When not to rely on AI
- judgment
- verification
Delegate by consequence: use AI freely where being wrong is cheap and reversible, let it assist where the stakes rise, and keep the decision with a human where the outcome is irreversible or someone has to be accountable. Three tiers, one question each.
The three tiers
- Delegate. Reversible, low-stakes, verifiable: drafting, summarizing, brainstorming, formatting, transforming. A mistake costs minutes and gets caught by a skim. Most everyday use lives here — enjoy it without ceremony.
- Assist. Real stakes, human decides: research synthesis to be verified, client-facing drafts to be reviewed, analyses to be checked. AI produces material; you produce the judgment.
- Keep human. Legal, medical and financial decisions about people; safety; anything irreversible; anything with a name on it. AI can still prepare: “list the questions to ask my lawyer” is a perfect request. “Tell me whether I am legally right” is not — a confident verdict is exactly the wrong output shape when the consequence is real.
Why the line sits there
Three properties mark the “keep human” zone: consequence (how much a mistake costs), reversibility (can it be undone), and accountability (who signs). The model cannot hold the third one — responsibility does not travel to a language model, whatever its tone suggests. And confident fluency is everywhere: the same mechanism that makes it make things up — a hallucination in calm prose — also makes it sound right about them.
The uncomfortable middle
Professionals use AI for exactly these topics — legally, well and daily. The discipline is not avoiding the tool; it is routing the parts correctly:
- AI → materials: summaries, question lists, draft language, options with trade-offs.
- You → the decision, informed by those materials.
- A qualified human → the verdict, whenever the consequence earns one.
A bad example
Ask: “Am I legally entitled to my full deposit back? Answer yes or no, and write the final reply.”
You are asking a system that cannot be accountable for a verdict — and inviting it to sound certain regardless.
A better example
Ask: “Summarize this letter in three points. List what I should verify under the tenancy rules. Draft a neutral request I can review — and flag anything that reads as legal advice so I check it with an advisor.”
AI does the compression and construction; the verdict goes to the professional; the decision and the signature stay with you.
Practice
Choose the sound delegation
Your landlord kept part of your deposit and sent a letter explaining why. You want AI help responding — and you are deciding what to actually ask it for.
Pick the request that delegates the right things to AI.
Hint
Which request treats the model as an assistant, and which treats it as a lawyer?
After each reply is sent: who is accountable if it is wrong?
Why this is the answer
The three-tier rule decides it: drafting and summarizing sit in 'delegate or assist', a legal verdict sits in 'keep human'. Prompt B keeps the material-production on the AI side and the consequence-bearing judgment where a professional can own it — with flags so advice-shaped output gets caught before it is acted on.
Transfer
- Health: organize information and questions; the diagnosis belongs to the doctor.
- Money: model scenarios as worksheets; the commitment is yours to make.
- People decisions: hiring, firing, discipline — AI drafts nothing that carries consequences for a person without full human judgment in the loop. Run your rubric before anything is said to a person.
Next
Next: Using AI at work: judgment calls — the last safety skill: your personal usage plan, written once and reused everywhere.