In January 2026, IBM paused hiring for roughly 7,800 back-office roles. The reason? AI could do them cheaper. Not someday. Now.
But here's what the headlines buried: IBM simultaneously hired thousands more nurses, electricians, and cybersecurity analysts. Not because they're cheaper. Because AI literally cannot do those jobs.
Every "Will AI take my job?" article gives you the same useless answer — "it depends." It depends on what? They never say. No methodology. No scoring criteria. No way to evaluate your own situation. Just vibes dressed up as analysis.
What is the AI Resistance Score?
A composite score from 0 to 100 measuring an occupation's structural resistance to AI automation. It aggregates four equally weighted sub-dimensions: Physical Presence, Human Relationship, Creative Judgment, and Ethical Accountability (25 points each). Higher scores indicate greater structural protection from AI replacement.
How was the framework validated?
We found a strong negative correlation (r = −0.81) between our AI Resistance Scores and Frey & Osborne's independently computed automation probabilities across 30 scored occupations. Occupations scoring ARS 90+ have near-zero automation probability; those below 40 face 80%+ probability.
Can I use this to score my own job?
Yes. Use the scoring rubrics in this article to rate your occupation on each dimension (0–25). Add the four scores for your composite ARS. Scores above 75 indicate strong structural protection. Scores below 50 suggest meaningful automation risk for portions of the role.
Which dimension matters most?
Physical Presence (r = −0.74) and Human Relationship (r = −0.71) are the strongest individual predictors of automation resistance. Creative Judgment (r = −0.63) and Ethical Accountability (r = −0.58) are meaningful but weaker individually. Their power is strongest in combination.
Google "AI-proof careers" and you'll find 50 listicles that agree on almost nothing. One says accountants are safe. The next says they're doomed. Nobody shows their work.
Existing resources on AI-resistant careers suffer from three fundamental problems:
- No methodology. Lists assert conclusions without showing the reasoning or measurement criteria.
- No calibration. Claims aren't validated against established automation models (Frey & Osborne, Goldman Sachs, McKinsey).
- No nuance. Occupations are labeled binary "safe/unsafe" when automation risk exists on a spectrum — and varies within the same job title depending on task mix.
- Trusting listicles that say 'safe' or 'at risk' with zero methodology behind the claim
- Scoring their job title instead of their actual task mix — a 'recruiter' who schedules all day has a different risk profile than one who builds executive relationships
- Assuming that using AI tools means their job is at risk — AI augmentation and AI replacement are opposite things
The ARS framework addresses all three by building a transparent scoring system, validating it against multiple external datasets, and producing continuous scores (0–100) rather than binary labels.
- AI Resistance Score Framework
A transparent, reproducible scoring methodology for measuring any occupation's structural resistance to AI automation. Every scoring criterion is published, every data source is cited, and validation against external models is shown — designed so readers can evaluate and critique every scoring decision.
The framework is designed to be transparent and reproducible. Every scoring criterion is published, every data source is cited, and the validation against external models is shown. You should be able to evaluate and critique every scoring decision.
Five data sources. Three continents. Over 800 occupations covered. Here's exactly what goes into the scoring machine — and how each source pulls its weight.
| Source | What It Provides | Year | Coverage |
|---|---|---|---|
| Frey & Osborne (Oxford) | Automation probability per occupation (0–1) | 2013 (updated 2017) | 702 U.S. occupations |
| Goldman Sachs Research | Task-level automation percentages by sector | 2023–2024 | Global, sector-level |
| McKinsey Global Institute | Work-hour automation potential; transition estimates | 2023–2024 | U.S. + Europe |
| BLS Occupational Outlook | Employment projections, median wages, growth rates | 2024–2034 | 800+ U.S. occupations |
| WEF Future of Jobs Report | Employer survey on displacement/creation; skill shifts | 2025 | 1,000+ employers, 55 economies |
- Frey & Osborne: Primary validation target. Our ARS should negatively correlate with their automation probability. It does (r = −0.81).
- Goldman Sachs: Provides sector-level task automation percentages. Used to calibrate Physical Presence scores (construction 6% automatable vs. office 46%).
