How can I tell whether an AI answer is correct?
Check the answer one claim at a time against evidence you can inspect. Open the cited source, find the supporting passage, check its date and scope, then independently repeat any calculation. Fluent writing, a confident tone and a list of links are not substitutes for those steps.
NIST describes confabulation as false or erroneous content expressed confidently. OWASP also identifies misinformation and overreliance as risks in language-model applications. The useful response is a repeatable checking process, not a promise that a particular prompt or model will never make a mistake.
This guide is for checking an AI-generated explanation, comparison or draft before you use it. You can complete the exercise below without an AI subscription, uploading private documents or trusting a live price quote.
Start by splitting the answer into claims you can test
A sentence such as “The service launched in May, is available everywhere and costs $12 a month” contains three different claims. Verify the event date, availability and price separately. A source for the launch date might say nothing about the other two.
Use four simple labels: supported, contradicted, not established and calculation. “Not established” means you have not found sufficient evidence; it is not the same as proving the claim false. Mark predictions and personal recommendations separately from observations.
Check consequential claims first: the cost you will pay, whether a feature is available to your account, what a command changes, or what information an app stores. For a creative draft, stylistic choices can stay subjective; factual statements inside that draft still need evidence.
- Copy the exact claim rather than a vague topic.
- Identify the original source and relevant passage.
- Record the date, plan, region, version and units where applicable.
- Label uncertainty instead of filling gaps with a guess.
Does a real citation prove that the answer is supported?
No. First check that the reference exists. Then check that it says what the answer attributes to it. Finally check whether it is the right evidence for this question. A genuine pricing page for an old plan does not establish today’s price for a different plan.
Open the full source rather than relying on a search snippet. Search within it for the quoted words or the relevant figure. Read the surrounding qualification: “up to,” “in preview,” “for eligible accounts” and “excluding tax” can change the conclusion. If a document says a feature is planned, keep it planned in your answer.
If a link fails, search the publisher’s own site for the title or identifier. An April 2026 research preprint distinguishes apparently fabricated references from once-existing URLs that no longer resolve. A broken link alone does not prove fabrication. Conversely, a working URL alone does not prove that its content supports the claim.
Keep a short evidence note beside the claim. For a vendor feature, use current product documentation; for a research result, read the study and its limitations. For a financial figure, check the original filing or dataset and reporting period. Do not convert a vendor’s promotional claim into an independently established fact.
Do web search and RAG eliminate hallucinations?
Retrieval-augmented generation, or RAG, supplies retrieved material to a model when it prepares an answer. Anthropic’s explanation describes how documents are split into chunks and how missing context can make the wrong material appear relevant. Retrieval can help provide evidence, but it still requires checking the chosen source and the generated conclusion.
Ask the assistant to separate what it found from what it inferred. “The documentation states X” and “I infer Y from X” should not look identical. If it cannot access a page, it should say that rather than produce a supporting quotation from memory.
A second model can help identify questions to investigate, but agreement between two generated answers is not an independent source. Return to the underlying document, measurement or calculation. The same approach applies to outputs labeled research, reasoning or expert.
Five-minute practice: find three errors in a fictional AI answer
Use this fictional source card: the Cedar Notes trial is available to invited accounts only. It includes 200 requests. Additional requests cost $0.04 each. Taxes are excluded. The trial ends November 1, 2026. Cedar Notes is invented for this exercise, and these are not real product terms.
Now check this fictional AI answer: “Cedar Notes is available to everyone, and 350 requests cost $14 including tax.” Split it into availability, usage cost and tax treatment before looking at the solution.
Solution: availability is contradicted because invitations are required. The chargeable usage is 350 minus 200, or 150 requests; 150 multiplied by $0.04 is $6 before tax. The $14 figure charges all 350 requests and ignores the included allowance. “Including tax” is contradicted by the source card. No final tax-inclusive total can be calculated without the relevant tax information.
A supported rewrite is: “For an invited account within the trial period, 350 requests incur $6 in usage charges before tax under these fictional terms.” Notice that the correction preserves the conditions instead of turning a trial example into a general price.
A prompt that makes verification easier
Give the assistant the source material or ask it to retrieve accessible originals, then request an evidence table. Use the prompt below as a structure, not a truth guarantee. Check the returned passages yourself; asking for citations does not make invented references impossible.
Prompt to adapt: Check this draft against the supplied sources. For each factual claim, return: exact claim; source title and URL or document identifier; supporting passage and location; relevant date/version; verdict (supported, contradicted, or not established). Separate inference from reported fact. Recalculate numbers with units and assumptions. Do not invent links or quotations. If a source is inaccessible, mark it unverified. Finish with a corrected draft that retains important qualifications.
What extra checks do code and technical instructions need?
A command or package name that looks plausible may still be wrong. OWASP explicitly includes insecure or nonexistent software-library suggestions in misinformation risks. Check the exact package and maintainer against the project’s official documentation before installing it, and read what a proposed command will change.
For generated code, use a disposable test environment and sample data. Test expected input, missing input and a failure case. A successful test on one example does not establish correctness for every input, and merely importing a package does not prove it is trustworthy.
Keep this separate from prompt injection: an inaccurate answer can happen without an attacker, while an outside document can also try to redirect an assistant. Source checking and permission controls solve different parts of the problem.
When is the answer ready to use?
Use the answer when its important claims have traceable support, calculations reproduce, and unresolved limitations are clear for the intended task. If a key source is unavailable, remove the unsupported assertion or label the gap. A shorter answer with visible uncertainty is more useful than a complete-looking answer built from guesses.
For decisions with material consequences, verification may require a qualified person or a direct check with the relevant provider. Keep the evidence date so a later reader can see what was checked. If the product terms change, repeat the affected checks rather than assuming yesterday’s result still applies.
Sources & further reading
- NIST AI 600-1: generative AI risks including confabulation
- OWASP LLM09: misinformation and verification controls
- Anthropic: how retrieval works and why context matters
- April 2026 research preprint: fabricated references versus broken links
Educational content. Examples are illustrative. Protocols, products and rules can change; consult the linked documentation for current details.