Shapley analysis reveals why AVSR models keep trusting corrupted audio, exposing a hidden bias in multimodal speech recognition.