A candidate-screening model has acceptable overall accuracy, but an audit finds substantially worse error rates for one demographic subgroup. Which conclusion is most appropriate?

1. Aggregate performance can hide unequal outcomes; investigate subgroup metrics, data, labels, and the intended fairness requirements.
2. Removing subgroup evaluation results makes the model fairer.
3. Acceptable overall accuracy proves that every subgroup receives equally good outcomes.
4. A lower inference latency automatically resolves the observed subgroup error gap.