On the epistemic costs of implicit bias