Graduation Disruption

A data based look at changes in career opportunities in the light of AI

Published

October 24, 2026

import {make_exposure_hist} from "./components/exposure_hist.js";

make_exposure_hist(degree_data, {
  width, height: width / 3, legend: false, animate: true,
  yAxis: null, xLabel: null, marginBottom: 28,
  title: html`<div class="splash-plot-title">Histogram of Majors by GPT-4 Beta<br>Shaded by Employment Δ</div>`
})

Just last month, the US Census published Graduating into Disruption (Orr et al. 2026), which explores labor market outcomes for AI-exposed college majors. A number of papers have previously explored the impact of AI on professions. As far as I know, though, this is the first longitudinal study of the impact of AI on early employment outcomes by university major.

As my own university considers its curriculum, I thought I’d promote this paper with my own interactive look at its main data set. While this post mostly just highlights the work of the Census, I also consider to what extent AI exposure has affected raw employment outcomes - as opposed to just changes in those outcomes.

One notable observation is that, while computer related majors have seen negative impact, they still earn strong wages after graduation compared to most other majors.

AI exposure data

The term AI exposure generally means to what extent can a task be performed more efficiently using AI. That idea can be easily extending to occupations by considering what types of tasks an occupation performs. It can be further extended to college majors by considering the types of occupations those majors generally go into.

There are several ways quantify AI exposure. The key measures used by Graduating into Disruption, as described by the paper itself, are

GPT-4 Beta (Eloundou et al. 2024): accounts for whether the length of time it takes to complete a task can be cut in half by AI.
Core task exposure (Eisfeldt et al. 2023): is also task focused but concentrates on tasks considered in the “core” of the occupation,
Observed task usage (Handa et al. 2025): uses Anthropic data to measure how much Claude.ai is actually used for various tasks, and
Share in top exposure quintile (Orr et al. 2026): measures the proportion of graduates within a major who work in in the top 20% of AI exposed occupations.

The visualizations below allow you to select which AI exposure measure you’d like to see. All AI exposures are reported as number of standard deviations from the mean. For example a GPT-4 Beta score of 1.57 for Applied Mathematics means that group’s score is 1.57 standard deviations past the mean for all majors.

A key question, of course, is how these AI exposure values affect employment. We can compare against one of two employment measures:

Employment: percentage of graduates who’ve found work and
Log earnings: logarithm of the average annual earnings of graduates.

Exposure and outcome changes

The key point behind the paper is to study how AI exposure has impacted employment outcomes using the measures described above. The figure below is effectively an interactive version of figures 10 and 11 from the paper. Each point corresponds to a major; its position and shading can be interpreted as follows:

the horizontal position is AI exposure value as described in the previous section,
the vertical position the percentage change in the outcome from 2022 to 2024, and
the shading is determined by the number of graduates in the major over the years 2018 -2024.

This version allows you to explore more variations, though, through the radio selectors. In addition, you can hover over a dot to get more information or search for a major to find it.

import {make_exposure_scatter} from "./components/exposure_scatter.js";
import {exposure_measures, make_major_search} from "./components/exposure_hist.js";

degree_data = FileAttachment("./data/ai_exposure_by_field_of_degree.csv").csv({typed: true})

viewof scatter_exposure = Inputs.radio(exposure_measures, {
  label: "AI exposure",
  value: "GPT4Beta"
})

viewof scatter_outcome = Inputs.radio(new Map([
  ["Employment Δ", "EmploymentCoefficient2024"],
  ["Log earnings Δ", "LogEarningsCoefficient2024"]
]), {
  label: "Outcome",
  value: "EmploymentCoefficient2024"
})

viewof scatter_search = make_major_search(degree_data, "scatter-degree-field-suggestions", {showLabel: false})

html`<div style="height: 20px"></div>`

make_exposure_scatter(degree_data, {
  exposure: scatter_exposure, outcome: scatter_outcome,
  search: scatter_search, width
})

Note that, in every choice of selector configuration, increased AI exposure generally implies a negative effect on employment outcomes. You can hover over the dots in the lower right to find which majors have been impacted the most. Not surprisingly, many subfields of computer science and engineering have a lot of AI exposure.

Every field that’s more than one standard deviation past the GPT-4 Beta mean, though, has had negative employment changes.

Not all fields with negative employment outcomes have been unduly impacted by AI - Intercultural and International Studies, for example. Perhaps, there is something else impacting that field?

Earnings and AI exposure

Graduating into Disruption focuses only on change in employment and earnings; it does not explore current values. Thus, it’s possible that a degree like computer science might still offer high wages relative to other majors, even if they’ve come down some. Unfortunately, the authors don’t offer any data on that so I thought I’d try to find some of my own.

The scatter plot below is based on data from the Federal Reserve (Federal Reserve Bank of New York 2026), in addition to the data from Graduating into Disruption. The Fed provides data on early career median earnings by major. The Fed data and the Census data do not match perfectly but there is significant overlap between labels. Thus, we’re able to plot median pay vs AI exposure for 2024, as shown below.

import {make_earnings_exposure} from "./components/earnings_exposure.js";
import {exposure_measures as earnings_exposure_choices} from "./components/exposure_hist.js";

matched_major_data = FileAttachment("./data/fed_exposure_matched_majors.csv").csv({typed: true})

viewof earnings_exposure_measure = Inputs.radio(earnings_exposure_choices, {
  label: "AI exposure",
  value: "GPT4Beta"
})

viewof earnings_major_search = make_major_search(matched_major_data, "earnings-major-suggestions", {
  showLabel: false, nameKey: "FedMajor"
})

html`<div style="height: 20px"></div>`

make_earnings_exposure(matched_major_data, {
  exposure: earnings_exposure_measure, search: earnings_major_search, width
})

One thing we see quite clearly for this group of professions is that increased AI exposure generally comes with higher pay.

References

Eisfeldt, Andrea L., Gregor Schubert, and Miao Ben Zhang. 2023. Generative AI and Firm Values. Working Paper No. 31222. National Bureau of Economic Research. https://www.nber.org/system/files/working_papers/w31222/revisions/w31222.rev0.pdf.

Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. 2024. “GPTs Are GPTs: Labor Market Impact Potential of LLMs.” Science 384 (6702): 1306–8. https://doi.org/10.1126/science.adj0998.

Federal Reserve Bank of New York. 2026. The Labor Market for Recent College Graduates. https://www.newyorkfed.org/research/college-labor-market.

Handa, Kunal, Alex Tamkin, Miles McCain, et al. 2025. Which Economic Tasks Are Performed with AI? Evidence from Millions of Claude Conversations. arXiv. https://doi.org/10.48550/arXiv.2503.04761.

Orr, Cody, Lee C. Tucker, and Lawrence Warren. 2026. Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors. Working Paper CES-26-56. U.S. Census Bureau, Center for Economic Studies. https://www2.census.gov/library/working-papers/2026/adrm/ces/CES-WP-26-56.pdf.

Comments

Anyone with a BlueSky account can leave comments on this site. Just hit the Reply on Bluesky button and post on that BlueSky discussion.