Policy, workers, and firms who make decisions with any one study's exposure scores are likely to get it wrong! These studies employ different definitions of "AI" and perform variably depending on the US region or sector. User beware!

So which exposure studies are the most predictive of job loss? Individually, exposure scores from Frey/Osborne, Brynjolfsson, Felten et al, and the recent Eloundou et al studies are not predictive of unemployment risk. However, when combined, they are very predictive!

How could we even test this? Commonly-used employment and wage data don't directly capture job loss. So we instead use data from each state's unemployment insurance program to calculate workers' unemployment risk based on their occupation and local labor market