AI exposure estimates increasingly shape public debate, workers’ career decisions, and government reskilling policies. Yet new research shows these measures are far less stable than commonly assumed. Using identical occupational data but four leading AI models, estimates of “high exposure” occupations ranged from 14% to 51%, and even reversed predictions about employment effects. The differences reflect deeper uncertainty about what AI exposure actually measures, from task automation to productivity gains or job displacement.
While such indicators remain useful for tracking technological change, researchers and policymakers should present ranges rather than single figures, communicate uncertainty clearly, and avoid basing labour market interventions on any one model’s estimates.
