A hallucination is when the AI states something false with complete confidence. Not a hedge, not a maybe. A specific figure, a regulation, a price, a quote from someone who never said it, delivered in exactly the same tone as everything it gets right.
That last part is the whole problem. Fluency and accuracy have nothing to do with each other in these systems. A wrong answer reads precisely as smoothly as a correct one, so your instinct for "this person sounds like they know" is useless here, and your instinct is what you have been using your whole life.
The response is not "never trust AI". It is knowing which of your tasks carry the risk, and having a checking habit for those. That gets you the 80 per cent where it is safe, at full speed.
Why it happens, briefly
These systems produce the most plausible continuation of the text so far. That is the mechanism. Most of the time the most plausible continuation is also the true one, which is why they work at all. Sometimes it is not, and nothing inside the system flags the difference, because it is not looking things up and then reporting. It is generating.
Once that lands, the pattern of when it goes wrong makes sense: it is worst on specific, verifiable, low-frequency facts. Exact figures. Recent events. Citations. Anything where being roughly right is the same as being wrong.
The risk map
| Risk | The work | Why | What to do |
|---|---|---|---|
| Low | Rewriting, shortening, restructuring, summarising something you pasted, brainstorming, drafting from your own notes | It is working from material in front of it | Read it. That is the whole check |
| Low | Explaining a general concept, talking a decision through, arguing with your reasoning | No specific facts being asserted | Read it |
| Medium | How-to steps for other software, general knowledge, industry norms | Usually right, often out of date, occasionally invented | Spot check anything you will act on |
| High | Any number you did not supply. Prices, rates, dates, statistics | It will produce a plausible figure rather than say it does not know | Verify every one, or supply them yourself |
| High | Tax, employment law, consumer law, licensing, insurance, anything regulated | Confidently wrong in exactly the places that cost money | Treat as a starting point for a question to a human, never as an answer |
| High | Quotes, citations, sources, statistics with a reference attached | Fabricated references are common and they look real | Click the link. If there is no link, assume it does not exist |
| High | Anything you will publish, print, or spend money on | The cost of being wrong is external | Full check before it leaves the building |
Most everyday business use sits in the top 2 rows. That is worth saying plainly, because people who read about hallucinations sometimes conclude the whole thing is unusable, and then hand-write emails for another year.
4 prompts that catch most of it
One is not enough, because they do different jobs.
1. The confidence audit
Run this after any answer with facts in it.
Go back over your last answer. List every factual claim you made. For each one, say how confident you are and whether it is something I should verify independently before relying on it. Where you can search the web and give me a source, do it. Where you cannot, say so plainly rather than guessing.
The claims it flags are your checking list. Models respond well to being asked about their own uncertainty, and the list is usually shorter than you fear.
2. The source test
For anything it has cited.
For every source, statistic or quote in that answer, give me the exact URL and the date. If you cannot produce a working link for one, remove that claim from the answer entirely and tell me which ones you removed.
The last clause is the one that works. Asking for sources gets you sources, some of which are invented. Asking it to remove what it cannot support gets you an answer you can use.
3. The adversarial pass
Best on plans, strategies and anything you are about to commit to.
Argue against your last answer. What is wrong with it, what have you assumed that I did not tell you, and what would make it fail? Be specific and do not soften it. Then tell me which of your original points still stand.
This is the one that catches the confident plan built on an assumption you never made.
4. The did-you-know-this test
Before you ask a factual question at all.
Before you answer: do you actually know this, or would you be inferring it? If your information might be out of date or you are unsure, say so first and then search rather than answering from memory.
Cheap, and it front-loads the honesty instead of making you audit afterwards.
Constructed to show the shape. A cafe owner asked what the current casual loading rate was and got a confident percentage with a reference to a Fair Work page.
Running prompt 2 returned: "I cannot produce a working link for the rate I quoted. I have removed it. Casual loading is set by the relevant modern award and I should not have stated a figure without checking the award that applies to you."
The figure it originally gave happened to be a number that has been correct at some point. That is the dangerous kind of wrong: familiar enough to pass.
The habit that beats all 4 prompts
Give it the facts yourself.
A supplied fact cannot be hallucinated. If you paste your actual price list, your actual policy, last month's actual numbers, the entire category of risk disappears for that task. The model stops needing to know and starts needing to read, which it is extremely good at.
This is why projects matter more than they look. Loading your real documents once means every conversation in that project is working from your facts instead of its guesses.
The short version: never ask it a question you could answer with a paste.
2 mistakes worth naming
Treating confidence as a signal. It is not one. There is no relationship between how sure an answer sounds and whether it is right. This is hard to internalise, because in humans the correlation is real.
One bad experience, total write-off. The people getting value are not the ones who never got a wrong answer. They are the ones with a checking habit, applied to the specific tasks that need it.
What the checks do not cover
The checks are not proof. A model can be confidently wrong about its own confidence. These prompts shift the odds a long way, and they do not replace verification on anything that matters.
Search does not make it true. Asking it to search reduces invention, and it can still misread a source or cite one that is itself wrong. Click the link on anything high risk.
None of this covers regulated advice. Tax, employment, legal and medical questions need a person who carries the liability. AI is useful for getting your question clear before you ask them, which is worth something, and that is where it stops.
Audit your last answer
Take the last AI answer you acted on that contained a number, a date, or a rule. Paste prompt 1 underneath it and see what comes back flagged.