Step 1: Sort claims by risk
Not all claims deserve equal scrutiny. A quick triage:
High risk — always check. Specific numbers, dates, statistics, citations, quotes, names, legal or medical claims, anything you'll act on or publish.
Medium risk — check if it matters. Descriptions of how something works, comparisons between products, historical sequences.
Low risk — usually fine. General explanations of well-established concepts, structure and organization, phrasing and tone.
The pattern: specific and verifiable = higher risk. General and conceptual = lower risk. Models are far more likely to invent a precise-sounding detail than to get the broad shape of a common concept wrong.
Step 2: Check citations properly
This is where AI output fails most often and most invisibly. A fake citation looks exactly like a real one.
Search for the exact title. If a paper, check whether it exists in a real database. If a URL, open it — don't assume a plausible-looking URL resolves. If the source exists, confirm it actually says what was claimed, because a real source attached to a claim it doesn't support is also a failure.
Step 3: Verify independently, not with the same AI
Asking the same model "are you sure?" is close to useless — it may simply agree with itself, or reverse course without either answer being grounded in anything. Verification has to come from outside the system: a search engine, primary sources, documentation, or a person who knows the field.
Step 4: Watch for confident specificity on obscure topics
The combination worth noticing: a niche question answered instantly with precise details and no hedging. Genuine expertise on obscure topics usually comes with caveats. Fluent precision without caveats on something rarely written about is a signal to check.
Step 5: Ask it to flag its own uncertainty
Adding "flag anything you're not confident about, and say explicitly where you're unsure" to your prompt often surfaces weak points upfront. It's not a guarantee — a model can be wrong about its own confidence — but it's a cheap extra signal.
The time-efficient version
If you only have time for one thing: check every specific number, date, and citation. That single habit catches the large majority of consequential AI errors.
Why this matters more than it sounds
The risk with AI output isn't that it's obviously wrong — obviously wrong is easy to catch. It's that it's fluently wrong, which slides past exactly the instincts that normally protect you. A deliberate checking habit substitutes for instincts that don't fire here.
Related reading
See AI Hallucinations Explained for why this happens, and the Hallucination glossary entry for a quick definition.
A proportionate approach
Checking everything exhaustively would erase the time saved, so the honest method is to scale checking to consequence.
No checking needed. Rewording your own text, formatting, brainstorming, explanations of things you already understand well enough to spot an error. The risk is near zero.
Quick check. Anything you will repeat to someone else. One search to confirm the central claim, thirty seconds.
Careful check. Anything going into work you are accountable for — a report, a client email, published writing. Every number, name, date, and citation verified at source.
Do not rely on it at all. Medical, legal, and financial decisions. Not because the answer will definitely be wrong, but because being confidently wrong there is expensive in ways that cannot be undone.
What to check first
Not everything in an answer carries equal risk. In order:
- Citations and links. Highest fabrication rate of anything. Open them. A dead link or a real paper that says something different is the most common finding.
- Numbers. Statistics, dates, percentages, quantities. Precision is easy to generate and hard to eyeball.
- Names and attributions. Who said or did a thing gets confidently mixed up, particularly among similar figures.
- Causal claims. "X happened because of Y" is often a reasonable-sounding guess presented as established.
- Anything conveniently supporting what you wanted. Models are agreeable. If the answer is exactly what you hoped, that is a reason for more scrutiny, not less.
Practical techniques
Search the exact claim. If a specific statistic returns nothing, it very likely does not exist.
Go to the primary source. Not a page repeating the claim — the study, the filing, the original report.
Ask in a fresh conversation. Without the earlier context, a genuine fact usually survives and an invention often does not.
Ask what would make it wrong. "What are the strongest objections to this?" produces more useful checking than asking whether it is sure.
Check the date. Much wrong output is outdated rather than invented, which is easier to spot once you are looking for it.
The part people skip
Verification is part of the cost of using these tools, not an optional extra. A task is only faster with AI if it is faster including the checking — and for some tasks it genuinely is not.
Being honest about that is what separates using AI well from using it fast. See AI Hallucinations Explained for why this is a permanent feature rather than a temporary flaw.