A classifier flags unsafe messages. Only 2% of messages are unsafe, and an always-safe baseline achieves 98% accuracy. What should the team conclude?

1. The dataset must contain exactly 98 classes.
2. Accuracy alone is misleading; inspect unsafe-class precision, recall, and the confusion matrix.
3. The baseline is suitable because 98% accuracy proves unsafe messages are detected.
4. Remove every unsafe example before evaluating.