More troublingly, those who identified as agender, genderqueer or non-binaryy*, were misclassified 100 percent of the time because these gender identities have not been built into the algorithms.
The training data for the AIs presumably doesn't have these nebulous categories marked. And if they did - and they were present in statistically significant numbers - then we'd probably just find lots of women and some men being miscategorised.
The piece doesn't say what the hit rate for transwomen was - I'd guess fairly high because an AI may well be able to detect sex Male + external 'feminine' markers such as makeup, and categorise them as transwomen.
I'd be curious to know if the hits and misses for transmen were broadly beard or no beard...