# When You Need a Human > [!abstract] What this note is > The other half of writing about AI: the situations where it has nothing to offer, and a person does. Written after a day when the tool was useless and the problem still had to be dealt with. Part of the [[AI Reading List]] and [[AI Security for Users]] section. ## The short version **Seek professional help when seeking to accomplish your goals.** AI is good at a first draft, a summary, a checklist, a tireless pass over something mechanical. It is not a substitute for a person who is qualified, present, and accountable. Knowing which situation you are in is most of the skill. ## Where a person is required, not preferred **Anything with a licence behind it.** Law, medicine, tax, accounting, structural and electrical work, anything inspected or regulated. The licence is not bureaucracy. It means someone carries duty of care, insurance and consequences. A model carries none of those and never will, no matter how well it explains the subject. **Anything that has to be witnessed or signed.** A model can prepare a document. It cannot take responsibility for one. **Anything that needs a body in a room.** Some problems only move when someone shows up: the physical check, the site visit, the machine that has to be touched. Work can sit stuck for weeks in messages and then close in a single afternoon in person. When something has been stalled a long time, "go there" is often the whole answer. **Anything that depends on trust between two people.** Apology, repair, a hard conversation, grief, a negotiation where your standing in the room is the point. A model can help you find words. It cannot carry them, mean them, or be the one the other person forgives. **Anything where the risk is yours to live with.** Advice costs a model nothing, because it has no stake in the outcome. A professional who will still be there next year gives different advice, and better advice, for that reason alone. **Anything that requires knowing what is actually true right now.** A model works from what it was given: a copy, a cache, a summary. It will describe the live world with complete confidence and be wrong. Only someone who looks can confirm. > [!warning] If you are in crisis > A chatbot is not care. If you are struggling with your mental health, in danger, or thinking about harming yourself, talk to a person: a doctor, a therapist, a crisis line (in the US and Canada, call or text **988**), or someone you trust. That is not a fallback. That is the correct tool. ## What it is actually good for Worth being fair about, because the boundary works both ways: - A first draft you will rewrite - Summarising material you will still read - Repetitive, mechanical work with a checkable result - A second opinion before you talk to the person who actually knows - Finding the question you should be asking a professional Every one of those ends with a person deciding. ## The honest failure mode The tool doesn't announce that it can't help. It produces something plausible instead, at length, immediately. That is its worst property. Hours can go into a confident answer to a problem that needed one phone call. The tell is simple: **if the next step requires standing, presence, accountability, or somebody's trust, stop typing and go find the person.** --- *I write all of this for myself, and in the hope of inspiring others to succeed. Best wishes and one love.* **— Jeff, 18 September 2026** ## Related - [[AI Security for Users]] — where the tools fail dangerously - [[AI Reading List]] — the long view on what these machines are - [[Disclaimer]] — what these notes are, and what they are not