How is an AI job-exposure score calculated?

What is the score built from?

The AI-exposure score is built from three sources, and from nothing else.

Occupational task data. The occupations and the specific task statements come from O*NET, the U.S. Department of Labor's occupational database. Browse it at O*NET OnLine, or read the database documentation at the O*NET Resource Center.

The AI task-exposure framework. Each task's exposure to current AI comes from the peer-reviewed framework in "GPTs are GPTs." Read it free as a preprint on arXiv, or see the published version in Science. Full citation: Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306 to 1308.

The AI capability trajectory. The outlook is anchored to METR's^* finding that the length of task an AI can complete with 50 percent reliability has been doubling roughly every seven months. See METR's^* write-up and the underlying paper.

Matching a path to a real occupation

We start by matching each career path you are considering to a standardized occupation, because AI does not replace job titles. It affects the specific tasks inside a job, and standardized occupations are how those tasks are defined.

This matters more for a student than for someone already working. You are not rating a job you hold. You are comparing paths you might take, so each path has to resolve to a real occupation before its tasks can be scored at all.

Throughout, you see the plain occupation name and how close the match is. The underlying federal occupation codes stay behind the scenes.

Weighting the work

An occupation is a set of tasks, and those tasks do not carry equal weight. The score reflects the relative emphasis across the work, so the tasks that take up the most of a working week move the number the most.

This is why two paths that sound similar can score very differently. The number follows the actual mix of work inside each occupation, not the job title on the outside.

How is the exposure score calculated?

Every task carries an exposure value drawn from the framework in "GPTs are GPTs":

• Minimal. Today's AI cannot meaningfully do the task. Exposure value 0.
• Partial. AI can do it with the right software or tools. Exposure value 0.5.
• Full. AI can already do it on its own. Exposure value 1.0.

Each task's exposure value is weighted by how much of the work it accounts for. Those weighted values are averaged and placed on a 0 to 100 scale. In plain terms: multiply each task's exposure by its share of the work, add the results together, divide by the total, and scale to 100.

A worked example. Say half of an occupation's work goes to a task AI can already do on its own, a quarter to a task AI can handle with the right tools, and the last quarter to a task AI cannot touch. That gives (0.5 x 1.0) + (0.25 x 0.5) + (0.25 x 0), which is 0.625, or about 63 on the 0 to 100 scale.

Why pay and growth never touch the score

Salary, job growth, location, your network, and your portfolio are all shown beside the exposure number. None of them can move it by a single point. That separation is the reason the number means anything: if wages could raise or lower an exposure score, the score would no longer be telling you about AI at all.

Pay and growth figures are shown as context, drawn from the U.S. Bureau of Labor Statistics: the Occupational Outlook Handbook for growth projections and Occupational Employment and Wage Statistics for pay. They are display only, and by design they cannot shift the exposure figure.

The outlook, and why there is no single future number

Alongside the current figure, we give a direction of travel: rising, steady, or already at the ceiling. It uses METR's^* established long-run time-horizon trend.

We deliberately do not publish one percentage for a specific year in the future. Pinning down a single future figure would imply a precision nobody honestly has. The useful and truthful read is the trajectory, not a false decimal.

We also use the established long-run trend rather than the faster 2026 estimates. Those newer figures come from a task suite near saturation, carry very wide confidence intervals, and measure software tasks specifically, as the title of the METR paper^* makes clear. Extrapolating them across all occupations is not justified, so we do not do it.

What does the score not measure?

It does not predict that any particular job will disappear, and it does not predict anything about you. It measures task exposure, meaning what current AI can do.

It is directional by design. It is built on today's models and the public research behind them, and it will move as both the models and the research move.

It also cannot see the things that often matter most: your employer, your skill, your judgment, and the relationships you build. Treat it as an informed starting point for thinking about a career, not as career, financial, or legal advice.

Producing it asks very little of you. No resume and no work history, just the paths you are weighing and the work you have already done. We collect only what is needed to generate your result. See the Privacy Policy for the details.

Sources and attribution

• Occupational and task data: O*NET OnLine and the O*NET Resource Center, U.S. Department of Labor.
• Task exposure framework: Eloundou et al. (2024), Science 384(6702), 1306 to 1308, also available as an arXiv preprint.
• Capability trajectory: METR^* and its paper on arXiv.
• Pay and growth context: U.S. Bureau of Labor Statistics, Occupational Outlook Handbook and Occupational Employment and Wage Statistics.

This product includes information from the O*NET Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA). Used under the CC BY 4.0 license. O*NET is a trademark of USDOL/ETA. Phronesis Labs LLC has modified some of this information. USDOL/ETA has not approved, endorsed, or tested these modifications.

Questions about the method

If something here is unclear, or you think we have a source wrong, tell us on the contact page. Corrections to this page are welcome and we would rather fix an error than defend it.