Skip to main content

Which College Majors Are Safest From AI (and Do They Pay)?

This week we’re looking at which college majors are safest from AI, and whether the safe ones actually pay well.

The dataset joins three public sources at the level of the college major. Starting salaries come from the U.S. Department of Education’s College Scorecard field-of-study files, using median earnings one year after graduation for bachelor’s degrees. AI disruption risk comes from the AI Occupational Exposure (AIOE) index of Felten, Raj and Seamans (2021), which scores how much each occupation’s required abilities overlap with what AI can do. It also includes separate scores for language tasks and image tasks. To connect the two, I used the NCES/BLS CIP-to-SOC crosswalk, which maps fields of study to the occupations their graduates tend to enter.

The result is 285 majors, each with a starting salary and an AI exposure score. What stands out is that the two barely relate to each other (r is about 0.01). Hands-on technical fields like nuclear, marine, and aerospace engineering technology show low exposure and high pay. Registered nursing sits near the 20th percentile of AI exposure and pays around $75,000 to start. On the other end, several humanities majors are both highly exposed and among the lowest paid.

One thing to know about the exposure score: the AIOE is standardized to a mean of 0 and a standard deviation of 1. A value of 0 is an average occupation, and negative values fall below average. High exposure does not necessarily mean a job will be automated away, since AI can also assist the work rather than replace it. So that you don’t need a stats background to read it, the main table also includes ai_exposure_percentile, a plain 0 to 100 rank across all majors.

The AIOE links ten common AI applications with fifty-two distinct occupational abilities used by O*NET and uses the relative prevalence of these abilities within an occupation to generate an occupational measure of AI exposure. (Felten, Raj and Seamans 2021; description rephrased for compliance with licensing restrictions.)

Some questions to explore:

  • Which majors are both safe from AI and well paid, and which are both exposed and low paid?
  • Does the picture change when you look at language-task exposure versus image-task exposure?
  • Do broad fields like Engineering, Health Professions, Business, and Humanities cluster together?
  • The AIOE measures exposure, not automation. How would the story change if high exposure mostly meant AI assisting the work rather than replacing it?

Thank you to Tony Galvan, Golden Dome Data Science for curating this week’s dataset.

The Data

# Using R
# Option 1: tidytuesdayR R package 
## install.packages("tidytuesdayR")

tuesdata <- tidytuesdayR::tt_load('2026-10-20')
## OR
tuesdata <- tidytuesdayR::tt_load(2026, week = 42)

ai_salary_majors <- tuesdata$ai_salary_majors
major_occupations <- tuesdata$major_occupations

# Option 2: Read directly from GitHub

ai_salary_majors <- readr::read_csv('https://raw.githubusercontent.com/rfordatascience/tidytuesday/main/data/2026/2026-10-20/ai_salary_majors.csv')
major_occupations <- readr::read_csv('https://raw.githubusercontent.com/rfordatascience/tidytuesday/main/data/2026/2026-10-20/major_occupations.csv')
# Using Python
# Option 1: pydytuesday python library
## pip install pydytuesday

import pydytuesday

# Download files from the week, which you can then read in locally
pydytuesday.get_date('2026-10-20')

# Option 2: Read directly from GitHub and assign to an object

ai_salary_majors = pandas.read_csv('https://raw.githubusercontent.com/rfordatascience/tidytuesday/main/data/2026/2026-10-20/ai_salary_majors.csv')
major_occupations = pandas.read_csv('https://raw.githubusercontent.com/rfordatascience/tidytuesday/main/data/2026/2026-10-20/major_occupations.csv')
# Using Julia
# Option 1: TidierTuesday.jl library
## Pkg.add(url="https://github.com/TidierOrg/TidierTuesday.jl")

using TidierTuesday

# Download datasets for the week, and load them as a NamedTuple of DataFrames
data = tt_load("2026-10-20")

# Option 2: Read directly from GitHub and assign to an object with TidierFiles

ai_salary_majors = read_csv("https://raw.githubusercontent.com/rfordatascience/tidytuesday/main/data/2026/2026-10-20/ai_salary_majors.csv")
major_occupations = read_csv("https://raw.githubusercontent.com/rfordatascience/tidytuesday/main/data/2026/2026-10-20/major_occupations.csv")

