The eight challenges below are the most-likely critiques the calculator will receive on LinkedIn, in journalist articles, and in academic comment threads. Each is paired with the calibrated response. The responses are not defensive; they are the honest version of the methodology's limits and where the calculator does and does not hold up.
Challenge 01
OECD's index uses ISCO-08, you are applying it to O*NET. The crosswalk introduces error.
Acknowledged. The crosswalk used here is the BLS-published SOC-to-ISCO mapping. Where the mapping is one-to-many, the calculator uses the dominant match. The methodology page documents the crosswalk inline at /methodology/#algorithm. The error is real and is one reason the calculator outputs four bands rather than a continuous percentile.
Challenge 02
Brookings's task-level rubric is built on OpenAI's task-completion data. You are using the disruptor's own classifier.
Acknowledged. This is the methodology's most-significant single limitation, and Brookings 2024 itself flags it. The calculator does not use Eloundou et al. (also OpenAI-affiliated) as primary methodology for the same reason. The ILO 2025 refined index is the primary score and serves as a triangulation against the Brookings task layer. Where the two diverge for a given occupation, the per-occupation page calls it out.
Challenge 03
Frey-Osborne 2013 estimated 47% of US employment was at high risk. It did not happen. Why is this different?
Frey-Osborne 2013 was pre-LLM, used a Gaussian process classifier on aggregate occupation features, and conflated exposure with displacement. The 2024 methodology used here (ILO 2025 plus Brookings 2024) is task-level, post-LLM, and explicitly distinguishes exposure from displacement. The improvements address the specific failures of the 2013 prediction. The site does not claim displacement; it reports exposure. The distinction is documented at /glossary/#ai-exposure and /glossary/#ai-displacement.
Challenge 04
BLS projections do not explicitly model AI as a discrete disruption variable.
Correct. BLS Employment Projections are aggregate workforce projections covering demographics, demand, policy, and technology together. The calculator uses BLS for the growth panel (where AI is one factor among several), not for the risk score (which uses ILO 2025 exposure data). The split is intentional: aggregate workforce projection where the data is aggregate, AI-specific exposure where the data is AI-specific.
Challenge 05
Aggregate predictions about technology and employment have a poor track record.
Aggregate predictions about displacement have a poor track record. Aggregate measurements of exposure (the share of tasks AI can technically do) are a different category. The calculator outputs exposure, not displacement. The distinction matters because the data Brookings 2024 published is task-level technical-feasibility data; the data the calculator does NOT publish is displacement prediction. Where the public discussion conflates the two, the calibrated answer is to keep them separate.
Challenge 06
Your task tags (Displaceable / Changing / Growing) are subjective.
The tags are derived from the Brookings 2024 published rubric (technical feasibility plus contextual feasibility) applied to O*NET task statements. The applied rubric is documented at /methodology/#task-rubric. A reader can examine any task tag, the underlying O*NET task statement, and the Brookings rubric, and challenge the classification on the merits. The tags are reproducible from the source data; they are not subjective in the colloquial sense.
Challenge 07
You should show a single percentage like other calculators do.
Most other calculators show a Frey-Osborne-derived computerisation probability or an unsourced single percentile. The ILO 2025 refined index outputs four exposure gradients, not a continuous percentage. Showing a fake-precise single number when the underlying data is band-level would be fabrication. The four-band output reflects the source data; it is calibrated, not under-engineered.
Challenge 08
You are cherry-picking sources to make AI sound less scary than it is.
The included sources are the most-cited and most-defensible primary sources currently in the public domain (ILO 2025, OECD AI work, Brookings 2024, BLS 2024-2034, O*NET 30.2, WEF 2025). The excluded sources are documented at /sources/ along with the reasons (Frey-Osborne pre-LLM, Eloundou et al. OpenAI-derived). McKinsey and Goldman Sachs are referenced for context but not used as primary methodology, with the reasons stated. Suggestions for additional defensible primary sources are welcomed.
Where the calculator's limitations actually are
Three honest weaknesses. First, the ILO 2025 gradient is one of four bands and cannot resolve fine-grained differences within a band; two High-band occupations may differ substantially in actual exposure. Second, the Brookings task-level rubric depends on O*NET task lists, which are updated quarterly but lag actual workplace tasks by months to years. Third, the BLS projection cycle is decade-scale; rapid 2024-2026 shifts may not yet be in the published projection.
How to verify the calculator's claims independently
Every primary source is linked at /sources/. A reader can pull the ILO 2025 refined index, the BLS 2024-2034 projections, the WEF 2025 report, and the Brookings 2024 article, and reproduce the calculator's logic for any occupation by hand. The calculator is not a black box; it is a typeset application of public data.
What this page is not
This page is not a complete defence. There are reasonable critiques of the calculator that no methodology can fully address (the inherent uncertainty of forecasting AI capability trajectories, the limits of any classification system applied to inherently variable jobs). The page is the calibrated pre-empt; it is not a claim that the calculator is beyond challenge.