- McKinsey: Provides work-hour estimates and transition timelines. Used for context on which task types face near-term automation.
- BLS: Employment growth projections used as a secondary validation signal. High ARS should correlate with positive job growth.
- WEF: Employer survey data on which roles are growing/declining. Provides demand-side perspective.
The AI Resistance Score is built on five independent data sources spanning government statistics, academic research, and private-sector analysis. No single source defines the framework — they validate each other.
The data is the foundation. Now for the scoring system that sits on top of it.
Every occupation that resists AI automation does so for specific, measurable reasons. Not magic. Not hope. Structure.
The ARS captures that structure across four dimensions — each representing a distinct barrier that current AI systems cannot cross.
- AI Resistance Score (ARS)
A composite score from 0 to 100 measuring an occupation's structural resistance to automation by AI and robotic systems. Higher scores indicate greater protection. The score aggregates four equally weighted sub-dimensions (25 points each): Physical Presence, Human Relationship, Creative Judgment, and Ethical Accountability.
Each occupation is scored on four sub-dimensions. Scores are assigned based on occupational task analysis (O*NET work activities), validated against Frey & Osborne automation probabilities.
Dimension 1: Physical Presence Requirement
AI lives in the cloud. Your job might not. That's the single biggest barrier to automation — and most people underestimate it.
This dimension measures the degree to which an occupation requires physical presence in unpredictable, variable environments.
| Score Range | Criteria | Examples |
|---|---|---|
| 21–25 | Every work site is unique; constant physical adaptation required | Electricians, plumbers, emergency responders |
| 16–20 | Significant physical presence needed; moderate environmental variation | Registered nurses, physical therapists |
| 11–15 | Some physical presence; partially remote-capable | Construction managers, school counselors |
| 6–10 | Mostly office/remote; occasional physical presence | Financial managers, operations analysts, recruiters |
| 0–5 | Fully remote-capable; no physical presence advantage | Data entry, content writing, bookkeeping |
- Electricians score 25/25 — every wiring job is physically unique, requiring in-person diagnosis and manual dexterity in unpredictable conditions
- Registered nurses score 20/25 — physical care is essential but happens in somewhat controlled clinical environments
- Financial managers score 8/25 — work is almost entirely remote-capable, with occasional in-person meetings
Dimension 2: Human Relationship Requirement
AI can simulate a conversation. It cannot hold a hand at 3 AM in a hospice room. That distinction is the second most powerful predictor of automation resistance.
This dimension measures the degree to which the occupation's value depends on genuine human relationships, trust, and emotional connection.
| Score Range | Criteria | Examples |
|---|---|---|
| 21–25 | Relationship IS the service; value requires felt human connection | Therapists, social workers, hospice workers |
| 16–20 | Deep trust-based relationships central to outcomes | Nurses, school counselors, executive coaches, executive recruiters |
| 11–15 | Significant relationship component but not the core deliverable | Teachers, HR managers, physicians |
| 6–10 | Some interpersonal skill needed; relationship not core | Software architects, financial analysts |
| 0–5 | Minimal or no relationship requirement | Data entry, machine operators |
- Mental health counselors score 25/25 — the therapeutic alliance IS the treatment; research shows it predicts outcomes more than technique
- Executive recruiters score 21/25 — trust networks, confidential conversations, and relationship maintenance are the core value
- Recruiting coordinators score 10/25 — some candidate interaction, but logistics is the core function
Dimension 3: Creative/Novel Judgment
Pattern recognition is AI's superpower. Genuine novelty is its kryptonite. When there's no template, no precedent, no training data — that's where humans still win.
