Home / AI Risk Index / Methodology
How we score AI job risk
Every occupation in the AI Risk Index carries two scores from 0 to 10. This is how each is made, the exact rubric the model scored against, and the limits of what they can tell you.
1. The two measures
Automation exposure is how much of the occupation’s work current AI — language models, code and image generation, speech-to-text and similar widely deployed tools — can already perform or is plainly about to.
AI assistance potential is how much those same tools make a worker in the occupation more capable without displacing them. It measures upside, not risk, and it is not the inverse of exposure: software development scores high on both, commercial diving low on both.
2. How a score is built
Exposure is scored task by task. Each of the occupation’s O*NET tasks gets its own 0–10 score, and the occupation’s score is their average weighted by how important O*NET rates each task:
Σ(task score × task importance) ÷ Σ(task importance)
Assistance potential is scored once for the whole occupation, from its title, description and most important tasks, and comes with the model’s one-paragraph reasoning, which each occupation’s page shows.
3. Reading a score
Both scores share one scale, and the band words follow the rubric’s own two-point steps:
- Minimal — 0–2
- Low — 2–4
- Moderate — 4–6
- High — 6–8
- Very high — 8–10
Exposure scores cluster in the middle of the scale — the most exposed occupation in the index scores 6.8. That is why each occupation’s page also says what share of the index it is more exposed than: the position is often more telling than the number.
4. The model and the run
Every score comes from one scoring run with gemini-3.6-flash, completed on August 23, 2026. Exposure used rubric version 1.0.0 and assistance version 1.0.0. A rubric that changes gets a new version, and only then are the occupations re-scored.
5. The exact rubrics
These are the prompts sent to the model, word for word. The braces mark where each occupation’s own data goes.
Automation exposure — per task
You are an expert analyst evaluating how exposed specific work tasks are to current AI capabilities (LLMs, code-generation tools, image-generation, voice-to-text, and similar widely-deployed digital systems as of 2026).
You will be given a numbered list of tasks, all from a single occupation. Rate EACH task independently from 0 to 10:
- 0-1: Current AI cannot meaningfully perform or assist with this task.
- 2-3: AI can assist with peripheral aspects but the task itself requires human capability.
- 4-5: AI can perform substantial portions of this task with human review.
- 6-7: AI can perform most of this task; humans verify outputs.
- 8-9: AI performs this task at near-human level today; humans intervene only for edge cases.
- 10: This task is fully automated by deployed AI systems already.
Consider what AI can do today and what is plainly within reach in the next 12-18 months given current capability trends. Do not speculate about long-term futures. Do not assume zero deployment friction; consider whether enterprises are actually deploying AI for tasks of this kind.
Return one object per task, in the SAME ORDER as the input, with "task" as the 1-based task number.
Tasks:
1. {each of the occupation’s O*NET tasks}AI assistance potential — per occupation
You are an expert analyst evaluating how much current AI tools can make a worker in this occupation more productive, valuable, or capable — without displacing them.
This vector measures upside, not risk. A worker with high augmentation potential becomes more valuable when they adopt AI tools well; a worker with low augmentation potential gains little from those tools.
Rate 0-10:
- 0-1: No meaningful AI augmentation available; tools don't apply to the work.
- 2-3: Marginal improvements available (scheduling, paperwork) but core work is unaffected.
- 4-5: Meaningful productivity gains available for portions of the work.
- 6-7: AI substantially extends the worker's reach; an AI-fluent worker is materially more capable than a non-fluent peer.
- 8-9: AI tools transform what an individual practitioner can accomplish; AI-fluency is becoming a key differentiator in the field.
- 10: AI tools are now a precondition for performing the work at the current standard of practice.
Augmentation potential is meaningfully different from inverse-of-displacement. A job can score high on both Digital AI Displacement AND Augmentation (e.g., software developer — transformed but the augmented worker is more valuable). A job can also score low on both (e.g., commercial diver — AI doesn't displace and doesn't help much either).
Consider deployed tools, real workflow integration, and the productivity ceiling — not just theoretical capability.
Occupation: {occupation title}
Description: {O*NET description}
Top tasks: {its most important tasks}6. What these scores do not tell you
- Exposure is not job loss. Whether an exposed task is automated depends on cost, regulation and how employers choose to work, none of which is scored here.
- Only AI is scored. Robotics and offshoring are separate pressures on physical and remote work, and the index does not measure either yet.
- Occupations are not weighted by how many people do them. The index says how exposed each occupation is, not what share of the workforce is exposed.
- The task scores are not kept. Only the weighted average and the number of tasks behind it are stored, so the index cannot show which tasks drive a score.
- It is one model’s judgement at one point in time. The rubric asks about today and the next 12–18 months, not the long-term future.
How the assessment itself matches you to occupations is a separate method — see the assessment methodology.
Exposure is half the question
The other half is fit. Take the assessment to see which occupations match your traits. A preview of your results is free; the full report is $10.00.
Start the assessment