The short answer: a task counts as exposed when a language model, or software built on one, could cut the time it takes by at least half without lowering quality. That is a claim about speed on one task. It says nothing about whether anyone buys the tool, whether the law allows it, or whether the job goes away.
Where the definition comes from
The rubric behind almost every serious exposure estimate comes from Eloundou, Manning, Mishkin and Rock, published in Science in 2024. For each task in the federal O*NET catalogue, they asked one question: would access to a language model, or to software built on top of one, reduce the time required to complete this task by at least half, while keeping the output at the same quality?
Every clause in that sentence is load-bearing, and the coverage tends to drop most of them.
"At least half the time" is a threshold, not a vibe
The vaguer question, "could AI do this," has no useful answer. Almost anything can be attempted. The Science rubric replaces it with a measurable bar: a fifty percent reduction in time. A task where the output arrives faster and worse is not rated as exposed at all, because the quality condition fails.
That matters for how you read your own work. Drafting a first version of a routine document clears the bar comfortably. Producing a document that a client will sign, at the standard your firm is held to, often does not, even when the drafting step inside it does.
The clause that moves the number most
The phrase about software built on top of the model matters more than anything else in the definition. The ratings separate what a chat window can do unaided from what becomes possible once real tooling is wrapped around the model: integrations, retrieval, checks, the plumbing that turns a demo into a product.
Those two assumptions produce dramatically different numbers. In the Science paper, the share of occupations with more than half their tasks exposed runs from about 1.8% for models alone to about 46% once complementary software is assumed.
From 1.8% to 46%, on the same underlying ratings, depending entirely on what you assume gets built. That range is the honest state of the evidence. Anyone quoting one end of it without the assumption attached is quoting half a sentence.
Why the same title gets different scores
The rubric is applied to tasks, not to job titles. O*NET lists roughly eighteen tasks for a typical occupation. Some of them clear the fifty percent bar, some sit near the floor, and the score for a person depends on how their week is split across that list.
Two project coordinators can share a title while one spends the week on status reports and scheduling and the other on vendor negotiation and site visits. The first has a week full of exposed tasks. The second does not. Twenty points of difference is ordinary once you measure the week instead of the title. The AI Job Risk Check score weights each task rating by the hours you report on it, which is the whole reason it can tell those two people apart.
What to do with this
When you next read a number about AI and jobs, look for two things. First, the definition: is it the Science rubric, or something looser? Second, the assumption: models alone, or models plus software? If either is missing, the number is not wrong so much as incomplete, and you cannot compare it with anything.
Then apply the same discipline to your own week. Pull up your occupation on O*NET, read the task list, and ask the Science question of each line: could a model, or software built on one, halve the time without losing quality? The tasks where the answer is yes are your exposure. The rest is the part of your job the headlines are not describing.
Frequently asked questions
What does it mean for a task to be exposed to AI?
Under the Science rubric, a task is exposed if a language model, or software built on one, could cut the time to complete it by at least half without reducing quality. It is a statement about speed on a single task, not a prediction about the job.
Why do exposure estimates range from 1.8% to 46%?
The lower figure assumes people use a bare language model. The higher figure assumes complementary software gets built around it. Both come from the same task ratings in Eloundou et al. (2024); only the tooling assumption changes.
Does high exposure mean my job will be automated?
No. Exposure measures what is technically possible on individual tasks. Whether your employer adopts the tools, whether regulation requires a human, and whether demand for the occupation grows are separate questions that the rating does not address.
Get your free AI Job Risk Score. Tell us your job title and how you actually spend your time, and we will show you which of your tasks are exposed today. Free. 60 seconds.
Get my AI risk now