Even if we narrowed the scope of training data exclusively to professionals, we would have issues with, for example, racial bias. Doctors underprescribe pain medications to black people because of prevalent myths that they are more tolerant to pain. If you feed that kind of data into an AI, it will absorb the unconscious racism of the doctors.

And that's in a best case scenario that's technically impossible. To get AI to even produce readable text, we have to feed a ton of data that cannot be screened by the people pumping it in. (AI "art" has a similar problem: When people say they trained AI on only their images, you can bet they just slapped a layer of extra data on top of something that other people already created.) So yeah, we do get extra biases regardless.