Preprint tests a sharper way to audit what AI models use
An arXiv preprint finds ICON reduced false positives in model audits and exposed shortcuts in image and brain-imaging tasks, with limits.
Developing Light ยท https://developinglight.com/editorial/developing-light