AI Guides › Judgement & Guardrails
What "Hallucination" Actually Costs You If You Don't Catch It
By Nigel Guy · 2 min read
Most people treat a hallucinated answer as a funny story — the AI invented a
court case, made up a statistic, insisted a person held a job they never
held. Told as an anecdote afterwards, it's harmless. The actual cost isn't
the invention itself; it's however far that invention travelled before
someone caught it, and that distance is the thing worth paying attention to.
The rule: the cost of a hallucination is not fixed — it multiplies with
every step it takes before you catch it, so the only real question is where
in that chain you're planning to catch yours.
The mechanism: the cost ladder
Every hallucination sits somewhere on this ladder. The cost of the same
underlying error is completely different depending on the rung.
- Caught before use. You asked, checked, found it was wrong, moved on.
Cost: a few minutes. This is the target outcome, not a bonus one.
- Caught at use, before sharing. You acted on it briefly — drafted an
email, built a section of a document — then caught it on review before
anyone else saw it. Cost: rework time, nothing else.
- Caught after sharing, before anyone relied on it. A colleague, client,
or reader saw the error, and it happened to not matter yet, or they
flagged it. Cost: a correction, some credibility, an awkward message.
- Caught after someone acted on it. A decision got made, money moved,
a claim got repeated further, on the strength of the wrong thing. Cost:
whatever that decision cost to make, plus whatever it costs to unwind.
- Never caught. The error is now load-bearing somewhere — in a filed
document, a public claim, a system built on the assumption it was true.
Cost: unknown, and unknown is the worst version of this list, not the
best.
Nothing on this ladder is about how the hallucination happened or how
sophisticated the model was. It's entirely about which rung you stopped it
on, and that's the part within your control.
What to skip
Skip trying to predict which topics or prompts are hallucination-prone
before you start — every topic is a candidate, and building a mental list of
"safe" categories just moves your blind spot rather than removing it. Skip,
too, treating a hallucination you caught early as proof the tool is
generally reliable enough to skip the same check next time. Each output
earns its own check; the last one being fine doesn't transfer.
Guardrails
- This ladder assumes you eventually catch the error somewhere. It says
nothing about the errors that are never caught at all — those don't show
up on anyone's cost accounting, which is exactly why they're worth taking
seriously rather than dismissing as rare.
- The higher up the ladder you catch something, the less it costs — so the
actual lever here is where in your own workflow you put the check, not
how good you get at spotting invented content by feel.
- Anything already shared, filed, or acted on deserves a real look back, not
just a resolution to check more carefully going forward. If you suspect
something slipped through, it's worth the time to trace it.
All 751 AI guides · JulieMango plans from £17/mo