When not to use AI
The five no-go categories, the gray zone in between, and how to say no gracefully when a use is tempting but wrong.
Knowing when not to reach for the tool is the mark of someone who actually understands it. Five categories cover nearly every genuine no-go at work.
- When the point is that you did it. Performance self-reviews, personal apologies, condolences, recommendation letters you signed. The value of these is the human attention behind them; outsourcing it is detectable and corrosive.
- When you can't verify and it matters. If you lack the expertise or time to check output in a domain with real consequences (legal language, medical guidance, structural calculations), fluent-but-wrong is worse than nothing.
- When the data can't leave. Regulated data without an approved tool. Full stop — see the last lesson.
- When accountability must be traceable to judgment. Firing decisions, disciplinary findings, credit approvals. Delegating judgment breaks the accountability chain even where it's legal.
- When policy or contract says no. Client agreements sometimes prohibit AI processing of their material; some regulators restrict it. Boring, binding.
The gray zone — and the question that resolves it
Most real situations aren't in the five categories; they're 'can I use AI to help with X?' The resolving question: which part is mechanical and which part is the judgment? Drafting the layoff FAQ from decided facts: mechanical, fine. Deciding who's affected: judgment, never. Summarizing interview notes: mechanical. Concluding who to hire: judgment. Split every task this way and the gray zone mostly disappears.
I'm considering using an AI assistant for this task: [describe]. Argue both sides briefly: the case that this is an appropriate use, and the case that it isn't — considering verification ability, data sensitivity, accountability, and whether the human element is the point. End with the single question I should ask myself to decide.
Using the model as an ethics rubber duck is legitimate — it knows these frameworks. Just remember Module 1: it will agree with a stated preference, so ask neutrally.
When AI helped substantially with something that carries your name, the emerging professional norm is simple: be comfortable saying so if asked, and volunteer it when the audience would feel misled otherwise. If disclosure would embarrass you, that's usually the five categories talking.