I still don't get what solution they are proposing to the problem. I have also always assumed that any bias stems from data since the system uses it to learn. What's the alternative?
They aren't proposing one. Most people would rather complain than fix.
I guess it's possible to suggest that, given the lack of diversity in training sets, different training techniques should be adopted to account for it. Is that actually a viable approach that people are suggesting, and LeCun is ignoring/downplaying?
Do they have any practical suggestions on where to find a huge number of suitable qualified people? The people in prestigious academic positions often come from not just non-poverty backgrounds but wealthy backgrounds.
If I had to summarize it in a sentence, the solution proposed is "consider the ethical implications of misapplication of your work, and attempt up front to mitigate the damage that would result." It's more nuanced than that, but it's the reason some people don't accept Yann LeCun's "just debias the inputs" as sufficient.