3.2 million replies, 88 candidates, and the gender gap that prevalence metrics erase

Perspective API returns one number per message, from zero to one, and what that number estimates is the chance a reader would find the comment rude enough to leave the discussion ^[3]. That is a serviceable instrument for asking whether a thread has gone bad. It is a weak one for the question the Joint Research Centre team ended up asking, because a reply bundling a sexist insult, a sexual threat, a reference to mental illness and a demand that the candidate quit politics sits on the same single axis as a reply calling a man incompetent ^[15][9].

The count that survived contact with human eyes is the interesting one. Of the roughly 2,700 highest-confidence toxic replies the team read by hand, 71% went to women ^[14][6]: about 1,900 replies against about 780 for the men, a little over two to one ^[19][20]. Part of that follows from attention, since the women in the sample posted less and were replied to more ^[5], and the adjustment that makes the gap vanish is a division, toxic replies over total replies ^[4]. Volume can explain a count. It cannot explain a vocabulary. Receiving more replies does not convert generic hostility into sexually explicit remarks and instructions to vacate political space ^[7][8], which is the material the authors connect to the literature on semiotic violence, harm done through language that works to delegitimise women as political actors ^[10].

The intensity result is narrower than it first reads. The team reports that 62% of toxic replies to women carried two or more abusive elements, and that 52% of those to men carried only one ^[16]. Read as complements, men sit at 48% with two or more, a gap of 14 percentage points ^[17], real but well short of what the 71-to-29 targeting split implies. The summary does not publish the men's two-or-more figure directly, which is the kind of asymmetry that makes a replication attempt guess.

Where this bites is reporting. Prevalence, the share of replies crossing a threshold, is the metric that scales to millions of messages, and it is the metric that came out flat here. Composition, what the abuse actually says, is what separated the sexes ^[7][9], and composition required hand-reading roughly 0.08% of the corpus ^[18]. A transparency regime built on a rate and a threshold will collect numbers that are accurate and almost empty of this finding. Concentration makes it worse: the JRC reports that women from left-leaning groups drew the higher volumes, matching what studies of Brazilian elections found ^[11], so an average across a national ballot spreads the load across candidates who were not carrying it.