How to Check an AI Answer Before You Pass It On

An AI tool will hand you a confident answer whether or not it has any basis for one. That is the whole problem, and it is not solved by asking the tool whether it is sure. This guide is a checking routine in plain English: which parts of an answer are worth checking first, how to tell a verifiable claim from an unverifiable one, what to do with numbers, names and quotations, and where to stop so the checking does not cost more than writing it yourself.

Why a confident answer is not a checked answer

The core thing to understand is that fluency and accuracy are produced by the same process, so they arrive together and feel identical.

An AI tool writes the most plausible next piece of text. A true statement is usually plausible, which is why these tools are useful. But a false statement can be equally plausible, and nothing in the writing of it flags the difference. There is no internal moment where the tool knows it has run out of knowledge and starts improvising.

This is why the usual human cues for doubt do not work here. With a person, uncertainty leaks: hedging, vagueness, a change of subject. An AI answer is often at its most specific and well-organised exactly where it is inventing, because invention is unconstrained by the awkward shape of real facts.

So the routine below does not try to detect doubt in the text. It sorts the answer by what kind of statement each part is, and checks the kinds that can be checked.

• Confidence carries no information about accuracy in a generated answer.
• Specific and tidy is not safer than vague - often the reverse.
• The tool cannot tell you where its knowledge ended, so it will not warn you.

Sort the answer into three kinds of statement

Before checking anything, read the answer once and mentally mark each claim as one of three types. This takes under a minute and it decides where all your effort goes.

Checkable facts are statements that something is the case in the world and could be confirmed at a source: a figure, a date, a name, who owns what, what a law requires, what a document says. These are the ones to check, and they are the ones that cause damage.

Reasoning is the argument connecting the facts: because this, therefore that. You do not check reasoning at a source; you read it and judge whether it follows. Flawed reasoning is common and usually visible on a careful read.

Opinion and framing is everything else - what matters most, what the reader should care about, how to characterise a situation. There is nothing to verify. Your only job is deciding whether you agree enough to put your name to it.

The mistake most people make is spreading attention evenly across all three. Almost all of the risk sits in the first category.

• Checkable fact - go to a source.
• Reasoning - read it and test whether it follows.
• Opinion or framing - decide whether you endorse it.

Check numbers at the source, never in the answer

Numbers deserve their own step because they are the most trusted and least reliable part of a generated answer.

A number in an AI answer can be right, stale, or invented, and all three look the same. Stale is the quietly dangerous one: a figure that was correct two years ago is not flagged as historical, and prices, limits, rates and thresholds move constantly.

The rule is that a number is unverified until you have seen it on a page that is responsible for it. The organisation that sets the figure publishes it, and that page is the source. A number repeated in an article about the organisation is a copy, and copies inherit errors.

When you confirm a number, write down where you saw it and when. A figure without a source and a date is not usable in anything that someone else will rely on, because neither you nor your reader can tell later whether it was ever true.

Be especially careful with anything that reads like a calculated result - a percentage, a total, an average. Check the inputs and redo the arithmetic yourself. Generated calculations are frequently plausible and wrong in ways that survive a glance.

• Find the body that sets the number and read it there.
• Record the source and the date you read it.
• Redo any arithmetic yourself from verified inputs.
• Assume undated figures are stale until shown otherwise.

Treat every name, title and quotation as a claim

Names are the second high-risk category, and the failure is distinctive: real components assembled into something that does not exist.

A generated citation can combine a genuine author, a genuine publication and a title that was never written. A quotation can be attributed to someone who holds the view but never said the words. A product can be credited with a feature belonging to a competitor. Each of these is wrong in a way that survives casual checking, because searching the author's name finds a real person and confirms nothing.

So check the whole unit, not the parts. The question is not whether the author exists but whether that author published that thing with that title. For a quotation, the question is whether the words appear, in that order, in a source you can open.

If a quotation cannot be traced to a specific document, do not paraphrase it into safety. Remove it. A paraphrase of a fabricated quotation is still a fabricated claim about what someone thinks, with the evidence trail removed.

The same applies to anything an AI answer says a document contains. Open the document and read the passage. A summary of a real document can misstate it completely while naming it correctly.

• Verify the whole citation, not just that the author is real.
• Find quotations verbatim in a document you can open, or cut them.
• Read the passage yourself when told what a document says.

Read the reasoning for the gap between the facts and the conclusion

Once the facts are checked, the remaining risk is an argument that sounds orderly but does not hold. This is a reading task, and a few patterns account for most of it.

The most common is a conclusion wider than the evidence: three specific examples supporting a general rule, or a finding about one group presented as applying to everyone. The facts survive checking individually; the leap between them does not.

The second is a missing alternative. An answer explains why something happened with a cause that is sufficient but not established, and never mentions the other explanations that fit the same facts.

The third is a definition that quietly changes. A term means one thing in the second paragraph and something broader by the fifth, and the conclusion depends on the broader sense.

