Published research · attributed

What independent research says (not RoleMath’s claim)

• Eloundou et al. estimate that about 80% of U.S. workers have at least 10% of their work tasks exposed to large language model capabilities (Science 2024). American Association for the Advancement of Science exposure = task overlap, not job loss.
• Eloundou et al. estimate that about 19% of U.S. workers have at least 50% of their work tasks exposed to large language model capabilities (Science 2024). American Association for the Advancement of Science exposure = task overlap, not job loss.
• Eloundou et al. explicitly disclaim any forecast of AI adoption or timing, describing their measure as capability overlap with tasks rather than a prediction of job loss (Science 2024). American Association for the Advancement of Science exposure = task overlap, not job loss.
• OECD reports that high-skill occupations are the most exposed to AI on task-overlap measures (OECD Employment Outlook 2023). Organisation for Economic Co-operation and Development exposure = task overlap, not job loss.
• OECD reports that, as of 2023, there is little empirical evidence of negative employment effects from AI (OECD Employment Outlook 2023). Organisation for Economic Co-operation and Development exposure = task overlap, not job loss.
• OECD and the AIOE research find that AI exposure and automation risk often run in opposite directions, with the most-exposed high-skill occupations tending to be the least at risk of automation. Organisation for Economic Co-operation and Development exposure = task overlap, not job loss.
• Felten, Raj and Seamans construct an occupation-level AI Occupational Exposure index by linking AI capabilities to O*NET occupational abilities (Strategic Management Journal). Strategic Management Journal (Wiley) exposure = task overlap, not job loss.
• Stanford Digital Economy Lab researchers find a roughly 16% relative decline in employment for workers ages 22-25 in the most AI-exposed occupations, based on high-frequency ADP payroll data (Canaries in the Coal Mine, working paper). Stanford Digital Economy Lab correlational usage data, not proof.
• The ILO notes that AI-exposure indicators measure potential task overlap and cannot by themselves establish job loss (Workers' exposure to AI). International Labour Organization exposure = task overlap, not job loss.
• The Anthropic Economic Index reports no measured systematic rise in unemployment attributable to AI in its usage data. Anthropic correlational usage data, not proof.

Tier A research stays attributed and separate from BLS outlook and employer-language samples.