Adding OpenAI Signals to Our Model

Q3 2026 update · 6 min read

Every AI exposure dataset answers one of two questions. The first is theoretical: given what this occupation involves, how much of it could AI do? That is the question behind the Will Robots Take My Job framework and our own internal deep-research model. The second is empirical: how much of this occupation's work are people already doing with AI? Our Anthropic and Microsoft sources answer that one, each from a different vantage point.

This quarter we added a new public dataset, drawn from the most widely used AI assistant in the world: OpenAI Signals, the data release accompanying OpenAI's “How People Use ChatGPT” research. It is now the fifth source in our ‘Meaningful Human Contribution’ (MHC) dimension, which measures AI and automation exposure.

What OpenAI Signals is

Signals is OpenAI's public dataset describing how people use consumer ChatGPT. It is built from a sample of 300,000 messages per month, spanning July 2024 through March 2026, from adult accounts on consumer plans. For each work-related message, an automated classifier asks: which real-world work activity is this person trying to accomplish? The categories come from O*NET, the Department of Labor's occupational database, which breaks all work in the US economy into a few hundred “intermediate work activities” like edit written materials, diagnose health conditions, or prepare foods or beverages.

The result is a month-by-month picture of which work activities people actually bring to ChatGPT.

Turning usage into an occupational resilience score

Signals tells us how much each activity happens in ChatGPT. It says nothing directly about occupations. To connect work activities to occupations, we take each occupation's full activity profile (i.e., the specific work activities comprising that occupation), weight each activity by how important it is using O*NET's importance ratings (e.g. a nurse's administer care counts for more than their maintain records), and then multiply by how much each activity actually shows up in ChatGPT usage. The result is a usage-grounded exposure score, answering “how much of this occupation's work overlaps with what people do with ChatGPT?”

We don't score occupations where too little of the work is visible in the usage data; those lean on the other four sources.

What the data says: ChatGPT is a writing machine

The single clearest finding in the Signals data is how concentrated work usage is. The top activity, edit written materials or documents, alone accounts for about 13% of all work-related ChatGPT usage. The top five together hold roughly a third:

Consumer ChatGPT, as observed at work, is overwhelmingly a writing, editing, research, and content-preparation engine. That shows up directly in the occupation-level scores. The highest-exposure occupations under this source are editors, technical writers, creative writers, and instructional coordinators. The lowest are machine operators, cooks, laundry workers, and similar hands-on roles, where essentially none of the job's work overlaps with what anyone does in a chat window.

What actually moved

We compare every quarterly snapshot of the model against the previous one. Between the June snapshot (before Signals) and the current one (after), across all 1,016 detailed occupations:

• 89% of occupations kept the exact same resilience label. 6% moved up a label, 5% moved down.
• The average occupation's top-line resilience score moved by only 1.6 percentage points in either direction, and not a single occupation moved more than 10.

The occupations that gained the most are the ones Signals scores at very low exposure: machine operators, postal mail sorters, security screeners, bartenders. The count of occupations labeled Vulnerable actually decreased from 34 to 23 in this update, mostly due to manufacturing and processing roles moving up a resilience label.

And the occupations that dropped the most are almost all in one field.

The teacher question

Signals ranks teaching occupations at roughly the 94th percentile of AI exposure. Our other four sources put them near the 52nd (very close to the middle). That is the single largest disagreement between this source and the rest of our model. In this update, teaching occupations saw the largest resilience declines in the report, with K-12 teaching occupations, in particular, dropping about 5 percentage points on average.

But an exposure score answers a specific question, and that question is not “will teaching disappear?” Signals measures how much of a job's work overlaps with what people do in consumer ChatGPT. No exposure source, this one included, measures whether a job is likely to erode. That is why exposure is one input to the resilience score rather than the score itself, and why the model's other dimensions continue to rank teaching occupations as at least somewhat resilient: the human core of the job is, for various reasons, hard to automate.

We tested several ways to reduce the divergence before shipping: reweighting, blending in other academic exposure datasets, compressing the heaviest-usage activities. Each one either double-counted data we already use, distorted the rest of the model, or eroded the independence that makes Signals worth adding. The signal is a faithful reflection of what people do in consumer ChatGPT, so we shipped it as-is.

Also worth knowing: what this source can't see

Every dataset has blind spots, and Signals' main one is the flip side of its strength. The public dataset covers consumer ChatGPT only: personal accounts on the free, Go, Plus, and Pro plans. It does not include enterprise or business workspaces, the API, or coding tools like Codex. That means it likely understates AI penetration in fields where the work happens in professional tools, like software, law, and finance. Our Microsoft and Anthropic sources cover some of that ground, which is a good illustration of why the model uses multiple sources.

What about the enterprise side? OpenAI does measure it, in a companion release called B2B Signals, introduced in May 2026. But it is a fundamentally different kind of release. B2B Signals reports firm-level aggregates: how the most AI-intensive “frontier” firms compare to typical firms, which tools show the biggest adoption gaps, and how usage breaks down across a handful of broad task categories by business function. What it does not offer is the thing our pipeline needs: usage broken down by detailed work activity, published as downloadable data. There is nothing in it we can join to occupations the way we did with the consumer release. If OpenAI ever publishes enterprise usage at the work-activity level, we will evaluate it exactly the way we evaluated this source. (For what it's worth, the aggregate enterprise picture rhymes with the consumer one: writing and communication is the most common use there too.)

One more wrinkle connects to the teacher discussion above. ChatGPT for Teachers, the dedicated free workspace for verified US K-12 educators, is a separate platform, and it is not one of the four consumer plans that the public dataset samples. So it is invisible to our score: whatever teacher usage Signals captures comes from ordinary consumer accounts, which means the data may understate AI use in and around classrooms.

What this means if you're using the report

If you checked an occupation last quarter, odds are about 9 in 10 its label is the same today. What did change:

• Every occupation page's sources list now shows “OpenAI Signals” alongside the other exposure sources, with its own score, wherever the occupation has one (that's about three-quarters of detailed occupations; the rest continue to lean on the other sources).
• A handful of occupations moved one label, in both directions, with teaching occupations moving down and several production and manual occupations moving up.
• The methodology page has been updated to reflect the five-source exposure model.

What's next

OpenAI has said Signals will continue to be updated, and our pipeline is built to refresh with it on our quarterly cycle. We are also watching how this source's picture of usage evolves: the current data reflects consumer ChatGPT through March 2026, and as usage patterns shift, so will the scores.

If you are using the report with students, in advising work, or for your own career thinking, we want to hear what's working and what isn't. Drop us a line at air@careervillage.org.