@james @inthehands I have seen some efforts to identify, quantify, and mitigate bias in the human-generated labels, if that's what you're getting at. I would say, yes, there will *always* be bias in manually tagged data. The question is, do the biases present in that data affect the job you want the model to do? Often the only source of truth for whether a task has been performed correctly is human judgment. In those cases, we can identify secondary biases (like gender or race in hiring decisions) that we want to specifically mitigate, but what we are training the model to learn is literally a bias itself, e.g. the bias towards candidates that hiring managers think will do well in the position.