Every role on AI Proof is scored using a structured, research-backed methodology. Here's how it works — and how you can help improve it.
Each score answers a single question: how likely is it that AI eliminates this job over the next decade? Not whether AI changes the job, automates some tasks, or makes workers more productive — but whether the job itself goes away.
Roles receive a score from 0 to 100. Higher scores mean greater resilience against AI displacement. The score is derived from six weighted factors, each measuring a different dimension of how protected a job is.
Labor demand, salary data, and employment growth projections are displayed on each role page but do not affect the score. A role can score low while still having strong hiring demand — both facts matter, and you'll see both.
O*NET (U.S. Department of Labor) — Canonical task lists for each occupation. Every role's tasks are pulled directly from O*NET and frozen. We don't invent or modify tasks.
Bureau of Labor Statistics — Salary data from the Occupational Employment and Wage Statistics survey (May 2024) and employment growth projections (2024–2034). These are the official U.S. labor market figures.
AI Industry Research — Documented AI deployments, layoff reports, robotics research, and regulatory developments. We track what's actually happening, not just what's theoretically possible.
We use frontier large language models from Anthropic, Google, and other leading providers as core tools in our research and scoring pipeline. These models help synthesize large volumes of labor market data, parse occupation task lists, evaluate the current state of AI capability against specific job functions, and produce structured factor assessments according to our methodology.
Every score is reviewed by a human before publication. The models accelerate research and enforce consistency across hundreds of roles, but the methodology itself — the factors, weights, calibration standards, and editorial judgment — is human-designed and human-governed.
Each role is evaluated across six dimensions. The factors carry different weights reflecting their relative importance to job displacement risk. Every factor is scored 0–100 and the weighted combination produces the final score.
How much of this role's daily work is beyond AI's current capabilities? We classify every task as exposed (AI can do it), augmented (AI assists but a human drives), or safe (human-essential). Augmented tasks protect jobs — a worker using AI to be more effective is more secure, not less.
Does this job require a human to physically be there? Three types of presence count: hands-on work (plumber, surgeon), field conditions (firefighter, paramedic), and embodied presence — being physically in the room for institutional functions like boardroom leadership, courtroom proceedings, or stakeholder negotiations.
How essential is it that a person — not just a machine — does this job? We evaluate four mechanisms: clinical trust (patients need a human), institutional accountability (governance requires a human), interpersonal judgment (reading people in real time), and customer preference (people prefer a human). The first two are most durable.
Does the law require a human in this role? Two types of legal protection: credential-based (licensure, exams, supervised practice) and mandate-based (law structurally requires a human person — fiduciary duty, corporate governance, constitutional requirements). Both are scored equally.
How fast is AI getting better at doing this work? Scored against the role's core function, not peripheral tools. If AI improves the analytics a CEO uses but no research targets autonomous organizational leadership, the core function gap isn't closing and the score stays high.
Are people in this role actually losing jobs to AI right now? This factor measures documented evidence — real layoffs, real headcount reductions in this specific role — not theoretical risk or displacement in adjacent roles. If nobody has been fired, the baseline score is high.
Every occupation has an official list of tasks from O*NET. We classify each task into one of three categories based on how AI interacts with that work today and in the near future.
AI can perform this task autonomously or nearly so. A human reviewing the output would approve it the vast majority of the time without modification.
AI meaningfully speeds up or improves this task, but a human still drives the process, makes key decisions, or handles exceptions. Augmented tasks actually protect jobs — they make the worker more productive, not more replaceable.
This task requires physical presence, trust, novel judgment, or real-time human adaptation that AI cannot replicate. AI involvement is minimal or absent.
The final weighted score maps to a resilience tier. These labels are derived directly from the score — they aren't assigned manually.
We don't think scoring should be a black box. Every role page shows its full task classification, and you can weigh in on whether we got it right.
For each classified task, you can:
Upvote
This task is classified correctly — you agree with where it landed.
Flag for Reclassification
You think this task belongs in a different category — exposed, augmented, or safe.
Votes help us identify classifications that real practitioners disagree with. If you work in the field and think we've misjudged how AI impacts a specific task, your input directly informs future rescoring.
AI is moving fast, and our scoring evolves with it. The methodology is versioned — when we update factor definitions, adjust weights, or refine calibration, every affected role gets rescored under the new version. Score history is preserved so you can see how a role's assessment has changed over time.
The current methodology version is v10, last updated February 2026.
Ready to see how your role scores?
Explore the Directory