Observed exposure is a metric developed by Anthropic economists Maxim Massenkoff and Peter McCrory in their March 2026 paper "Labor Market Impacts of AI: A New Measure and Early Evidence." Unlike prior AI displacement research that measured theoretical capability — what AI could do — observed exposure measures what AI is actually doing in professional settings by analyzing anonymized usage data from Claude across hundreds of occupations. The metric weights fully automated uses more heavily than collaborative or augmentative ones, and restricts analysis to work-related use. By combining theoretical capability data with real-world usage, observed exposure provides a more grounded picture of AI displacement risk. Computer programmers scored 75% observed exposure, meaning AI is actively performing three-quarters of that occupation's tasks in real professional workflows. By contrast, cooks, mechanics, and lifeguards scored 0%, because no meaningful professional use of Claude exists in those fields.