# Option 3: Read directly from Github and assign without Tidier dependencies
ai_salary_majors = CSV.read("https://raw.githubusercontent.com/rfordatascience/tidytuesday/main/data/2026/2026-10-20/ai_salary_majors.csv", DataFrame)
major_occupations = CSV.read("https://raw.githubusercontent.com/rfordatascience/tidytuesday/main/data/2026/2026-10-20/major_occupations.csv", DataFrame)

How to Participate

  • Explore the data, watching out for interesting relationships. We would like to emphasize that you should not draw conclusions about causation in the data. There are various moderating variables that affect all data, many of which might not have been captured in these datasets. As such, our suggestion is to use the data provided to practice your data tidying and plotting techniques, and to consider for yourself what nuances might underlie these relationships.
  • Create a visualization, a model, a Quarto report, a shiny app, or some other piece of data-science-related output, using R, Python, or another programming language.
  • Share your output and the code used to generate it on social media with the #TidyTuesday hashtag.
  • Submit your own dataset!

PydyTuesday: A Posit collaboration with TidyTuesday

  • Exploring the TidyTuesday data in Python? Posit has some extra resources for you! Have you tried making a Quarto dashboard? Find videos and other resources in Posit’s PydyTuesday repo.
  • Share your work with the world using the hashtags #TidyTuesday and #PydyTuesday so that Posit has the chance to highlight your work, too!
  • Deploy or share your work however you want! If you’d like a super easy way to publish your work, give Connect Cloud a try.

Data Dictionary

ai_salary_majors.csv

variable class description
cip4 character Four-digit Classification of Instructional Programs (CIP) 2020 code identifying the field of study.
field_of_study character Human-readable name of the college major (from the College Scorecard CIP description).
broad_field character Official NCES title of the two-digit CIP family the major belongs to (from the CIP 2020 code file), title-cased with the trailing period removed. For example, CIP family 52 is “Business, Management, Marketing, and Related Support Services”.
broad_field_short character A short, chart-friendly label for the CIP family, keyed to the two-digit family code. This is a convenience label for plotting and grouping, not an official NCES title (e.g. the family above is shortened to “Business”).
median_starting_salary double Median earnings in US dollars one year after graduation, pooled across institutions offering the major (bachelor’s degree graduates). This is the starting-salary measure.
median_salary_4yr double Median earnings in US dollars four years after graduation, pooled across institutions. Missing for a few majors without four-year follow-up data.
ai_exposure double Average AI Occupational Exposure (AIOE) score across the occupations this major feeds into. The AIOE is standardized (mean 0, standard deviation 1); higher values mean more of the occupation’s abilities overlap with what AI can do.
ai_exposure_percentile double Rank of this major’s AI exposure on a 0 to 100 scale relative to all 285 majors in the dataset. 0 is the least AI-exposed major, 100 the most.
ai_exposure_language double Average AIOE score for the language-modeling application of generative AI (text-based tasks) across the major’s occupations. Standardized like ai_exposure.
ai_exposure_image double Average AIOE score for the image-generation application of generative AI (visual tasks) across the major’s occupations. Standardized like ai_exposure.
n_occupations integer Number of distinct occupations (SOC codes) the major maps to in the CIP-SOC crosswalk; the exposure scores are averaged over these occupations.
n_institutions integer Number of institutions whose graduates contributed to the pooled starting-salary figure for this major.
n_graduates double Total count of graduates with measured earnings one year out, summed across institutions offering the major.

major_occupations.csv

variable class description
cip4 character Four-digit CIP 2020 code for the field of study; joins to cip4 in the ai_salary_majors table.
field_of_study character Name of the college major from the CIP-SOC crosswalk.
soc character Six-digit Standard Occupational Classification (SOC) 2018 code for an occupation this major can lead to.
occupation character Human-readable occupation title for the SOC code.
ai_exposure double Overall AI Occupational Exposure (AIOE) score for this occupation. Standardized (mean 0, SD 1); higher means more exposed to AI.
ai_exposure_language double AIOE score for the language-modeling application of generative AI for this occupation.
ai_exposure_image double AIOE score for the image-generation application of generative AI for this occupation.

