How we calculate AI risk

The data

Each occupation's tasks come from O*NET, the U.S. Department of Labor's occupational database. Researchers rated about 19,000 of those tasks for exposure to large language models in "GPTs are GPTs" (Eloundou, Manning, Mishkin & Rock, Science, 2024). We use two of their measures: the share of tasks an LLM alone could do at least 50% faster (alpha), and the share exposed once AI tools are built around it (gamma).

The model

Tasks AI can do in a given year = alpha × direct adoption + (gamma − alpha) × tool adoption. Both adoption rates rise along S-curves between 2026 and 2036, with tool-dependent tasks adopted more slowly.

The headcount index starts at 100 in 2025. Only automation gained after 2025 reduces it, scaled by how fully each task can be automated and offset by expected demand growth in some fields such as healthcare.

The chance of contraction is the probability that headcount falls by 20% or more, derived from the projected decline.

Limitations

• The task ratings date from 2023; AI capabilities have moved since.
• Exposure measures what AI could speed up, not what employers will actually cut.
• The occupation list follows the U.S. taxonomy.
• Demand assumptions are judgment calls, and robots replacing physical work are not modelled.

These are projections, not official statistics. Data version 2026-10.