What Frey-Osborne 2013 did
Carl Benedikt Frey and Michael Osborne, in "The Future of Employment: How Susceptible Are Jobs to Computerisation?" (Oxford Martin Programme, 2013), estimated the probability that each of 702 US occupations could be computerised within roughly two decades. The paper used a Gaussian process classifier trained on aggregate occupation features (cognitive demand, physical dexterity, social intelligence). The headline finding was that 47% of US employment was at high risk of computerisation.
What the data subsequently showed
By 2024, the predicted high-risk employment had not materially declined. The US unemployment rate in 2024 was historically low. Many occupations Frey-Osborne flagged as high-risk (including various clerical and customer-service roles) had grown rather than declined. The discrepancy is now well-documented in the academic literature (see Brookings 2024, the OECD work, and various retrospective analyses).
Reason one: pre-LLM
Frey-Osborne 2013 predates the 2022 generative-AI wave. The underlying classifier features (cognitive demand, social intelligence, physical dexterity) were calibrated against pre-LLM AI capabilities. Generative AI in 2024-2026 is most disruptive in a set of cognitive-task domains that pre-LLM AI was poorly positioned to enter (writing, summarisation, image generation, code completion). The 2013 classifier features do not map cleanly onto the 2024-2026 capability frontier.
Reason two: occupation-level rather than task-level
Frey-Osborne 2013 outputs a single computerisation probability per occupation. The 2024 methodology (Brookings 2024) operates at the task level within each occupation. Task-level analysis captures the empirical pattern that is now visible: most occupations have a mix of exposed, augmentation-prone, and human-only tasks, and overall occupation outcomes depend on the mix and on how the work is reorganised.
Reason three: exposure conflated with displacement
Frey-Osborne 2013's "computerisation probability" was interpreted (correctly or not) as a probability of displacement. The 2024 methodology distinguishes exposure (technical feasibility) from displacement (labour-market outcome) explicitly. The distinction matters because most exposed tasks have not led to displaced workers in the 2024-2025 data.
Why the paper still appears in references
Frey-Osborne 2013 is acknowledged as the foundational paper of the modern "AI-and-jobs-at-risk" research stream. It powered the first generation of automation calculators and remains a frequently-cited reference. The site cites it for completeness at /sources/, classified as Excluded with the reasons documented inline. It is a piece of intellectual history for the field; it is not the right primary source for a 2026 calculator.
What this calculator uses instead
The primary score is the ILO 2025 refined Generative AI Occupational Exposure Index (a 2025 update built specifically for the generative-AI wave at the ISCO-08 6-digit task level). The task tags are derived from the Brookings 2024 task-level rubric. The growth panel is BLS Employment Projections 2025-2035 plus WEF Future of Jobs Report 2025. None of the four primary sources are pre-LLM, and all distinguish exposure from displacement.