Does my multimodal model learn cross-modal interactions? It’s harder to tell than you might think!