A 2023 paper titled "GPTs are GPTs" by Eloundou et al. estimated that 80% of the U.S. workforce has tasks exposed to large language models (LLMs), with 19% having 50% or more of their work tasks impacted. These figures have been widely integrated into global policy discussions, including reports by the IMF, OECD, and various U.S. Senate proposals. Despite their prevalence, Cohere researchers argue that applying these static exposure scores—which measure the technical feasibility of AI performing discrete work tasks based on 2023 capabilities—to current labor market decisions creates a significant evidence gap.

The limitations of these original scores stem from several factors. First, they rely on a 2023 model, while current AI capabilities have advanced, creating an estimated 26 percentage point gap. Second, the metrics use an American occupational taxonomy (a classification system for job roles) that does not universally apply to other labor markets. Finally, the scores treat work as a series of itemized, discrete tasks, failing to account for essential human elements such as professional judgment, workplace relationships, and situational context. Researchers note that these constraints compound when policymakers use the data to inform decisions for 2026 and beyond.

To address these shortcomings, researchers are developing more dynamic, representative measurement tools. Emerging approaches include dynamic indexes that link AI capability evaluations to real labor market data, such as a study finding that a 10-point increase in exposure is associated with a 5.6 to 8.5 percentage point decline in employment. Other methods, such as ensemble approaches and task-framework extensions, are being used to create more nuanced estimates by weighting multiple frameworks and analyzing task sequencing.

The research team emphasizes that exposure scores should function as one signal among many rather than a definitive forecast. They advocate for policymakers to strengthen worker protections and invest in reskilling infrastructure while engaging employees as partners. Furthermore, researchers are encouraged to prioritize measurement tools that evolve alongside AI capabilities and incorporate worker-centered measures—such as evaluating tasks workers prefer not to automate—to provide a more complete picture of labor market vulnerability.

A 2023 paper titled "GPTs are GPTs" by Eloundou et al. estimated that 80% of the U.S. workforce has tasks exposed to large language models (LLMs), with 19% having 50% or more of their work tasks impacted. These figures have been widely integrated into global policy discussions, including reports by the IMF, OECD, and various U.S. Senate proposals. Despite their prevalence, Cohere researchers argue that applying these static exposure scores—which measure the technical feasibility of AI performing discrete work tasks based on 2023 capabilities—to current labor market decisions creates a significant evidence gap.

The limitations of these original scores stem from several factors. First, they rely on a 2023 model, while current AI capabilities have advanced, creating an estimated 26 percentage point gap. Second, the metrics use an American occupational taxonomy (a classification system for job roles) that does not universally apply to other labor markets. Finally, the scores treat work as a series of itemized, discrete tasks, failing to account for essential human elements such as professional judgment, workplace relationships, and situational context. Researchers note that these constraints compound when policymakers use the data to inform decisions for 2026 and beyond.

To address these shortcomings, researchers are developing more dynamic, representative measurement tools. Emerging approaches include dynamic indexes that link AI capability evaluations to real labor market data, such as a study finding that a 10-point increase in exposure is associated with a 5.6 to 8.5 percentage point decline in employment. Other methods, such as ensemble approaches and task-framework extensions, are being used to create more nuanced estimates by weighting multiple frameworks and analyzing task sequencing.

The research team emphasizes that exposure scores should function as one signal among many rather than a definitive forecast. They advocate for policymakers to strengthen worker protections and invest in reskilling infrastructure while engaging employees as partners. Furthermore, researchers are encouraged to prioritize measurement tools that evolve alongside AI capabilities and incorporate worker-centered measures—such as evaluating tasks workers prefer not to automate—to provide a more complete picture of labor market vulnerability.