3 ms·

That’s not the meaningful comparison, because Twitter has no control over what people do on their own, whereas their algorithm was introducing bias into photos

That’s not the meaningful comparison, because Twitter has no control over what people do on their own, whereas their algorithm was introducing bias into photos with no recourse for the people posting the images. I would argue that a 7% preference for White women over Black women is pretty darned significant. But beyond that, cropping a photo in a way that disrespects the intent of the poster is not the right approach. Fortunately, Twitter realized that it goes beyond the numbers.

ALittleLight 5y ago
How is that not a meaningful comparison? If your intent is to reduce bias in image cropping this research doesn't even suggest that changing to manual cropping would do that. What if humans are even more biased than the algorithm? It seems like Twitter, and you, are saying they should try to minimize responsibility for bias in image cropping. Right now Twitter is responsible for it because their algorithm does it. If they foist the responsibility to users, then users will be responsible, but is there more or less bias in image cropping? No idea. I think, if bias in image cropping is a problem, then they should check to see whether their change increases or reduces the problem. If bias in image cropping isn't a problem, then who cares. If their problem is that they don't want to be responsible for bias in image cropping then of course making the users responsible is the right way to go.
- URSpider94 5y ago
  It's not meaningful because the question here is not whether the algorithm reduces overall bias. The question is whether the algorithm introduces unwanted bias into the presentation of people's posts, with no ability for them to opt out. The bar would be different, in my mind, if we let people choose whether to use this algorithm or opt out. Also, as Twitter finally came to at the end of their post, the much bigger issue is that this algorithm doesn't respect people's desire to curate their own public image. Taking out the bias aspect, if I post a group photo that includes my best friend, and the algorithm crops them out of the thumbnail entirely, then that's going to be really frustrating to me. If I want to center a person of color or a woman (or an LGBTQ individual, which we don't know how the algorithm would rate - we don't have the data), and the algorithm makes a different choice, then my message is damaged.

ALittleLight 5y ago

How is that not a meaningful comparison? If your intent is to reduce bias in image cropping this research doesn't even suggest that changing to manual cropping would do that. What if humans are even more biased than the algorithm? It seems like Twitter, and you, are saying they should try to minimize responsibility for bias in image cropping. Right now Twitter is responsible for it because their algorithm does it. If they foist the responsibility to users, then users will be responsible, but is there more or less bias in image cropping? No idea. I think, if bias in image cropping is a problem, then they should check to see whether their change increases or reduces the problem. If bias in image cropping isn't a problem, then who cares. If their problem is that they don't want to be responsible for bias in image cropping then of course making the users responsible is the right way to go.

URSpider94 5y ago
It's not meaningful because the question here is not whether the algorithm reduces overall bias. The question is whether the algorithm introduces unwanted bias into the presentation of people's posts, with no ability for them to opt out. The bar would be different, in my mind, if we let people choose whether to use this algorithm or opt out. Also, as Twitter finally came to at the end of their post, the much bigger issue is that this algorithm doesn't respect people's desire to curate their own public image. Taking out the bias aspect, if I post a group photo that includes my best friend, and the algorithm crops them out of the thumbnail entirely, then that's going to be really frustrating to me. If I want to center a person of color or a woman (or an LGBTQ individual, which we don't know how the algorithm would rate - we don't have the data), and the algorithm makes a different choice, then my message is damaged.