> it doesn't infer gender from signals such as profession I'm struggling to tell from the article if Smart Compose ever actually did that. It's certainly impli

> it doesn't infer gender from signals such as profession I'm struggling to tell from the article if Smart Compose ever actually did that. It's certainly implied, but the explanation given about gender balance in different fields is offered by the reporter. The parts actually attributed to Googlers only say they're being cautious about gendering in their outputs and couldn't find a fully-acceptable solution. Which leaves me curious they found specific semantic biases or instead noticed general issues like "our corpus taught the NLG to use a masculine default and simple fixes like randomizing pronouns created new issues". (Either way, it seems like a reasonable decision. I'm just wishing for a more technical breakdown out of engineering curiosity.) The other interesting point I saw is that Smart Reply apparently runs on the same system as Smart Complete. Since that offers fully-formed ideas, I imagine gender issues there could be vastly more embarrassing than single-word autocompletions.