Spot and reduce hallucinations in your workflow
Where fabrication concentrates, what actually reduces it, and what only appears to.
Before you start
- Any workflow where output reaches other people
What you will be able to do
- Predict which parts of an output are worth checking
- Apply the two techniques that measurably reduce fabrication
- Stop relying on confidence as a quality signal
Treating every sentence as equally suspect is exhausting and people stop doing it within a week. Treating none as suspect is how a fabricated figure ends up in a board pack.
The workable middle relies on the fact that fabrication is concentrated rather than uniform.
Know where it concentrates
Numbers, names, dates, citations, quotes, URLs. Specific and verifiable.
Fabrication clusters in exactly the material that looks most authoritative: statistics, proper nouns, dates, references, direct quotes, links.
General explanation is comparatively reliable, because it is drawn from material repeated thousands of times. The specific detail is the risk, and it is also the part readers most trust.
Give it the source instead of asking it to recall
Grounded in supplied text, fabrication drops sharply.
Recall from training is where invention happens. Summarising a document you pasted is a fundamentally different and far more reliable task.
Whenever a workflow can supply the source — paste the page, attach the file, use a search-grounded tool — do so. This is the single biggest reduction available, and it is bigger than any prompt technique.
Explicitly permit not knowing
The default is to produce something. Override it every time.
Add a standing instruction: if you are not confident, say so and say what you would need. Unprompted, a model fills the gap, because producing a plausible answer is what it does.
It will not volunteer uncertainty on its own, and the improvement from one sentence is disproportionate to the effort.
- Keep the permission in your reusable prompts rather than adding it ad hoc. The times you forget are the times it matters.
Warning Stop reading confidence as accuracy
The register is identical whether it knows or is inventing.
There is no tonal difference between a fact and a fabrication. Both arrive fluent, well-formed and unhedged, and human judgement of AI output correlates with fluency rather than with truth.
Asking "are you sure?" is also weak: it often triggers a change of answer rather than a genuine assessment, which tells you the model is agreeable, not that the first answer was wrong.
- Using a second model to verify the first. Shared training data means shared blind spots, and you get a confident second opinion on the same error.
Check specifics, ground what you can, and let it say it does not know. Confidence is not a signal — it is the default register for everything.
Common questions
Was this guide useful?
84% of readers found this useful
Read next
Handle personal data in AI tools without breaking the law
Most AI data problems are not exotic. They are a consumer account processing customer records, and they are avoidable with four de…
Disclose AI use in your work the right amount
Disclosure norms are unsettled and the test that works is simple: would this person feel misled if they found out how it was made?
Build your first agent that does one thing reliably
Agent demos look magical and agent deployments mostly fail on the same three things: unbounded loops, unverified tool output, and…
Estimate what an AI feature will cost you per month
Per-token pricing looks trivially cheap and routinely surprises people at the invoice. The gap is almost always retries, context a…
