How to read your AI exposure score

What the number is made of

The score comes from a Microsoft Research study that analysed roughly 200,000 anonymised conversations between United States users and Bing Copilot, collected between 1 January and 30 September 2024.¹ It matched two separate things against the work activities that make up each occupation: what the user asked for, and what the assistant actually did. Your score is the average of those two sides.²

Three things were classified for every activity, by a language-model pipeline rather than by a person reading each conversation: how often it showed up at all, whether the assistant finished the job, and how much of the activity it handled rather than a corner of it. The authors report that pipeline agreeing with human annotators at Cohen's kappa between 0.34 and 0.53, a range they themselves call generally low, because it is often arguable whether a given activity describes a conversation at all. Their own documentation says plainly that "LLM-based classifiers are imperfect, so all metrics should be interpreted with caution".¹ Treat individual rows as indicative, not exact.

Why the band matters more than the decimal

The study's own authors are blunt about the raw figure. Because an activity only counts as covered once it crosses a frequency threshold they chose, they write that "relative comparisons are more meaningful than absolute score values", and they demonstrate the point by showing that moving the threshold lets you conclude almost anything.¹

So a score of 0.195 does not mean 19.5% of anything. Read the band and the percentile instead. The five bands are quintiles: each one holds exactly one fifth of the 785 occupations, so "elevated" means "in the fourth fifth", not "in danger". The exact cut-offs and the arithmetic behind them are on the methodology page.³

What "not observed" really means on your task table

Of the 332 work activities measured, 205 score zero on the user-goal side, which is the side the task table shows, and the table labels those "not observed".² That reads like "AI cannot do this". It does not mean that.

Checking the published dataset directly: 203 of those 205 activities did appear in the conversations.² They simply fell below the study's cut-off of 0.05% of all classified activity, so they score zero by construction. That cut-off measures share of activity, not share of conversations: each conversation is divided equally between the activities it matches. Earlier versions of the paper glossed that cut-off as roughly 100 to 300 conversations in samples of 100,000, before conversion to activity share; version 6, of 22 December 2025, drops that gloss.¹

Two of the 205 did not appear on the user-goal side at all: fabricating medical devices and supervising activities in a correctional facility. Of those, only supervising a correctional facility is absent from both sides of the dataset. That is one activity out of 332 that never appeared anywhere.²

The honest reading of a zero is therefore: this activity almost never came up in this dataset. There are at least four reasons that can happen, and the number cannot tell you which applies. The work may be genuinely hard for AI. It may happen in a warehouse rather than a chat window. Nobody may have thought to ask. Or, and the study's authors name this one themselves, the people who do that work may use an entirely different tool: programmers work inside AI-enhanced development environments, and legal, medical and financial work runs on compliance-approved systems. None of that traffic appears in a consumer chat log.¹

Three ways people misread the score

Reading it as a probability. A score of 0.25 is not a 25% chance of anything. There is no time horizon anywhere in the underlying study and no model of employment in it at all.

Comparing it to numbers from other studies. Other well-known figures count entirely different objects. Eloundou and colleagues estimate that around 80% of United States workers could have at least 10% of their work tasks affected by large language models, an estimate of potential exposure rather than a measurement of what has happened; the International Monetary Fund put almost 40% of global employment in an exposed category, with the high and low cut-offs set at the median.⁴ Neither is comparable to a score built from usage frequency. Exposure, automation and replacement sets out what each layer of evidence can and cannot support.

Treating a zero as a guarantee. Zero means the activity was rare in this dataset, not that AI cannot do it.

What this means for you

Use the band and percentile to place yourself, not to predict yourself. Then use the task table, which is the genuinely useful part: it shows which parts of your work overlap with AI use and which do not, and that distinction is what any sensible plan is built on. If your own job differs from the occupation average, and most people's do, the six questions on your occupation page adjust the estimate and show you every adjustment they make.

Start from the calculator if you have not looked up your occupation yet, or work from your actual weekly tasks with the task self-assessment.