Adding Google ATLAS to Our Model

Q4 2026 update · November 4, 2026 · 7 min read · CareerVillage AIR Team

Key takeaways

  • We added a sixth source to the Meaningful Human Contribution (MHC) subscore of the AI Resilience Report: Google’s AI & Economy ATLAS, built from 14.7 million Gemini interactions in April 2026. It measures how much of each occupation’s work people already bring to Gemini.
  • ATLAS is the broadest usage-based source we have: it scores 84% of detailed occupations, compared with 75% for OpenAI Signals and 47% for Anthropic. 304 occupations now have all six exposure sources.
  • It carries new information (a 0.36 correlation with the average of the other five), and it has a distinct point of view: ATLAS reads office and administrative work as the most AI-saturated, and hands-on work as the least.
  • 93% of occupations kept the exact same resilience label. The average score moved 1.1 points, and no occupation moved more than 9. Teaching occupations recovered some of the ground they lost in August. Nursing and the skilled trades slipped a little, and we explain why below.

When we added OpenAI Signals in August, we made the case that usage data answers a different question from a feasibility model: not what AI could do to a job, but what people are already doing with it. ATLAS is the newest public dataset of that kind, this time from Gemini rather than ChatGPT, and it is now the fourth usage-grounded source in the model alongside Microsoft, Anthropic, and OpenAI.

What Google ATLAS is

ATLAS (Activity, Task, Landscape, and Adoption Study) is Google’s ongoing study of how people use its AI products. The first release, published in July 2026 with a research paper and a public dataset, draws on 14.7 million de-identified interactions from the Gemini app, AI Mode, and the Gemini API, collected April 6 to 19, 2026. Each work-related interaction is classified to a task in O*NET, the Department of Labor’s occupational database, across more than 800 occupations and 4,000 tasks.

Google’s headline is that AI use at work is “broad but shallow.” Work-related use shows up in occupations covering roughly nine-tenths of U.S. employment, but in a typical job AI touches about 21% of tasks, and fewer than 10% of work interactions fully automate a task. Most are collaborative: ideation, information retrieval, and learning.

The figure we use is task saturation: the share of an occupation’s O*NET tasks that saw non-negligible Gemini use during the window. An occupation where Gemini touched 4 of its 40 tasks scores 10%. As with every exposure source, we convert that to a percentile across occupations, where a higher percentile means more of the work overlaps with what people do in Gemini. The dataset is licensed CC BY 4.0, and every career page that shows the score credits Google.

One value per occupation group

ATLAS has one limitation the other sources don’t. Google publishes occupation figures for broad occupation groups, not individual careers: 375 groups in all. “Registered Nurses” is one group, so registered nurses, critical care nurses, and clinical nurse specialists all start with the same ATLAS value. In our catalog, 183 groups contain a single career and 658 careers share their group’s value with at least one other.

Rather than treat a shared value as if it were a precise one, we weight it by how well it fits. Where the other five sources already show that the careers in a group differ a lot from one another, the shared ATLAS value is a worse description of any one of them, and it counts for less. Where the other sources agree that the careers are alike, ATLAS counts in full. In practice the median weight in multi-career groups is 0.69, and it never drops below about a third. Careers in single-career groups get the full weight. The formula is on the methodology page under Broad-Level Source Weighting.

Two kinds of careers get no ATLAS score at all: “All Other” catch-all occupations, which Google excluded when computing group figures, and small groups Google suppressed for privacy. Those lean on the other five sources, as before.

Career coverage by AI exposure source

CV (internal model)model
100%
WRTMJfeasibility model
91.3%
Google ATLASobserved usage
84.4%
Microsoftobserved usage
77.5%
OpenAI Signalsobserved usage
75.2%
Anthropicobserved usage
46.5%

Percentage of the 997 detailed careers in the report with a valid score from each source. Model-based sources can assign a score to every career by construction; usage-based sources only score careers whose work is visible in the usage data.

What the data says: ATLAS is an office-work detector

Signals told us ChatGPT is a writing machine. ATLAS tells a related but different story. The occupations it rates most AI-saturated are the ones built on administrative, financial, and record-keeping tasks: office clerks, customer service representatives, procurement clerks, budget analysts, payroll clerks, and accountants and auditors. At the bottom are flight attendants, butchers and meat cutters, shoe and leather workers, and fence erectors, where no task cleared the bar.

That point of view is distinct from every other source. ATLAS correlates 0.46 with Anthropic, 0.39 with our internal model, 0.34 with Microsoft, 0.27 with OpenAI Signals, and slightly negatively (−0.09) with Will Robots Take My Job. Against the average of all five, it’s 0.36. That is the range where a source adds information instead of repeating it.

How Google ATLAS correlates with each existing source

Anthropicobserved usage
0.46
CV (internal model)model
0.39
Microsoftobserved usage
0.34
OpenAI Signalsobserved usage
0.27
WRTMJfeasibility model
−0.09

Spearman rank correlation between Google ATLAS’s exposure percentiles and each other source’s, computed on the occupations both sources score. A value near 1 means the two sources rank occupations nearly identically; near 0 means they carry independent information. ATLAS correlates 0.36 with the average of all five combined.

What actually moved

Between the October 7 snapshot (before ATLAS) and the November 4 snapshot (after), across the 997 detailed occupations:

Where 997 occupations landed after adding Google ATLAS

93%Stayed same927 occupations
3.2%Moved up32 occupations
3.8%Moved down38 occupations

Comparison of the October 7 and November 4, 2026 snapshots across all 997 detailed occupations in the report. “Stayed same” means the occupation kept the same one of six resilience tiers.

