How CareerForge AI calculates AI risk scores

The short version

Median wages and BLS employment projections on our role pages are published government figures, reproduced as published. The 0-100 exposure score, the automation trajectory percentages, the projected 2030 salary and the jobs-at-risk count are editorial estimates informed by the automation literature. They are not computed from a public dataset, and we do not present them as measurements.

Section 1, What the scores measure

A CareerForge exposure score describes task composition. We look at what a role actually spends its hours doing, ask which of those tasks current AI systems can perform to an acceptable standard, and express the result on a 0-100 scale with a band: low, moderate, high or critical. It is a statement about the work, not about you and not about your employer. Two people holding the same title can carry very different exposure, because the mix of judgment, relationship and routine processing in their weeks differs. That is why we score tasks rather than titles. The score estimates from published research. It does not read your company's plans, your performance record or your industry's capital budget, because we hold none of that.

Section 2, Data sources

O*NET, U.S. Department of Labor (https://www.onetonline.org/): task statements used as the reference vocabulary for describing each role; consulted as reference, not machine-ingested. BLS Occupational Outlook Handbook, projections 2024-2034 (https://www.bls.gov/ooh/): published median annual wage and 10-year projected employment change; directly sourced and reproduced as published. Frey and Osborne (2017), The future of employment: foundational reference for automation probability tracking task composition rather than job title; conceptual basis only, their occupation-level probabilities are not reproduced as our scores. Eloundou, Manning, Mishkin and Rock (2023), GPTs are GPTs (https://arxiv.org/abs/2303.10130): reference for language-model exposure at the level of individual O*NET tasks and for the augmentation-versus-displacement distinction; conceptual basis only, their per-task ratings are not loaded into our scoring. McKinsey Global Institute, Generative AI and the future of work (2023): directional reference for the pace and shape of work-hour automation by occupational group; our per-role percentages are our own estimates, not McKinsey figures. World Economic Forum, Future of Jobs Report 2025: employer-reported expectations about which roles grow and decline, used when choosing adjacent roles; directional reference. Stanford Digital Economy Lab, Canaries in the Coal Mine? (August 2026) (https://digitaleconomy.stanford.edu/news/canariesaug26/): evidence that the strongest adverse employment signal so far is concentrated among young workers in highly exposed occupations rather than experienced workers; directly cited claim.

Section 3, Methodology

Step 1, define the role's task set: take the occupation's task statements and work activities from O*NET as the reference description, then reduce them to the task clusters accounting for most of the hours a practitioner actually spends. Step 2, categorise each cluster as automatable, augmentable or durable, using the Eloundou et al. framing, with routine information processing and pattern-based drafting at the automatable end and work whose value comes from accountability, negotiation, physical presence or contested judgment at the durable end. Step 3, assign the composite score: the share of typical hours falling in the automatable and augmentable classes, weighted toward automatable, then adjusted against the direction of the McKinsey and WEF findings for that occupational group and against the BLS employment projection. That adjustment is an editorial judgment by our research team; there is no published coefficient behind it and we will not pretend otherwise. Step 4, estimate the timeline: the 2025 / 2027 / 2030 trajectory expresses how much of the task set we judge technically automatable at each point, in the direction the literature indicates. The 3-6 year window quoted on assessment results is a planning horizon, long enough for a real transition and short enough to act on. It is not a forecast date, and no study we cite publishes one. Step 5, attach the published figures: median wage and 10-year employment change come straight from the BLS Occupational Outlook Handbook and are shown as published, and adjacent roles are chosen for overlap with the durable clusters and screened against BLS growth so we never point you at a shrinking destination.

Section 4, What the scores are NOT

Not an employer forecast: we hold no data about your company's headcount plans, budgets or restructuring decisions. Not a guarantee of job loss or of safety: a low score is not protection and a high score is not a redundancy notice. Not real time: the dataset is reviewed quarterly and between reviews the numbers do not move, even if the news does. Not uniform across individuals: two people with the same job title can hold very different task mixes, and the score describes the role's typical composition rather than your personal week. Not a labor-market projection: where you want an official employment projection, use the BLS figure we cite and link, not our score.

Section 5, The 40 to 60 professional distinction

General AI risk calculators score an occupation for everyone in it. Our role pages do something narrower on purpose. Task-level analysis is applied to long-tenure career profiles: five role pages (paralegal, accountant, operations manager, data analyst and project manager) currently carry a section written for someone with fifteen to twenty-five years in the work, naming the latent expertise that tenure produces and that a general calculator has no way to represent, and we are writing these role by role rather than generating them, so the remaining pages do not have one yet. Legibility Debt is layered on top of exposure: exposure asks whether the work can be automated, while Legibility Debt, from the RECAST framework at /recast, asks whether current screening systems can still read what you are worth. An experienced professional can hold low exposure and high Legibility Debt at the same time, and only the second one is losing them interviews. Age-tell detection sits in the resume tooling: our analyzer flags the wording, formatting and date patterns that mark a record as dated to a screener, and we are not aware of another AI risk calculator that does this. One thing we will not do: claim experienced workers are being displaced faster than everyone else. The Stanford Digital Economy Lab's August 2026 work points the other way. The mid-career problem is legibility, not obsolescence.

Section 6, Version history and updates

Version 1.0, published 15 September 2026. First published methodology for the 57-role exposure dataset. Sources: O*NET, BLS Occupational Outlook Handbook 2024-2034, Frey and Osborne (2017), Eloundou et al. (2023), McKinsey Global Institute (2023), WEF Future of Jobs 2025, Stanford Digital Economy Lab (August 2026). This version states plainly which figures are directly sourced and which are editorial estimates. Next scheduled review: 15 December 2026, at which we plan to replace our editorial trajectory estimates with values computed from the Eloundou et al. per-task exposure ratings joined to O*NET task importance weights, and to remove the projected 2030 salary figures unless they can be tied to a published projection. Report an error to methodology@careerforgeai.com; if a figure cannot be defended we change it and record the change here.

Section 7, Limitations

The composite score is an editorial estimate, not a computed statistic: it is assigned by our research team after reading the task composition of the role against the literature, it is not the output of a published formula, and two reasonable analysts could assign scores several points apart. The automation trajectory percentages are directional: they describe the share of typical tasks we judge technically automatable, and technical feasibility is not employer adoption, which regulation, liability, union agreements, capital cost and organizational inertia all slow. The projected 2030 salary figures are estimates: only the current median wage is a published BLS figure. The jobs-at-risk counts are order-of-magnitude, useful for scale and not for citation. Timelines are estimates, not findings: no study we cite publishes a per-occupation date, and any site that gives you one is overstating what the evidence supports. Legibility Debt is a framework construct under test, a published proposition rather than a validated instrument, and our own outcome dataset is still being collected so we publish no benchmark distributions yet. Coverage is 57 roles, not the whole economy: titles outside the 57 resolve to the nearest researched occupation or an occupational-family average, and the tool tells you which of those happened rather than hiding the fallback. Read together, our exposure scores are a defensible way to decide where to spend the next ninety days. They are not a measurement and should not be quoted as one.

Get your free AI risk score →

Published by The CareerForge Institute for Mid-Career Intelligence. © 2026 The CareerForge Institute for Mid-Career Intelligence. All rights reserved.