Skip to main content

About

Careers built to outlast AI.

Every week there's a new headline about AI replacing jobs. Most of it is fear or hype. This site cuts through it by focusing on the opposite: the careers that still need a human, and aren't going anywhere anytime soon.

For every career on the site, you get the facts that matter: real median pay, the 10-year job outlook, the education and licensing you actually need, the day-to-day, and a concrete path to get started, plus an honest read on how safe it is from automation.

The numbers come from the best free source there is, the U.S. Bureau of Labor Statistics and O*NET. That data is excellent but buried in a clunky government interface, so this site puts it somewhere clean, focused, and easy to read at a glance.

The leaders of the companies building AI now say the disruption lands first on desk work. We map the other side of that: what AI is and isn't likely to replace, and why.

BLS · O*NET

Official U.S. data sources

May 2025

Latest BLS wage data

Free

No signup, fully sourced

Where the numbers come from

Median pay, salary ranges, job outlook, and typical education for every career come from the U.S. Bureau of Labor Statistics Occupational Outlook Handbook ↗ (May 2025 wages and 2024 to 2034 projections). Job tasks, skills, and tools draw on O*NET ↗, the U.S. Department of Labor's occupational database.

Every career page links its exact source so you can check it yourself. The BLS updates these figures each year, and we refresh them when it does.

Pay, outlook, and education are U.S. figures from the BLS. The AI-resistance score reflects the nature of the work itself, so the ranking applies broadly beyond the U.S., even though the dollar amounts do not.

The AI-resistance score is the WontReplace Index (WRI), our own two-axis model. Capability Gap is computed from real O*NET ratings; Deployment Friction is coded from BLS facts and state law. Full methodology, including every weight and O*NET element ID, is published in the methodology section below.

The WontReplace Index, in full

WRI 2026.1

The WRI is a two-axis, time-stamped, transparent model. Most published automation indices measure only whether AI is capable of doing a job. WRI measures both that and whether AI can actually be deployed in the role, which is the half the labor market really turns on.

Formula

WRI = 0.55 × Capability Gap + 0.45× Deployment Friction

Both axes are scored on the same 0 to 10 scale, normalized so the AI-safe threshold maps to 9.0 and the most resistant work approaches 10. The score is then exactly that blend of the two axes, so every career we list reads 9.0 or higher and the breakdown adds up. For calibration, occupations below the safe threshold (shown further down) score under 9 on the same scale.

Axis 1: Capability Gap

What current AI cannot do in the work. Pulled from real O*NET descriptor ratings (0 to 100) across three human bottlenecks. Each descriptor below is named by its O*NET element ID so anyone can verify it.

How the three are combined

A job is hard to automate if it leans heavily on at least onehuman bottleneck, so the Capability Gap is driven by a career's two strongest bottlenecks (0.6 × strongest + 0.4 × next), not a flat average across all three. A flat average would unfairly penalize a job that is hard for AI on one axis but light on the others. Being grounded in two bottlenecks (like a respiratory therapist, who is both relational and judgment heavy) scores higher than leaning on just one.

Physical and embodied work and sensing

  • 4.A.3.a.1Performing General Physical Activities
  • 4.A.3.a.2Handling and Moving Objects
  • 4.C.2.d.1.bSpend Time Standing
  • 4.C.2.d.1.gSpend Time Using Your Hands
  • 4.C.2.b.1.eExposed to Cramped Work Space, Awkward Positions
  • 1.A.2.a.2Manual Dexterity
  • 1.A.2.a.3Finger Dexterity
  • 1.A.2.a.1Arm-Hand Steadiness
  • 1.A.2.b.2Multilimb Coordination
  • 1.A.3.c.3Gross Body Coordination
  • 1.A.4.a.6Depth Perception

Real-time relational work with stakes

  • 4.A.4.a.5Assisting and Caring for Others
  • 4.A.4.a.4Establishing and Maintaining Interpersonal Relationships
  • 4.A.4.a.7Resolving Conflicts and Negotiating with Others
  • 4.A.4.b.5Coaching and Developing Others
  • 4.C.2.a.3Physical Proximity
  • 4.C.1.a.2.lFace-to-Face Discussions
  • 4.C.1.d.2Dealing With Unpleasant, Angry, or Discourteous People
  • 2.B.1.aSocial Perceptiveness
  • 2.B.1.fService Orientation
  • 2.B.1.dNegotiation

