Recognition algorithms often make assumptions about the way people "should" be shaped according to their recorded gender or the way the machine attempts to "guess" said gender, and a clash in expectations gets flagged and generates false negatives because things aren't in the "right" places it thinks they should be.

For a really simple example if this problem, the TSA body scanner machines flag on trans people constantly. It detects as an "unknown mass" if gendered body parts (ie. breasts/genitals) don't match its binary expectations for what gender is marked on their passport. Often they get it coming and going: a trans femme still marked M will get flagged for having breasts, but if they've changed it to F, they get flagged for their genitals instead.

Now you get a full grope patdown from a suspicious TSA officer, because a machine was coded to make binary assumptions about bodies.