The winning bounty for Twitter's AI based Image-crop tool reveals it favoured certain body types

Earlier this month, Twitter had announced a bounty of up to $3,500 (approx. Rs. 2,60,000) to people who would be able to spot biases in their saliency algorithm for the image crop tool that they used.

This was after an unintentional bias towards certain body types and skin colors were detected within the tool. Users discovered that the AI would tend to crop out people of colour in image cropped previews and it had strong preference for keeping women in picture as opposed to men.

Bogdan Kulynych has now won that bounty. He generated a series of fake faces, tweaking their appearance with each pass to see what the algorithm preferred. He found the tool veered towards more conventionally attractive features. The test revealed that the model was biased towards people that were slim, young or had lighter or warmer skin tone. It also had a strong preference for stereotypically feminine traits.

Halt AI, a Canadian startup, who came second and took home funds of $1,000 (approx. Rs 74,260) found that the tool was more likely to crop out people in wheelchairs. They also found that people with grey or white hair were more likely to be left out.

Roya Pakzad came third and she translated texts from English to Arabic, feeding them to the AI tool. The tool showed a strong preference for English over Arabic.

Speaking with The Register, volunteer researcher at AI village gave Twitter props for sponsoring the challenge, saying that, "Twitter has the same problem now that anyone who runs a bug bounty contest runs: how to fix what the participants found.”