What 'Observed Exposure' Establishes Institutionally
The methodological move Anthropic makes here is consequential beyond the headline numbers. Prior AI labor impact studies worked from capability assessments — what can models theoretically do — and mapped those capabilities to occupational task descriptions. Massenkoff and McCrory's approach inverts that logic: they start from what Claude users are actually requesting and build exposure estimates from observed behavior. The result is a measure that tracks adoption, not potential.
That distinction matters institutionally because it shifts the policy frame. If exposure were purely theoretical, the buffer against displacement would be technological — waiting for models to improve. Under observed exposure, the buffer is behavioral: the gap exists because workers and organizations have not yet integrated AI into their workflows. That buffer will close through ordinary process change, not a capability breakthrough. The research puts the locus of disruption in adoption dynamics, which HR functions and workforce planners control, not in the labs.