Cleaning Script

# =============================================================================
# TidyTuesday dataset: College majors, AI disruption risk, and starting salaries
# =============================================================================
#
# This script is fully self-contained. It downloads three public data sources,
# joins them at the level of the college major (4-digit CIP code), and produces
# two tidy data frames:
#
#   1. ai_salary_majors   - one row per major: starting salary + AI exposure
#   2. major_occupations  - the major -> occupation crosswalk with per-occupation
#                           AI exposure scores (the detail behind each major's
#                           exposure)
#
# SOURCES
# -------
# A. AI Occupational Exposure (AIOE), Felten, Raj & Seamans (2021), Strategic
#    Management Journal 42(12):2195-2217. Occupation-level (SOC) exposure to AI,
#    plus generative-AI variants for language modeling and image generation.
#    https://github.com/AIOE-Data/AIOE
#
# B. U.S. Department of Education, College Scorecard, Field of Study data files
#    (updated 2026-06-10). Median earnings one year after graduation, by
#    institution x 4-digit CIP code x credential level.
#    https://collegescorecard.ed.gov/data/
#
# C. NCES / BLS CIP (2020) -> SOC (2018) Crosswalk. Maps fields of study to the
#    occupations their graduates enter.
#    https://nces.ed.gov/ipeds/cipcode/resources.aspx
#
# D. NCES Classification of Instructional Programs (CIP) 2020 code file. Supplies
#    the official title for each 2-digit CIP family, used for the broad_field
#    grouping.
#    https://nces.ed.gov/ipeds/cipcode/resources.aspx
# =============================================================================

library(readr)
library(readxl)
library(dplyr)
library(stringr)
library(tidyr)

# -----------------------------------------------------------------------------
# 0. Download raw files to a temporary directory
# -----------------------------------------------------------------------------
tmp <- tempdir()

aioe_url <- "https://github.com/AIOE-Data/AIOE/raw/main/AIOE_DataAppendix.xlsx"
lm_url   <- "https://github.com/AIOE-Data/AIOE/raw/main/Language%20Modeling%20AIOE%20and%20AIIE.xlsx"
ig_url   <- "https://github.com/AIOE-Data/AIOE/raw/main/Image%20Generation%20AIOE%20and%20AIIE.xlsx"
xw_url   <- "https://nces.ed.gov/ipeds/cipcode/Files/CIP2020_SOC2018_Crosswalk.xlsx"
cip_url  <- "https://nces.ed.gov/ipeds/cipcode/Files/CIPCode2020.csv"
sc_url   <- paste0("https://ed-public-download.scorecard.network/downloads/",
                   "Most-Recent-Cohorts-Field-of-Study_06102026.zip")

aioe_file <- file.path(tmp, "AIOE_DataAppendix.xlsx")
lm_file   <- file.path(tmp, "LanguageModeling_AIOE.xlsx")
ig_file   <- file.path(tmp, "ImageGeneration_AIOE.xlsx")
xw_file   <- file.path(tmp, "CIP2020_SOC2018_Crosswalk.xlsx")
cip_file  <- file.path(tmp, "CIPCode2020.csv")
sc_zip    <- file.path(tmp, "fos.zip")

download.file(aioe_url, aioe_file, mode = "wb", quiet = TRUE)
download.file(lm_url,   lm_file,   mode = "wb", quiet = TRUE)
download.file(ig_url,   ig_file,   mode = "wb", quiet = TRUE)
download.file(xw_url,   xw_file,   mode = "wb", quiet = TRUE)
download.file(cip_url,  cip_file,  mode = "wb", quiet = TRUE)
download.file(sc_url,   sc_zip,    mode = "wb", quiet = TRUE)

sc_csv <- unzip(sc_zip, exdir = file.path(tmp, "fos"))
sc_csv <- sc_csv[grepl("Field-of-Study\\.csv$", sc_csv)][1]

