An index that only explains what it measures, and never what it does not, is doing marketing. This page is the other half.
It is not a prediction of job losses
Exposure measures what current AI systems can produce and what structurally stands in the way. It says nothing about whether any employer will act on that.
Whether exposed work becomes a lost job depends on the cost of the alternative, on procurement cycles, on union agreements, on regulation, on how work is organised inside a particular firm, and on plain institutional inertia. This index models none of those, on purpose. They are real, they are decisive, and they are not measurable from an occupational taxonomy.
A score of 60% does not mean 60% of those jobs will go. It means 60% of the weighted task load in that occupation is work a current system could produce with little standing in the way.
It is not a measure of how skilled or valuable the work is
Some of the least exposed work in the index is among the lowest paid. Healthcare support sits at a median of 11.9%. Handling and moving objects, the largest activity family in the taxonomy, scores 2.4%.
The frictions that protect work are embodiment, presence, licensure, private context and the cost of error. None of those is a measure of difficulty, contribution or pay. Reading low exposure as a compliment, or high exposure as an insult, gets the index exactly backwards.
It is not a forecast
Every release is scored against a stated capability reference date and frozen. Release v2026.Q3 describes September 2026. It does not extrapolate, and no figure on this site is a projection.
We publish a new release each quarter and publish what moved. The value of the series is that it is a record of what was true when, which is only worth anything if we resist the temptation to guess ahead.
The friction ratings are judgments, not measurements
Capability and the five frictions are rated by people against a published rubric. That is a judgment. We do not describe it as anything else.
What we do instead is make the judgment checkable. The rubric is published. The ratings are published. The agreement between raters is published: 84.9% exact agreement across eight raters on a shared calibration set, 100% within one point, with verification cost the noisiest dimension at 77.1%. You can take any score apart into its six components and dispute the one you think is wrong.
An index whose judgments cannot be located and argued with is not evidence, whatever its sample size.
It does not describe your job, only your occupation
O*NET describes occupations as surveyed across an economy. Your actual role is some subset of that task list, in some particular organisation, with some particular division of labour.
That is why the occupation number is the smallest thing on every page and the task table is the largest. The number is an entry point. The list is the thing worth reading, because you are the only person who knows which rows are actually yours.
Release v2026.Q3, built on O*NET 31.0. The full rubric, rating process and reliability figures are on the methodology page. Corrections and disputes: corrections@taskexposure.org.