To be representative, the models need a stronger bias towards what the last person they talked to said was the truth rather than what was in their training data.