Conversation

It's extremely disturbing to read these ML stories after working in the field for a while. So many ways to bias training data. When they compound it by going for medians, you end up with THIS IS HOW A BALTIMORE CRIMINAL LOOKS and it's just a picture of a black man. Obscene. twitter.com/RealSexyCyborg

This Tweet is unavailable.

Cinder

@Triquetrea

Since I work in linguistics, I could even produce similarly ridiculous results: "Preposition 'to' implicated in 80% of error sentences!" (Because it appears in bad machine translated data.)

9:13 PM · Mar 7, 2019Twitter Web Client

Replying to

ML output essentially disregards confounding variables entirely. It's unfiltered data. You can't just use it uncritically. But by a mixture of 'science' 'journalism', lazy scientists and extremely dubious marketing strategies, that is in many cases exactly what happens.