A hiring model performs well overall but has a much higher false-rejection rate for one demographic group. Which responsible AI concern should be investigated first?

1. Token caching
2. Endpoint naming
3. Fairness
4. Storage redundancy