Improvisational judgment in novel contexts

  • 4.A.2.b.1Making Decisions and Solving Problems
  • 4.A.2.b.3Thinking Creatively
  • 4.A.2.a.1Judging the Qualities of Objects, Services, or People
  • 4.C.3.b.7Importance of Repeating Same Tasks(inverted)
  • 4.C.3.a.4Freedom to Make Decisions
  • 4.C.3.a.2.bFrequency of Decision Making
  • 4.C.3.a.1Consequence of Error
  • 2.B.2.iComplex Problem Solving
  • 1.A.1.b.2Originality
  • 1.A.1.b.1Fluency of Ideas
  • 1.A.1.b.3Problem Sensitivity

Axis 2: Deployment Friction

Whether AI can actually be put into the role, regardless of capability. These are our own structured estimates, informed by BLS "how to become one" facts and state licensing law, not a government statistic. The four dimensions are weighted equally, though in practice licensing, accountability, and public trust move together, so read them as one legal-and-trust requirement plus a separate capital-and-scale factor.

Licensing

weight 0.25

D1. Does the law require a credentialed human to perform the work? (Higher = stronger legal requirement.)

Accountability

weight 0.25

D2. Is a real person on the hook for outcomes (malpractice, board sanctions, criminal liability)?

Public trust

weight 0.25

D3. Does the role inherently require a person the public can hold accountable?

Capital and scale

weight 0.25

D4. Even if AI is capable, how hard is it to economically deploy? (Per-site vs scaleable.)

Validation against prior research

A credible AI-resistance score should line up, in the right direction, with the best-known prior automation index. Across 19well-known reference occupations (none listed on the site), the WRI correlates with Frey and Osborne's (2013) probability of computerisation at a Spearman rank correlation of about -0.65 (Pearson -0.83): the more AI-resistant a career scores here, the lower its computerisation probability there.

These 19 are well-known reference occupations scored only to test the index against prior research; none are careers listed on the site. The WRI shows strong convergent validity with the best-known prior automation index: a Pearson r of -0.83 and Spearman rho of -0.65 confirm the two measures rank careers in opposite directions, exactly as a valid AI-resistance score should. Caveats temper this. (1) Frey-Osborne is a 2013 estimate built on pre-LLM, mechanical/routine-task computerisation and is widely criticised as miscalibrated (e.g. it scores dental hygienists at 0.68 and welders at 0.94, which generative AI has not borne out), so agreement is partial validation, not ground truth. (2) The set is range-restricted: all 19 careers were curated as AI-safe, so most cluster at low F&O probabilities and the Pearson value is sensitive to the few mid/high-risk trades (welder, carpenter, HVAC) that anchor the slope; the more robust Spearman of -0.65 is the fairer headline. (3) Three SOCs are 2018-revision codes mapped to F&O's 2010 codes, and nurse-practitioner has no separate F&O entry and was proxied to registered nurses, though dropping it barely changes the result.

Frey & Osborne (2013), The Future of Employment ↗

Calibration: the index can score low, too

Because we only list careers that score as safe, it is fair to ask whether the formula can produce a low score at all. It can. Here is how the exact same model rates occupations widely considered highly exposed to AI. They are shown for calibration only and are not careers we recommend.

  • Customer service rep4.8 / 10 · Some risk
  • Telemarketer4.5 / 10 · Some risk
  • Bookkeeping clerk4.1 / 10 · At risk
  • Data-entry keyer4.3 / 10 · At risk
  • Proofreader2.4 / 10 · At risk

Calibration and limits

  • The Capability Gap is objective: it reads O*NET descriptor ratings directly. The Deployment Friction is our own structured judgment, informed by licensing law and BLS "how to become one" facts, not a government statistic. We label it as ours.
  • A living index. The facts (pay, outlook, education) come from the latest BLS release and update once a year. The WRI is a dated model (this is WRI 2026.1); as AI gets more capable we re-score and publish a new vintage, rather than pretend a number is permanent. That is the honest way to track a fast-moving target.
  • There is no objective "true" AI-resistance number. Major published indices (Frey and Osborne 2013, Felten et al. AIOE, Eloundou et al. GPT exposure) disagree, and most predate large language models. The WRI is a transparent model, not a measurement of truth.
  • Fully reproducible: scripts/refresh-onet.mjs pulls the O*NET data and scripts/refresh-bls.mjs the wages; src/data/wri.generated.json, bls.generated.json, and deployment.json hold the inputs; src/lib/wri.ts is the formula.
Browse the careers →View all categories