When Microsoft debuted its AI chatbot "Tay" last year, she greeted Twitter users excitedly, gushing that she was "stoked" to be on the social network and that "humans are super cool". Within 24 hours Tay, which was designed to emulate a teenage girl, was telling followers to "f*** her", calling them "Daddy" and declaring "I f***ing hate feminists". Microsoft subsequently abandoned the project and deleted her from the internet. Of course, Tay's offensive outbursts were partly due to internet users' determination to interfere with a corporate PR stunt. But they also highlighted a major problem faced by the AI industry: if robots learn from humans, there's a good chance they'll also adopt the biases – gender, racial and socio-economic – that exist in society.

Robots are picking up sexist and racist biases based on information used to program them predominantly coming from one homogenous group of people, suggests a new study from Princeton University and the U.K.'s University of Bath. Lead study author, Aylin Caliskan says the findings surprised her. "There's this common understanding that machines are supposed to be objective. But robots based on artificial intelligence (AI) and machine learning learn from historic human data and this data usually contain biases," Caliskan tells The Current's Anna Maria Tremonti. Machine learning takes statistics and information that has been inputted and Caliskan argues it's only until humans become completely unbiased that the possibility of an unprejudiced robot can exist.