> LeCun mentioned that there would have been opposite problem if it were trained on a dataset from Senegal. But why wasn't it trained on a dataset from Senegal? Why do we always see these errors where white-centric datasets produce white-centric results?

Probably because it was a research AI not a production AI. Having a very diverse dataset at that stage doesn’t help with your research, so it is fine to use whatever is easily available.

At that stage, you are trying to show that your approach can work in some cases. Once you've got that, it is time to expand the research with a wider range of inputs to find out what the limits of your approach are.

For example, if I were trying to make a US English speech to text transcription system, I might start with recordings of assorted NPR programs, because NRP often makes the recording available online along with they transcripts.

That would be great for determining if my basic approach has promise. Once I have determined that, so know that the whole endeavor is not just a waste of time, I could go looking for data that includes speech that has characteristics that would be missing from the NPR data, such as heavy regional accents.