Andrew Yang on Jobs Safe From AI: How to Assess Your Career

The short answer

Andrew Yang's central point is directionally useful: jobs tend to be safer from near-term AI replacement when the work requires physical presence, trusted human relationships, hands-on judgment, responsibility for people, or operation in institutions that cannot change overnight. He contrasts those conditions with highly digitized, repetitive knowledge work inside organizations under strong pressure to cut costs. That is a better starting point than assuming every office job is doomed or every hands-on job is untouchable.

But 'safe' should mean more resilient for the time being, not immune. AI exposure measures what technology might help perform; they do not by themselves predict layoffs, wages, hiring, or the disappearance of an occupation. Demand, regulation, budgets, customer preferences, management choices, and the cost of reorganizing work all affect what happens next. A teacher, trainer, nurse, electrician, analyst, or designer can each have some tasks changed while the larger role persists.

Use Yang's argument as a prompt for a specific audit of your work. Break the role into tasks, identify who values the human contribution, study your employer's incentive and capacity to automate, and move your evidence toward decisions, relationships, physical-world execution, revenue, and accountable outcomes. The goal is not to guess the one perfect title. It is to become harder to substitute and more able to move if your current role changes.

What Andrew Yang says makes a job safer

In his essay 'What Jobs Are Safe?', Yang begins with a personal trainer worried that an app could replace him. Yang argues that the trainer has several defenses: clients value live encouragement and physical assistance, his market contains customers willing to pay for personal service, and older clients may prefer a person to a screen. He makes a similar case for teachers, whose work includes supervision, motivation, safety, and classroom authority rather than content delivery alone.

His higher-risk signals include work that happens mainly at a computer, routine coding or analysis, large departments, public-company cost pressure, customer service, media production, and roles that do not clearly touch revenue. His safer signals include working with people, children, patients, animals, tools, repairs, homes, or the outdoors; union or institutional protection; and being the person others consult about AI. He also notes—sometimes with deliberate humor—that employer culture and customer willingness to pay can matter as much as a task score.

The valuable idea underneath the list is that technology does not enter an empty room. A company must have an incentive to deploy it, access to workable systems and data, permission to change the process, and a plan for who remains accountable. Customers and employees also have to accept the new arrangement. Your risk therefore depends on both what software can do and the environment in which your job exists.

Where the evidence supports the argument

The International Labour Organization's refined global index found that about one in four workers worldwide is in an occupation with some generative-AI exposure. Clerical work remains the most exposed, while exposure has expanded in highly digitized professional and technical roles. Crucially, the ILO concludes that transformation is more likely than full redundancy for most occupations because jobs contain varied tasks and continue to require human input.

Microsoft Research reached a compatible but narrower finding by studying 200,000 anonymized Copilot conversations. Generative AI showed the most applicability to information gathering, writing, teaching, advising, and other knowledge or communication activities; direct applicability was lower for physical tasks. Microsoft later emphasized that its score measures where AI may be useful, not which jobs will be eliminated. That distinction matters when a headline turns an applicability ranking into a layoff forecast.

Together, the research supports several parts of Yang's intuition. Work conducted entirely through digital information is easier for current generative AI to touch than unpredictable physical work. Communication-heavy jobs may be highly exposed even when the relationship, authority, or accountability surrounding the communication remains human. And an occupation can be exposed without vanishing: the mix of drafting, research, review, coordination, and decision-making can change inside the same title.

Where the checklist needs more nuance

A checklist can reveal risk signals, but counting yes answers does not produce a reliable probability of job loss. A well-paid analyst at a public company may use AI to expand the team's capacity rather than lose the role. A lower-paid service worker may face scheduling software, self-service systems, robotics, outsourcing, or budget cuts even if generative AI cannot perform the whole job. Institutional slowness can delay adoption without guaranteeing long-term protection.

Human contact is not a shield by itself. Some customers will accept less personal service for lower prices, while others will pay a premium for trust, convenience, taste, discretion, or a real relationship. Ask who pays, what they value, and what failure costs them. A physical therapist providing licensed assessment and hands-on care has a different moat from a generic workout plan. A teacher responsible for a classroom has a different role from a recorded lesson.

