What jobs are safe from LLMs?

No job is 100% safe from LLMs forever, but some are far more resistant than others. The pattern is simple: LLMs are great at generating text, code, and analysis from patterns in data, but bad at physical manipulation, legal accountability, and earning trust face-to-face — so the jobs built on those human strengths hold up best.

Which jobs hold up best?

Four categories are the most resistant:

• Physical trades in unpredictable settings — plumbers, electricians, HVAC techs, and mechanics work in someone's messy basement or a car engine nobody photographed. Nobody's trained a model on that particular mess.
• High-stakes accountability and licensure — surgeons, judges, and pilots carry legal and professional liability a chatbot can't hold. Someone has to be the one who signs, operates, or rules, and answers for it.
• Deep interpersonal trust — therapists, hospice nurses, and salespeople closing complex B2B deals succeed because a human believes another human understands them. That's not a text-prediction problem.
• Skilled trades that combine diagnosis with hands-on fixes — reading a real, physical problem and then repairing it with your own hands, whether that's an electrician tracing a fault or an appliance repair tech.

Which jobs are feeling the pressure first?

Notice what's missing from that list.

• junior writing
• routine coding
• first-draft analysis
• basic customer support

Those already overlap heavily with what LLMs do well, which is why they're feeling pressure first. A rough rule of thumb: if the core of a job is producing text or analysis on a screen and the cost of a mistake is low, it's more exposed. If it needs hands, a license, or a relationship, it's more protected.

Why aren't these jobs safe forever?

Because "safest for now" beats "safe forever." Better robotics and tooling could close some of these gaps eventually, and the boundary has moved before. But for the moment, dexterity, accountability, and trust remain the moat — the three things a text-prediction system doesn't have, and can't fake by generating a better sentence.