> One study showed that software trained on Google News became sexist as a result of the data it was learning from. When asked to complete the statement, "Man is to computer programmer as woman is to X", the software replied 'homemaker".
> Dr Joanna Bryson, from the University of Bath's department of computer science said that the issue of sexist AI could be down to the fact that a lot of machines are programmed by "white, single guys from California" and can be addressed, at least partially, by diversifying the workforce.
Just curious but wouldn't this more likely be a result of a bias in the training data rather than the diversity of the programmers?
> When asked to complete the statement, "Man is to computer programmer as woman is to X", the software replied 'homemaker".
Seems they're putting a lot of weight on an unqualified question...the AI just fed back the easiest, most popular, stereotypical answer it came up with. A tee-ball response.
It didn't stop and think about who'd be offended, first.
If it was trained on Google News it would just make more sense to me that it is a reflection of our journalists and possibly whatever algorithm Google uses to identify popularity, presumably a reflection of our society.
Yes, but the intention is to demonstrate one of the issues with using ML. For example, translating into a language which uses genders in grammar. A human translator will read a passage to determine a doctor's gender. A machine translation will know that doctor is usually translated as a male doctor and apply that to the translation. If the bias in the training is strong enough it can override actual signs in the source content that the doctor is a woman.
If a person reads enough translated content, this just passes on the bias from the training data.
> Dr Joanna Bryson, from the University of Bath's department of computer science said that the issue of sexist AI could be down to the fact that a lot of machines are programmed by "white, single guys from California" and can be addressed, at least partially, by diversifying the workforce.
Just curious but wouldn't this more likely be a result of a bias in the training data rather than the diversity of the programmers?