An arXiv preprint finds ICON reduced false positives in model audits and exposed shortcuts in image and brain-imaging tasks, with limits.