  • 93% kept the exact same resilience label. 3% moved up a label, 4% moved down.
  • The average top-line score moved 1.1 points in either direction. Fourteen occupations moved more than 5 points, and none moved more than 9.
  • The Vulnerable tier shrank from 11 occupations to 8, and the Highly Resilient tier from 3 to 1.

The biggest gainers are clerical roles where ATLAS disagrees sharply with everyone else. Switchboard operators sit between the 67th and 100th exposure percentile under the other five sources and at the 14th under ATLAS; they gained 8.9 points and left the Vulnerable tier. Brokerage clerks (+8.3) and meter readers (+5.6) follow the same pattern. The model averages its sources, so a lone dissent pulls these careers toward the middle rather than flipping them.

The biggest declines come from two places: nursing and the skilled trades.

The occupations that moved most when Google ATLAS was added

Became more resilient
Switchboard Operators+8.9 pp
Vulnerable→Not Very Resilient
Brokerage Clerks+8.3 pp
Not Very Resilient→Not Very Resilient
Hydrologic Technicians+6.7 pp
Not Very Resilient→Somewhat Resilient
Stone Cutters and Carvers+6.2 pp
Not Very Resilient→Somewhat Resilient
Meter Readers+5.6 pp
Vulnerable→Not Very Resilient
Gem and Diamond Workers+5.4 pp
Vulnerable→Not Very Resilient
Became less resilient
Advanced Practice Psychiatric Nurses-6.2 pp
Highly Resilient→Resilient
HVAC Mechanics and Installers-6.2 pp
Resilient→Mostly Resilient
Food Service Managers-6.2 pp
Resilient→Resilient
Paramedics-5.6 pp
Resilient→Mostly Resilient
Emergency Medical Technicians-5.3 pp
Resilient→Mostly Resilient
Critical Care Nurses-4.6 pp
Highly Resilient→Resilient

“pp” = percentage points on the top-line resilience score. The gainers are clerical roles that ATLAS rates far lower than every other source does; the declines are nursing specialties and skilled trades, where ATLAS rates exposure far higher.

The trades question

ATLAS ranks hands-on technical work far higher than any other source does. Electricians land at the 93rd exposure percentile under ATLAS and between the 8th and 71st under the other sources. Auto body repairers are at the 92nd, industrial machinery mechanics the 87th, HVAC mechanics the 81st, and paramedics the 69th.

Google reads this as real adoption: in its write-up, workers in manual and technical trades use conversational AI for live diagnostics, troubleshooting, and learning (interpreting test results, debugging wiring, inspecting machinery for wear), and they are twice as likely as other workers to use image and video. We think part of it is something else. The data cannot tell an electrician debugging a panel from a homeowner asking about theirs; both conversations get classified against electrician tasks. For the trades, ATLAS probably overstates how much of the job is being done with AI.

Nursing is the same shape. The Registered Nurses group sits at the 66th percentile under ATLAS and between the 4th and 54th under the other sources, and because the four nursing specialties share that group value with registered nurses, four of the ten largest declines in this update are nursing roles, each down 4 to 6 points.

We shipped it anyway, for the reason we gave with Signals: an exposure source is worth adding because it’s independent, and trimming it wherever it disagrees with the others would just make the model agree with itself. The averaging does the moderating. Electricians moved 3.9 points, from the low end of Resilient to the high end of Mostly Resilient. No occupation in the report moved more than one label.

Teachers, revisited

In August, Signals placed teaching occupations near the 94th exposure percentile, and teaching saw the largest declines in that update. ATLAS places them near the 27th. Across the 51 postsecondary and K-12 teaching occupations, the exposure subscore rose 4 points on average and the top-line score 1.3, and three teaching occupations moved up a label while none moved down. Two usage datasets, two different answers about the same jobs: that disagreement is exactly why the model averages six sources instead of trusting one.

What this source can’t see

ATLAS covers Google’s consumer AI surfaces only: the Gemini app, AI Mode, and the free API tier. It does not include Gemini inside Google Workspace, enterprise Gemini, AI Overviews in Search, or agentic coding tools, so, like Signals, it likely understates AI use in fields where the work happens inside professional tools. The occupation figures are global rather than U.S.-only, and they come from a two-week window. Google describes v1.0 as an early view of a fast-moving picture and plans further releases; our pipeline refreshes with them on our quarterly cycle.

What this means if you’re using the report

  • Every career page’s sources list now shows “Google ATLAS” with its own score wherever the occupation has one.
  • Confidence levels were recalibrated for a six-source model. ATLAS counts as a full source for confidence, so occupations that picked it up as a fourth or fifth source gained about a point of confidence, and the roughly 120 occupations still scored by three or fewer sources lost about as much.
  • The methodology page describes the six-source exposure model and the group-level weighting. The public API (v2.3) and dataset download include the ATLAS score and its weight, and the historical trends chart tracks it from this snapshot forward.

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.

The AI Resilience Report is a project from CareerVillage.org®, a registered 501(c)(3) nonprofit.

Built with ❤️ by Sandbox Web

The AI Resilience Report is governed by CareerVillage.org’s Privacy Policy and Terms of Service. This site is not affiliated with Anthropic, Microsoft, or any other data provider and doesn't necessarily represent their viewpoints. This site is being actively updated, and may sometimes contain errors or require improvement in wording or data. To report an error or request a change, please contact air@careervillage.org.