Ownership can increase control, as Yang suggests, but becoming a business owner is not a universal escape hatch. Owners face demand risk, capital needs, regulation, sales, and the possibility that their own business model is automated. Treat ownership as one possible source of leverage—not a requirement for a secure future. Internal ownership also counts: controlling an outcome, customer relationship, regulated decision, scarce process, or revenue stream can make an employee more valuable without requiring entrepreneurship.

Run a five-part AI resilience audit

First, map tasks rather than judging your title. For two typical weeks, list what you produce, decide, coordinate, inspect, explain, repair, and approve. Mark each activity as routine or variable, digital or physical, independent or relationship-dependent, and reversible or consequential. Then test whether current approved tools can assist with it reliably. Do not assume a polished demo can handle your actual context, exceptions, data boundaries, and quality standard.

Second, assess substitution and complementarity separately. A task has higher substitution exposure when its inputs and outputs are digital, the rules are stable, errors are cheap to detect, and little context is needed. It has stronger complementarity when AI can accelerate preparation but a person still improves the result through domain judgment, trust, physical action, negotiation, or accountability. Your objective is to shift time from easily replicated production toward complementary decisions and outcomes.

Third, examine employer incentives. Is leadership publicly committed to reducing headcount, or is demand growing faster than the team can serve it? Is your department treated as a cost center or tied to retention, risk, service quality, or revenue? Does the organization have clean data, approved tools, integration capacity, and permission to redesign the process? Watch actual workflow pilots, staffing plans, vendor purchases, and changed performance measures—not executive slogans alone.

Fourth, measure your human and institutional moat. Identify relationships people would notice losing, licenses or decision rights the organization must retain, physical environments that require adaptation, and sensitive exceptions routed to you. Be honest: being pleasant, experienced, or busy is not the same as owning scarce value. Write down three examples where your judgment changed an outcome and two capabilities another employer would also value.

Fifth, assess mobility. If 30 percent of your current tasks disappeared, which adjacent role could use the remaining evidence? How long would it take to qualify, what income floor must you protect, and who could validate the move? Resilience includes the ability to transition; no employer-specific moat should be mistaken for a portable career asset.

Strengthen your position without panic-pivoting

Start with one real workflow. Learn how AI affects it, document the baseline, use only approved data and tools, and measure time, error rate, review effort, or stakeholder value. Keep the consequential decision with a qualified person. A credible case study showing that you improved a process while catching limitations is stronger than adding a list of AI products to your résumé.

Move closer to outcomes that somebody can name and defend. That might mean taking responsibility for customer retention, implementation, quality, safety, revenue operations, compliance, complex cases, or adoption. It can also mean becoming the translator between a tool and the people affected by it. Do not merely produce faster; show that you can decide what should be produced, verify whether it works, and improve the system when it fails.

Invest in relationships and external proof. Maintain contacts beyond your immediate team, collect non-confidential evidence of results, and ask practitioners how their roles are changing. If you need training, choose it after studying live postings and talking with employers. A broad certificate purchased from fear is less useful than a targeted skill gap tied to a tested direction.

Build a contingency plan before you need one

Yang recommends saving money, contacting professional connections, and making a plan before disruption becomes urgent. That advice is practical even if his timeline or risk estimate does not fit your role. Set a trigger for action: a hiring freeze, reorganization, new automation target, shrinking workload, removed responsibilities, or a documented change in your market. Decide in advance which trigger means update your portfolio, begin conversations, or start applying.

Create two paths. Path A strengthens your current position through complementary AI skill, outcome ownership, and visible evidence. Path B prepares one adjacent option through informational conversations, a small project, and a realistic financial model. Review both every quarter. You are not predicting the future; you are reducing the cost of being wrong.

CareerWing can help you break your role into evidence, compare adjacent directions, and turn uncertainty into a short experiment. Keep the final decision grounded in your income, benefits, health, caregiving, location, and risk tolerance. A provocative article can start the question, but your career plan should be built from the conditions of your actual life.