natural-language-based machine learning, and it shows. I've used the tool derived from their research and it's beyond useless: it classifies every single one of my posts wildly differently. Because it's solely based on a few dozen words' frequencies, it's more affected by the topic than anything else. Topic might actually account for a large portion of the bias in their corpus. But if we start claiming topic to be a sort of "gender identification" system, all we've done is enforce stereotypes.