Unique Research · Insight · October 18, 2024 · Source: Indeed Hiring Lab (CN editorial of ARK/Indeed research package)
Key findings
- Of 2,800+ U.S. job skills scored, none rated “very likely” to be replaced by GenAI. On a 5-point scale (very unlikely → very likely), 68.7% scored “very unlikely” or “unlikely.”
- If firms change practices and tools improve, about 28.5% of skills could become replaceable.
- GenAI is strong on theory, weaker on problem-solving; wherever heavy hands-on presence is required (aviation, cooking, nursing), usefulness stays limited.
- Occupation-level impact = share of skills GenAI can do well now or later. In hands-on roles GenAI can take repetitive docs; in office roles (incl. software) it supplies knowledge and small fixes — continuous upskilling still required.
Bottom line: this generation of GenAI tools is almost never going to fully replace skilled humans or master any one of thousands of common job skills alone. Successful workers flex across three domains — theory, problem-solving, physical/digital execution. Digital “workers” may excel on one or two axes but struggle to balance all three. Models can keep learning, but that needs big digitalization and workflow change first. Best fit today: simpler tasks with moderate problem-solving and no hands-on requirement.
Method
Indeed filtered hundreds of millions of postings into 2,800+ common skills (customer management → trauma care). Hiring Lab built a long (>1,000-word) prompt for GPT-4o to score each skill on three 1–5 dimensions: theory delivery; problem-solving with the skill; necessity of human physical presence. Combined score = replacement likelihood (1 very unlikely → 5 very likely). Months of human–AI iteration; final prompt run 15× and aggregated. Human researchers validated outputs. Irony: the method itself proves the thesis — humans supervising, directing, and correcting GenAI remain hard to replace.
Theory > practice
GenAI rates itself “good” (4) on theory for 79.7% of skills; “moderate” (3) problem-solving for 70.7%. For 54% of skills, human physical execution is “essential” (1) or “very necessary” (2). No skill scored 5 (very likely replaceable). 28.5% may become replaceable with tech + workplace change; 68.7% stay very unlikely/unlikely.
Theory is the only dimension that ever hit a 5 (aside from physical-necessity scores). Models summarize and reframe fast — but also invent and leap logically. Image/audio generators can amplify stereotypes; GenAI has no intrinsic ethics — human oversight of prompts and outputs remains essential.
Hands-on share drives replaceability
Problem-solving ability + physical necessity dominate the replacement score. ~70.7% of skills score 3 on problem-solving; of those, 27.7% also score 3 on replacement. Only 2.8% score 4 on both problem-solving and replacement. Even when problem-solving is “good,” GenAI cannot do physical tasks today — among good problem-solving skills, 101 need only slight/no physical labor. Across all skills, GenAI calls physical execution “essential” for 30.1% → those stay very unlikely/unlikely to replace.
Occupation impact
Tech roles (digital, low hands-on) show higher GenAI competence; nursing (wide skill mix, high physical) lower. Software development: >71% of common skills potentially replaceable; general nurse: ~32.9%. Lower physical-skill share → higher replaceable-skill share. Accounting tops replaceability (~78% of common skills); cooking sits near the bottom. Accounting is also the only occupation where GenAI rates itself “good” at problem-solving on ≥30% of skills (30.5%) — elsewhere problem-solving ceilings keep replacement shares lower. If problem-solving rises later, replacement shares can rise with it.
Software vs nursing skill clouds
Shared soft skills (comms, leadership) exist, but occupation-specific skills diverge. GenAI learns Python/JavaScript and can debug human code; ~40.6% of software postings ask for Java. It can help draft nursing care plans but cannot give meds or run emergency hands-on care. Top-25 skill word clouds for both roles are in CN charts (~18 imgs).
Not investment advice. Charts: violin of skill ratings; tile charts for problem-solving × replaceability and physical × replaceability; occupation barplot; two scatterplots; two word clouds — attach on import.
— END —