# -----------------------------------------------------------------------------
# 1. AI exposure by occupation (SOC)
#    AIOE scores are standardized (mean 0, SD 1). Higher = more exposed to AI.
# -----------------------------------------------------------------------------
aioe <- read_excel(aioe_file, sheet = "Appendix A") |>
  rename(soc = `SOC Code`, occupation = `Occupation Title`, aioe = AIOE) |>
  select(soc, occupation, aioe)

lm <- read_excel(lm_file, sheet = "LM AIOE") |>
  rename(soc = `SOC Code`, aioe_language = `Language Modeling AIOE`) |>
  select(soc, aioe_language)

ig <- read_excel(ig_file, sheet = "IG AIOE") |>
  rename(soc = `SOC Code`, aioe_image = `Image Generation AIOE`) |>
  select(soc, aioe_image)

exposure <- aioe |>
  left_join(lm, by = "soc") |>
  left_join(ig, by = "soc")

# -----------------------------------------------------------------------------
# 2. CIP (2020) -> SOC (2018) crosswalk
#    A major maps to one or more occupations. Drop the "NO MATCH" sentinel.
#    Normalize the dotted 6-digit CIP (e.g. "11.0701") to the 4-digit key
#    ("1107") used by the College Scorecard field-of-study file.
# -----------------------------------------------------------------------------
crosswalk <- read_excel(xw_file, sheet = "CIP-SOC") |>
  rename(cip6 = CIP2020Code, cip_title = CIP2020Title,
         soc  = SOC2018Code, soc_title = SOC2018Title) |>
  filter(soc != "99-9999") |>
  mutate(cip4 = str_sub(str_remove(cip6, fixed(".")), 1, 4)) |>
  distinct(cip4, soc, .keep_all = TRUE)

# -----------------------------------------------------------------------------
# 2b. CIP family (2-digit) labels, from the official NCES CIP 2020 code file
#     The CSV wraps every value as an Excel text guard (e.g. ="01"), so strip
#     that first. The family-header row is the one whose CIPCode equals its
#     CIPFamily. broad_field is the official NCES title (title-cased, trailing
#     period removed). broad_field_short is a chart-friendly label I maintain by
#     hand, keyed to the family code; it is a convenience, not an NCES label.
# -----------------------------------------------------------------------------
unguard <- function(x) str_replace_all(x, '^="?|"$', "")

cip_families <- read_csv(cip_file, col_types = cols(.default = col_character())) |>
  transmute(fam = unguard(CIPFamily), code = unguard(CIPCode), title = CIPTitle) |>
  filter(code == fam) |>
  distinct(fam, title) |>
  mutate(broad_field = title |>
           str_remove("\\.$") |>
           str_to_title() |>
           str_replace_all("\\bAnd\\b", "and") |>
           str_replace_all("\\bOf\\b", "of")) |>
  select(fam, broad_field)

broad_field_short <- tribble(
  ~fam, ~broad_field_short,
  "01", "Agriculture",              "03", "Natural Resources",
  "04", "Architecture",             "05", "Area & Ethnic Studies",
  "09", "Communication",            "10", "Communications Tech",
  "11", "Computer Science",         "12", "Culinary & Personal Svc",
  "13", "Education",                "14", "Engineering",
  "15", "Engineering Tech",         "16", "Foreign Languages",
  "19", "Family & Consumer Sci",    "22", "Legal",
  "23", "English",                  "24", "Liberal Arts",
  "25", "Library Science",          "26", "Biological Sciences",
  "27", "Mathematics & Statistics", "28", "Military Science",
  "29", "Military Tech",            "30", "Interdisciplinary",
  "31", "Parks & Recreation",       "38", "Philosophy & Religion",
  "39", "Theology & Religion",      "40", "Physical Sciences",
  "41", "Science Tech",             "42", "Psychology",
  "43", "Homeland Security",        "44", "Public Administration",
  "45", "Social Sciences",          "46", "Construction Trades",
  "47", "Mechanic & Repair Tech",   "48", "Precision Production",
  "49", "Transportation",           "50", "Visual & Performing Arts",
  "51", "Health Professions",       "52", "Business",
  "54", "History"
)

cip_families <- cip_families |> left_join(broad_field_short, by = "fam")

