AI assistants write fluent, confident text whether or not they are right. Fluency is not evidence. If an answer will influence a decision, a grade, a purchase, a medical choice or something you will say to other people, spend a few minutes checking it. This is a routine that works with any assistant.
Step 1: decide how much checking the answer deserves
Match effort to consequences. A synonym for "happy" needs no checking. A dosage, a tax rule, a legal deadline, a statistic you will quote, a piece of code that touches production data or an email to your boss deserves real attention. A useful test is: what happens if this is wrong, and who pays?
Step 2: separate claims from prose
Read the answer and underline the parts that could be false: names, dates, numbers, quotations, citations, legal or medical statements, "always" and "never". Smooth explanation is easy to judge by whether it makes sense. Specific, checkable claims are where invented details tend to hide.
Step 3: check the claims that carry the weight
For each important claim, find a primary or authoritative source yourself:
- Numbers and dates: the original report, the statute, the company filing, the official site.
- Quotes: the original text. Models sometimes produce plausible quotations that were never said.
- Citations and links: open them. Confirm the page exists, says what the answer says it says, and is from who it claims.
- Laws, medicine, taxes: the regulator, the agency, the clinical guideline, a qualified professional. Rules differ by place and change with time.
- Code: run it, test it, read it. A passing run on one input is not proof.
If a claim cannot be traced to anything you can open, treat it as unverified.
Step 4: test with a second, different source of reasoning
Ask the question again in a different way, or ask a different model. Compare the substance, not the wording. If two independent models give the same specific answer, that is a reason to take it more seriously. If they differ, you just learned where to dig. The key word is independent: a second sample from the same model, with the same training and the same blind spots, tells you less than a model from a different family.
Step 5: look for the usual failure signs
- Very specific details with no source.
- Perfectly formatted references that you cannot find.
- Confident answers to questions that depend on today's date or breaking news, from a model with no live lookup.
- Arithmetic done in prose rather than computed.
- An answer that matches exactly what you hoped to hear. Models can lean toward agreeing with the asker.
- Silence about uncertainty on a question where experts disagree.
Step 6: ask the model to attack its own answer, with limits
A prompt like "List the three weakest claims in your answer and what would show each is wrong" can surface soft spots. It is a prompt for finding things to check, not a substitute for checking them; the same model can be confidently wrong about its own weak points.
Step 7: record what you verified
If the answer matters, note what you checked and where. A short line such as "figures verified against the 2025 annual report, quote not verified" prevents you or a colleague from treating all of it as settled later.
A worked mini-example
Suppose you ask an assistant when a particular filing deadline falls and get a specific date. You would (1) mark the date and the rule it relies on as the load-bearing claims, (2) look up the rule on the agency's site, (3) check whether the answer assumed the right year and jurisdiction, and (4) ask a second model and see whether it names the same date and rule. If both match the agency page, you can rely on it. If the models differ, the agency page decides, and you have learned the question was less settled than the first answer implied.
What checking cannot do
Even a careful check has limits. You can miss an error you do not know to look for. A source can be wrong. Some questions have no single correct answer. And checking takes time, which is why matching effort to stakes in Step 1 matters.
FAQ
Can I just ask the AI whether its answer is correct?
You can, but treat the reply as another claim to check. A model that produced an error can repeat it when asked to confirm.
Does asking two AIs guarantee a correct answer?
No. Two models can share the same mistake, especially when their training data overlaps. Agreement raises your attention level; it does not prove correctness.
Are AI citations reliable?
Not automatically. Always open the link and confirm it supports the specific claim.
How Keplar approaches this
Keplar builds part of this routine into the answer, and leaves the rest to you. For a non-trivial question it asks several models from different families, compares what they actually claim, and flags where they disagree. A reviewer model then reads the draft against those responses and marks points it held, revised, removed or kept as a minority view. Exact arithmetic and counting are computed by code rather than guessed.
What Keplar does not do is open or verify sources. Links that models cite are shown labeled "Cited by <model>; Keplar didn't open or check", and the answer carries no confidence or accuracy score. So Steps 3 and 7 still belong to you. The Consensus and Disagreements sections tell you where to look first. See how verification works and what consensus means.