The fourth is a buried assumption presented as a given - usually a clause beginning with an obviously or a clearly, which is where unsupported premises prefer to live.

The test to apply is simple: cover the conclusion, read only the evidence, and ask what you would conclude from it. If you would not reach that conclusion yourself, the gap is the answer's, not yours.

• Conclusion broader than the evidence supports it.
• An alternative explanation that fits the facts and is never addressed.
• A term that shifts meaning between the setup and the conclusion.
• Assumptions wearing the word obviously instead of a source.

Watch for what the answer left out

Checking what is on the page misses a whole class of problem, because the most consequential error in a generated answer is often an omission.

An AI answer gives you a complete-sounding response to the question you asked. It does not tell you that your question left out the exception that applies to your situation, or that there is a legal requirement, a safety consideration or a cost you did not ask about and therefore did not get.

This matters most in exactly the areas where people most want a quick answer: anything involving money, law, health, employment, tax or safety. In those fields the omitted qualification is frequently the whole answer.

Two habits help. First, ask explicitly what would change this answer - what conditions, exceptions or jurisdictions make it wrong. You are not trusting the reply; you are using it to generate a list of things to check. Second, before you rely on the answer, ask yourself what a person who does this for a living would have told you that the answer did not.

If the subject has real consequences, a generated answer is a way to arrive at a competent question for a professional. It is not a substitute for the professional.

• Ask what would make this answer wrong, then check those conditions.
• Assume exceptions exist in anything legal, medical, tax or safety related.
• Use the answer to form the question you take to someone qualified.

Match the checking to what the work has to carry

The reason checking routines get abandoned is that they are written as though everything deserves the same scrutiny. It does not, and pretending otherwise guarantees that nothing gets checked.

Work that only you will see needs essentially no verification. A draft you are about to rewrite, a list of angles, a summary to orient yourself - the errors cost nothing because nothing is built on them. Checking here is wasted effort, and treating it as mandatory is what makes people give up on the habit entirely.

Work that an internal colleague will read needs the specifics checked and nothing more. Numbers, names and dates at source; skim the reasoning; leave the framing.

Work that leaves your organisation - a client deliverable, a published page, anything a decision rests on - needs every checkable claim verified at source, every quotation traced, every calculation redone, and anything unverifiable removed. This is also the category where it is worth asking whether generating it was the efficient route at all.

And there is a fourth category: regulated or high-consequence output, where a qualified human has to be accountable. There the generated text is raw material at most.

• Only you will see it - check nothing, rewrite freely.
• A colleague will read it - check numbers, names and dates.
• It leaves the building - verify every specific claim at source; cut what you cannot.
• Someone is accountable for it - a qualified person owns the output, not the tool.

A routine short enough to actually use

Put together, the whole thing is six passes and most of them are fast. The order matters, because each pass makes the next one cheaper.

First, decide what the work has to carry, using the four categories above. This sets your budget and prevents both over- and under-checking.

Second, read once and mark the checkable facts. Do not verify yet - just find them.

Third, verify the specifics at source: numbers where the setting body publishes them, citations in full, quotations verbatim, calculations redone.

Fourth, cover the conclusion and read the evidence alone to test whether the reasoning holds.

Fifth, ask what is missing - exceptions, jurisdictions, costs, the thing a professional would have said.

Sixth, delete or label everything still unverified. This is the pass people skip, and it is the one that keeps an unsupported claim from becoming yours.

If you want the other half of this - producing an answer that needs less checking in the first place - see how to write a prompt that gets a usable first draft and how to ask an AI tool for what you want. A request that specifies what it wants and what it will not accept produces fewer invented specifics to catch.

• Decide what the work has to carry.
• Read once; mark the checkable facts.
• Verify specifics at source.
• Test the reasoning without the conclusion.
• Ask what was left out.
• Cut or label whatever is still unverified.

What to keep a record of

One habit makes everything above durable: write down what you checked, where, and when.

This sounds like bureaucracy and takes about twenty seconds per claim. The payoff arrives the first time somebody questions a figure months later. With a note, the answer takes a moment. Without one, you cannot tell whether you verified it, assumed it, or inherited it from a generated draft - and the honest answer becomes that you do not know.

It also protects against the slow failure mode of working this way: a checked fact becomes a remembered fact, and a remembered fact gets reused in another document a year later, by which time it may no longer be true. A date next to the figure makes staleness visible.

Keep it minimal. The claim, the source, the date you read it. That is enough to re-check quickly, and enough to notice when something has aged.

If you are also thinking about what happens to the text you put into these tools in the first place, how to get your data out before you cancel covers the other end of the same relationship.

How we chose

This guide is method rather than measurement. It names no tool, quotes no price and describes no specific button or screen, because interfaces change every few weeks and a wrong instruction reads exactly like a right one. What does not change is the shape of the problem: which kinds of statement can be checked, which cannot, what order to check them in, and how much checking a piece of work actually deserves. Every section tells you what to look for in the answer in front of you rather than what any particular product does.

Frequently asked