Back to course overview
Module 4Responsible AI 10 min

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.

  1. 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.
  2. 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.
  3. When the data can't leave. Regulated data without an approved tool. Full stop — see the last lesson.
  4. When accountability must be traceable to judgment. Firing decisions, disciplinary findings, credit approvals. Delegating judgment breaks the accountability chain even where it's legal.
  5. 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.

Prompt to try

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.

Disclosure norms

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.