Jott2121/ai-career-threat-index

AI Career Threat Index

An open dataset scoring 300 US professions on AI displacement risk. Every role carries four published sub-scores — task automation potential, AI tool maturity, employer adoption, and agentic exposure (exposure to autonomous multi-step agents, distinct from copilot-style tools) — with a one-sentence written rationale per factor, combined by an open formula into a single 0–100 score (higher = more exposed). MIT licensed, reviewed quarterly, versioned on GitHub.

Most AI-and-jobs discussion is anecdotal. This dataset gives one numeric answer per role and publishes the factor judgments behind it so they can be argued with — for journalists, researchers, career tooling, and workforce analysis. To our knowledge the agenticExposure factor is the first published per-occupation scoring of exposure to autonomous agents specifically.

Files

• —ai-career-threat-index.csv — flat table, one row per role (loads as the default config)
• —ai-career-threat-index.json — full structured records: sub-score rationales, task lists, defense skills, industry modifiers, quarterly history
• —soc-crosswalk.csv — every BLS SOC 2018 detailed occupation (867 codes) mapped to its nearest scored role with a match-quality label

Columns (CSV, one row per role)

How the score is built (methodology v2, Q3 2026+)

score = taskAutomation × (0.45 + 0.30·toolMaturity/100 + 0.25·adoption/100) + 0.10·agenticExposure

Task automation potential sets the ceiling (you can't be displaced from work AI can't do); tool maturity and employer adoption determine how much of that ceiling is realized today; agentic exposure adds forward pressure from autonomous agents. Full rubric with scoring anchors and calibration examples: pipeline/RUBRIC.md.

Quarters through Q2 2026 were published under methodology v1 (50/30/20 composite, no agentic factor) and are retained as published. Quarterly moves over ±8 points carry a stated restatement reason (four in Q3 2026).

Sources: O*NET task lists, public AI capability evidence, employer surveys (BCG, Gallup, WEF, McKinsey), and BLS Occupational Employment Statistics.

Load it

from datasets import load_dataset
ds = load_dataset("Jott2121/ai-career-threat-index")
df = ds["train"].to_pandas()

# Roles most exposed to autonomous agents specifically
print(df.nlargest(10, "sub_agentic_exposure")[["title", "sub_agentic_exposure", "score"]])

# Most AI-resilient roles
print(df.nsmallest(5, "score")[["title", "score"]])

Interactive explorer: https://jott2121.github.io/ai-career-threat-index/

Honest limitations

The scores are structured editorial estimates made against a published rubric — not empirical measurements of realized job loss. The formula and every sub-score are published precisely so they can be challenged; treat the dataset as a structured, versioned prior for analysis and reporting, not ground truth. Scores within 5 points of a band boundary may retain their prior band across quarterly reviews to avoid noise.

Citation

@dataset{ai_career_threat_index,
  title  = {AI Career Threat Index},
  author = {Otterson, Jeff},
  year   = {2026},
  url    = {https://github.com/Jott2121/ai-career-threat-index},
  doi    = {10.5281/zenodo.21227026},
  note   = {Open dataset of AI displacement risk for 300 professions with published
            sub-scores including agentic exposure, MIT licensed}
}

Canonical repo: https://github.com/Jott2121/ai-career-threat-index — License: MIT.