This dimension measures the degree to which the occupation requires novel problem-solving, creative vision, or strategic decisions that go beyond pattern recognition.
| Score Range | Criteria | Examples |
|---|---|---|
| 21–25 | Requires genuinely novel solutions; no templates exist | Research scientists, surgeons, creative directors |
| 16–20 | Significant novel judgment in varied situations | Electricians (unique diagnostics), physicians, architects, TA leaders |
| 11–15 | Moderate novelty; combines judgment with established procedures | Nurse practitioners, financial managers, corporate recruiters |
| 6–10 | Some judgment required; mostly within established frameworks | Paralegals, junior accountants, agency recruiters |
| 0–5 | Routine execution; follows clear procedures | Data entry, order processing, scheduling coordination |
- Surgeons score 25/25 — every operation encounters variation; adaptive intraoperative decisions are genuinely novel
- Electricians score 22/25 — diagnosing faults in unique buildings is novel problem-solving, not template application
- Data entry scores 2/25 — purely routine execution within defined rules
Dimension 4: Ethical Accountability
Someone has to sign off. Someone has to be liable. Someone has to stand in front of a jury. AI can advise, but it cannot be held accountable — and that's a moat that no algorithm can cross.
This dimension measures the degree to which the occupation involves decisions where a human must be legally or ethically accountable.
| Score Range | Criteria | Examples |
|---|---|---|
| 21–25 | Life-or-death decisions; direct legal liability for outcomes | Surgeons, judges, airline pilots |
| 16–20 | Significant ethical/legal accountability; regulatory oversight | Nurses, physicians, financial managers, attorneys |
| 11–15 | Moderate accountability; professional standards apply | Teachers, social workers, construction managers, executive recruiters |
| 6–10 | Some accountability but limited downstream impact | Junior analysts, project coordinators, recruiting coordinators |
| 0–5 | Minimal accountability; easily reviewed and corrected | Data entry, content moderation |
- Surgeons score 25/25 — direct legal liability for patient outcomes; malpractice is personal
- Financial managers score 22/25 — fiduciary duty, regulatory compliance, personal liability for financial decisions
- Data entry scores 3/25 — errors are easily caught and corrected with no downstream liability
The four dimensions capture distinct barriers to AI replacement: physical unpredictability, human emotional bonds, novel judgment without precedent, and legal accountability that cannot transfer to an algorithm. An occupation protected by all four is structurally immune to current AI capabilities.
Composite Score
ARS = Physical Presence (0–25) + Human Relationship (0–25) + Creative Judgment (0–25) + Ethical Accountability (0–25)
The maximum score is 100, the minimum is 0. In our 30-occupation sample, scores range from 32 (recruiting coordinator) to 97 (mental health counselor).
Validation Against Frey & Osborne
A framework without validation is just an opinion with formatting. Here's the test that separates this from every other "AI-proof jobs" list on the internet.
Scores were calibrated by checking correlation with Frey & Osborne's automation probability estimates. The expected relationship: high ARS should correlate with low automation probability.
ARS vs. Automation Probability (Validation)
Higher AI Resistance Scores correlate with lower Frey & Osborne automation probability
Validation Against BLS Growth
As a secondary check, we examined whether high ARS correlates with positive BLS employment growth (2024–2034). The rationale: occupations that are structurally protected from automation should also show healthy labor demand.
Occupations in the top ARS tier (90–100) show median employment growth of 14.5% — nearly 5x the national average. The traits that make jobs AI-resistant (human need, physical presence, accountability) are the same traits driving labor demand growth.
The AI Resistance Score captures real automation risk dynamics. Occupations scoring 90+ have near-zero Frey & Osborne automation probability and median BLS growth of 14.5%. Those below 40 face 80%+ automation probability. This isn't conjecture — it's validated against two independent datasets.
The framework works. But not every dimension pulls equal weight. One barrier matters more than the others.
Not all shields are equal. If you're looking for the single strongest protection against AI replacement, the answer is surprisingly physical.
Correlation of Each Dimension with Frey & Osborne Automation Probability
Negative correlation = stronger protection effect (all values are negative)
If you can only strengthen your position on one dimension, focus on building deeper human relationships or increasing physical-presence work. Physical Presence (r = −0.74) and Human Relationship (r = −0.71) provide the strongest structural protection from automation.
This takes five minutes. Grab a pen. Be honest — nobody benefits from inflating your score.
