The ultimate proof is actual data that's steadily being collected as time moves forward. That surely culls off bad models as they're proven wrong. But I think it would only be helpful at reducing bias if we just look at the surviving old models rather than making new or updated ones that incorporate the new data or else the new models might still have a systemic bias in one direction or other.