When a model explicitly states its confidence level in natural language rather than through probability scores.