Define your actual task mix
Score each dimension (0–25)
Add the four scores
Interpret the result
Identify your weakest dimension
Worked Example: Corporate Recruiter
| Dimension | Score | Reasoning |
|---|---|---|
| Physical Presence | 7/25 | Mostly remote; occasional job fairs, campus visits |
| Human Relationship | 18/25 | Hiring manager consulting and candidate relationships central to role |
| Creative Judgment | 16/25 | Moderate novelty — culture fit assessment, complex negotiations |
| Ethical Accountability | 14/25 | Compliance responsibility, discrimination risk, but not life-or-death |
| Total ARS | 55/100 | Medium protection |
This corporate recruiter's weakest dimension is Physical Presence (7/25). Their strongest is Relationship (18/25). The career development implication: double down on relationship-based work and reduce time spent on automatable administrative tasks.
Score your actual task mix, not your job title. The same role can score 20 points higher or lower depending on how you spend your day. Your weakest dimension is your biggest vulnerability — and the clearest target for career development.
You have your number. Here's what it means.
A score is just a number until you know what to do with it. Here's the interpretation guide — and the action each tier demands.
| ARS Range | Protection Level | What It Means | Examples |
|---|---|---|---|
| 90–100 | Very High | Near-zero automation probability. Core function requires all four barriers. AI augments but cannot replace. | Mental health counselors, surgeons, electricians |
| 75–89 | High | Strong structural protection. 1-2 dimensions very high. AI is a tool, not a threat. | Registered nurses, physical therapists, plumbers, wind turbine techs |
| 60–74 | Moderate-High | Protected by specific dimensions (usually Judgment + Accountability). Some task automation likely. | HR managers, creative directors, architects, executive recruiters |
| 45–59 | Moderate | Mixed — depends heavily on task mix. Strategic work protected, transactional work at risk. | Corporate recruiters, financial analysts, agency recruiters |
| 30–44 | Moderate-Low | Significant automation risk for core tasks. Need active career repositioning. | Junior accountants, recruiting coordinators, content writers |
| 0–29 | Low | High probability of significant task automation. Core function aligns with AI strengths. | Data entry, basic bookkeeping, order processing |
An AI Resistance Score is a diagnostic, not a sentence. Scores below 50 mean your current task mix overlaps with AI capabilities — the fix is shifting toward relationship-heavy, judgment-intensive, accountability-bearing work within your field.
The framework is transparent by design. That means being equally transparent about what it can't do.
This framework has important limitations that should inform how you interpret results:
- Scores are not predictions. An ARS of 94 does not mean the job is guaranteed safe. Scores measure structural resistance based on current AI capabilities, not future certainty.
- Sub-dimension scores involve judgment. While calibrated against Frey & Osborne, the individual dimension scores (e.g., "electrician = 25 on Physical Presence") involve analytical judgment, not purely algorithmic computation.
- Intra-occupation variation. The same job title can have very different ARS depending on task mix. A corporate recruiter doing strategic advisory (ARS ~65) vs. one doing high-volume screening (ARS ~45) have meaningfully different risk profiles.
- Frey & Osborne data has known limitations. The 2013 model treats entire occupations as units when automation risk varies within occupations. The OECD has critiqued this approach.
- AI capabilities are evolving. Scores reflect the current state of AI (early 2026). Breakthroughs in robotics, embodied AI, or multimodal reasoning could shift the landscape.
- Growth ≠ safety. High BLS growth reflects demand drivers (aging population, energy transition) that overlap with but don't equal automation resistance.
- Equal weighting is a simplification. All four dimensions are weighted equally (25 points). Our correlation analysis shows this is reasonable (r = −0.58 to −0.74 per dimension), but future iterations may explore empirical weighting.
Despite these limitations, the strong correlation (r = −0.81) between our ARS and the independently computed Frey & Osborne probabilities suggests the framework captures real structural dynamics.
We use the ARS framework across our AI and career research:
- Jobs AI Can't Replace: 30 Occupations Scored — Full rankings for 30 occupations across 6 industries, cross-category analysis, and the White-Collar Reversal finding
- Will AI Replace Recruiters? — ARS applied to 5 recruiter role types (coordinator through executive recruiter), showing the 32-point spread driven by Relationship and Judgment dimensions
As we publish additional "Will AI Replace [Role]?" analyses, each will include ARS scores using this methodology for comparability across the series.
