A model performs excellently on its training data and poorly on new cases. Which pair of terms describes this, and what does it mean for subgroups thinly represented in the data?

• AUnderfitting and high bias, and thin subgroups are unaffected because the model is too simple to distinguish them.
• BDrift and decay, and subgroups are affected equally over time.
• CPoisoning and extraction, and subgroups are affected only where an attacker targets them.
• DOverfitting and high variance, and thin subgroups suffer first because the model has least evidence about them.