Using predictive multiplicity to measure individual performance within the AI Act

A new paper argues the EU AI Act's requirement to report individual performance exposes 'predictive multiplicity'—where many equally accurate models make conflicting predictions for the same person. This arbitrariness in high-risk systems (like hiring or loans) may violate the Act. The authors propose new metrics, like individual conflict ratios, to quantify this disagreement and suggest providers must report it to deployers, fundamentally changing how model performance is evaluated and disclosed under regulation.