AI Guides › Judgement & Guardrails
When To Not Use AI At All
By Nigel Guy · 2 min read
Most guides about AI, including most of the ones in this library, are about
using it well. This one is the necessary counterweight: a short, honest list
of situations where the right move is to not reach for it, not to use it
more carefully.
The rule: some situations aren't a "how to prompt this well" problem —
they're a "don't" problem, and no amount of careful prompting fixes that.
The short, real list
- Anything requiring licensed professional judgement with legal weight —
a contract you're about to sign, a diagnosis, a tax filing, specific
investment advice. Useful as a first pass to understand the shape of the
thing; not a substitute for the qualified person who's actually
accountable for getting it right.
- A moment that needs your presence, not your output. Comforting
someone through a hard moment, being physically present, showing up. The
task-shaped part of the situation might be automatable; the actual point
of it usually isn't.
- Anything where you can't verify the answer and getting it wrong is
expensive. If you have no way to check whether the output is right, and
being wrong costs real money, safety, or trust, that's a signal to get a
second, verifiable source — not to prompt harder.
- Decisions you're using it to avoid making yourself. Asking for a
second opinion is healthy. Asking it to make the call so you don't have
to own the decision is a different thing, and it doesn't actually remove
your responsibility for the outcome.
- Anything you'd be embarrassed to admit you didn't do yourself, where
the doing was the actual value — a hand-written note, a genuinely
personal piece of writing, work you're claiming as entirely your own
craft.
What this isn't
This isn't an anti-AI list. It's a short set of situations where the honest
answer is "not this one," sitting alongside the much larger set of
situations where AI genuinely helps. Most of this library is about that
larger set.
Guardrails
- This list is a starting point, not exhaustive — the underlying test is
simple: can you verify it, and does it matter if you can't?
- "I don't have time to do this properly" is not the same category as "this
genuinely shouldn't be automated" — the first is a scheduling problem, not
a guardrail.
- Revisit this list as your own judgement develops. What belongs on it
changes as your ability to verify AI output in a given domain improves.
All 751 AI guides · JulieMango plans from £17/mo