As someone who developed diagnostic support for several skin conditions, black skin is just plainly more difficult to classify in most cases. The reason for that is indeed partly due to lacking training data, but also to a large part because black skin reflects less light.
If you are in northern Europe, good luck finding a black woman who has psoriasis, which is much less common for black people. White people have magnitudes of higher risks here. Vitiligo on the other hand is trivial to detect on black skin, not always that easy on white skin.
Vitiligo is completely harmless and not a health threat. But since Ebola can have similar looking symptoms, there is stigmatization of people suffering from it in regions where Ebola is still prevalent. These are actual problems that need fixing.
The real ethical minefield in my opinion is reducing individual cases to numbers within any AI model. And elevating current AI technology, which I would classify as very dumb, is part of that problem.
No, the danger of AI probably lies in small specialized agents trying to classify people within their restricted model of reality. That is already reality in banks, were every customer has some form of performance score to determine credit worthiness. A score every bank employee has adhere to and is not allowed to use own judgement.
As I said, the software I developed was diagnostic support, not diagnostic. The medical sector seems to have a more mature approach to AI than finance. So if the software fails, there is still a doctor.
If you are in northern Europe, good luck finding a black woman who has psoriasis, which is much less common for black people. White people have magnitudes of higher risks here. Vitiligo on the other hand is trivial to detect on black skin, not always that easy on white skin.
Vitiligo is completely harmless and not a health threat. But since Ebola can have similar looking symptoms, there is stigmatization of people suffering from it in regions where Ebola is still prevalent. These are actual problems that need fixing.
The real ethical minefield in my opinion is reducing individual cases to numbers within any AI model. And elevating current AI technology, which I would classify as very dumb, is part of that problem.
No, the danger of AI probably lies in small specialized agents trying to classify people within their restricted model of reality. That is already reality in banks, were every customer has some form of performance score to determine credit worthiness. A score every bank employee has adhere to and is not allowed to use own judgement.
As I said, the software I developed was diagnostic support, not diagnostic. The medical sector seems to have a more mature approach to AI than finance. So if the software fails, there is still a doctor.