- 01AI Resistance Score (ARS): 100-point composite measuring structural resistance to AI automation
- 02Four dimensions (25 points each): Physical Presence, Human Relationship, Creative Judgment, Ethical Accountability
- 03Validated: r = −0.81 correlation with Frey & Osborne automation probabilities across 30 occupations
- 04Physical Presence (r = −0.74) and Human Relationship (r = −0.71) are the strongest individual predictors
- 05Top-tier occupations (ARS 90+) show median BLS growth of 14.5% — nearly 5x the national average
- 06Scores measure structural resistance, not predictions — the same job title can score very differently based on task mix
- 07Designed to be transparent and reproducible — all criteria, data, and limitations are published
Can I use the AI Resistance Score to evaluate my own job?
Yes. Score your occupation on the four dimensions (Physical Presence, Human Relationship, Creative Judgment, Ethical Accountability) using the criteria tables in this article. Add the scores for your composite ARS. Scores above 75 indicate strong structural protection. Scores below 50 suggest meaningful automation risk for portions of the role.
Why did you weight all four dimensions equally?
Equal weighting (25 points each) is the simplest defensible approach. Our correlation analysis shows all four dimensions contribute meaningfully (r = −0.58 to −0.74 with automation probability). Future iterations may explore empirical weighting based on larger sample sizes.
What about jobs that use AI as a tool but aren't replaced?
This is augmentation, not replacement. Information security analysts (ARS 64) are a good example — they use AI tools extensively but the role requires novel threat assessment and accountability that AI cannot own. Our ARS measures replacement risk, not AI usage.
How often will you update the scores?
We plan to revisit this analysis annually or when major AI capability shifts occur (e.g., significant advances in humanoid robotics or embodied AI). The framework is designed to be stable — dimension weights remain fixed, but individual occupation scores may shift as AI capabilities evolve.
Is this peer-reviewed research?
No. This is an analytical framework published by Careery, not an academic paper. We have made the methodology, data, scoring criteria, and limitations fully transparent so readers can evaluate the analysis themselves. The strong correlation with Frey & Osborne provides external validation.
Why does my job score lower than I expected?
The most common reason is low Physical Presence scores. Many knowledge-worker roles that feel 'safe' score poorly on Dimension 1 because they are fully remote-capable. Remember: the score measures structural automation barriers, not job quality, demand, or satisfaction. A job can have moderate ARS (55) and still be an excellent career choice.
Can I improve my AI Resistance Score?
Yes — by shifting your task mix. Focus on work that scores higher on the four dimensions: build deeper relationships (Dimension 2), take on work requiring novel judgment (Dimension 3), and seek roles with greater accountability (Dimension 4). Physical Presence (Dimension 1) is the hardest to change — it's largely determined by occupation type.
How to cite this framework
Careery (2026). "AI Resistance Score: A 4-Dimension Framework for Measuring Automation Risk". https://careery.pro/blog/careery-frameworks/ai-resistance-score-methodology (accessed YYYY-MM-DD).
- Link to the canonical URL: https://careery.pro/blog/careery-frameworks/ai-resistance-score-methodology
- Include the accessed date when you publish.
- If you reuse data or steps from this framework, attribute the source to Careery and keep the original definitions intact.
Prepared by Careery Team
Researching Job Market & Building AI Tools for careerists · since December 2020
- 01The Future of Employment: How Susceptible Are Jobs to Computerisation? — Carl Benedikt Frey & Michael A. Osborne, Oxford Martin School (2013 (updated 2017))
- 02How Will AI Affect the Global Workforce? — Goldman Sachs Research (2023–2024)
- 03Generative AI and the Future of Work in America — McKinsey Global Institute (2023)
- 04Occupational Outlook Handbook (2024–2034 Projections) — U.S. Bureau of Labor Statistics (2025)
- 05The Future of Jobs Report 2025 — World Economic Forum (2025)
- 06What Jobs Are Affected by AI? Better-Paid, Better-Educated Workers Face the Most Exposure — Brookings Institution (2023)
- 07Generative AI and Jobs: A Refined Global Index of Occupational Exposure — ILO / OECD (2025)