# -----------------------------------------------------------------------------
# 3. Starting salary by major (College Scorecard, bachelor's degrees)
#    CREDLEV == 3 is Bachelor's Degree. EARN_MDN_1YR is median earnings one year
#    after completion. Aggregate institution-level values to a national median
#    per 4-digit CIP. "PrivacySuppressed" cells are treated as missing.
# -----------------------------------------------------------------------------
scorecard <- read_csv(
  sc_csv,
  col_types = cols(.default = col_character()),
  na = c("", "NA", "NULL", "PrivacySuppressed")
) |>
  filter(CREDLEV == "3") |>
  transmute(
    cip4        = str_pad(CIPCODE, 4, "left", "0"),
    cip_desc    = str_squish(str_remove(CIPDESC, "\\.$")),
    earn_1yr    = suppressWarnings(as.numeric(EARN_MDN_1YR)),
    earn_4yr    = suppressWarnings(as.numeric(EARN_MDN_4YR)),
    grad_count  = suppressWarnings(as.numeric(EARN_COUNT_WNE_1YR))
  )

salary <- scorecard |>
  filter(!is.na(earn_1yr)) |>
  group_by(cip4) |>
  summarise(
    field_of_study         = first(cip_desc),
    median_starting_salary = round(median(earn_1yr, na.rm = TRUE)),
    median_salary_4yr      = round(median(earn_4yr, na.rm = TRUE)),
    n_institutions         = n(),
    n_graduates            = sum(grad_count, na.rm = TRUE),
    .groups = "drop"
  )

# -----------------------------------------------------------------------------
# 4. Major-level AI exposure
#    Average the occupation-level AIOE scores across every occupation a major
#    maps to. Attach the standardized scores to the salary table.
# -----------------------------------------------------------------------------
major_exposure <- crosswalk |>
  inner_join(exposure, by = "soc") |>
  group_by(cip4) |>
  summarise(
    n_occupations    = n_distinct(soc),
    ai_exposure       = mean(aioe, na.rm = TRUE),
    ai_exposure_language = mean(aioe_language, na.rm = TRUE),
    ai_exposure_image    = mean(aioe_image, na.rm = TRUE),
    .groups = "drop"
  )

# -----------------------------------------------------------------------------
# 5. Final major-level dataset
#    Keep majors that have BOTH a starting salary and an AI-exposure score.
#    Add a 0-100 percentile rank for AI exposure so a non-specialist can read
#    "how exposed is this major relative to all others" without needing to know
#    that the raw score is a z-score.
# -----------------------------------------------------------------------------
ai_salary_majors <- salary |>
  inner_join(major_exposure, by = "cip4") |>
  mutate(fam = str_sub(cip4, 1, 2)) |>
  left_join(cip_families, by = "fam") |>
  mutate(
    ai_exposure_percentile = round(100 * (rank(ai_exposure) - 1) / (n() - 1), 1)
  ) |>
  select(
    cip4, field_of_study, broad_field, broad_field_short,
    median_starting_salary, median_salary_4yr,
    ai_exposure, ai_exposure_percentile,
    ai_exposure_language, ai_exposure_image,
    n_occupations, n_institutions, n_graduates
  ) |>
  arrange(ai_exposure)

# Fail loudly if any major did not match an NCES CIP family label, rather than
# silently emitting NA or an "Other" bucket. This guards future data refreshes.
if (anyNA(ai_salary_majors$broad_field) ||
    anyNA(ai_salary_majors$broad_field_short)) {
  missing <- ai_salary_majors |>
    filter(is.na(broad_field) | is.na(broad_field_short)) |>
    distinct(fam = str_sub(cip4, 1, 2))
  stop("Unmapped CIP family codes (add them to the lookups): ",
       paste(missing$fam, collapse = ", "))
}

# -----------------------------------------------------------------------------
# 6. Supporting dataset: which occupations drive each major's exposure
# -----------------------------------------------------------------------------
major_occupations <- crosswalk |>
  inner_join(exposure, by = "soc") |>
  semi_join(ai_salary_majors, by = "cip4") |>
  transmute(
    cip4,
    field_of_study      = str_squish(str_remove(cip_title, "\\.$")),
    soc,
    occupation,
    ai_exposure          = round(aioe, 3),
    ai_exposure_language = round(aioe_language, 3),
    ai_exposure_image    = round(aioe_image, 3)
  ) |>
  arrange(cip4, desc(ai_exposure))

# round the standardized scores in the main table for readability
ai_salary_majors <- ai_salary_majors |>
  mutate(across(c(ai_exposure, ai_exposure_language, ai_exposure_image),
                \(x) round(x, 3)))