I’d phrase it that “bias” in the social sense can occur in (1) the task itself based on how it’s defined, (2) preexisting within the problem domain, so representative datasets could still be biased, and (3) in when the system and the people operating it choose to use the outputs of the model. Which is all to say that “bias” problems related to the model can extend beyond the model itself.

The solution isn’t necessarily to fix the data or the model. It might be banning use of facial recognition by police, for example, which some places are doing.