That AI answer has sources. Do they actually support it?

01 / Tech Ka Asli Sach

A source can be real and still be the wrong evidence

Imagine this buying question: “Can I edit my college video project on this laptop?” An assistant replies with a confident yes, a memory figure and a link to an operating-system requirements page. The link opens. The publisher is genuine. The answer still has a problem.

The source might tell you what the operating system needs to run. It does not, by itself, settle how a particular editing project will perform. This is an illustrative example, but the verification problem is ordinary: a relevant-looking page is doing more work in the answer than its contents justify.

That is where to begin. You do not need to decide whether an entire AI system is trustworthy. You need to check the few statements your decision depends on.

NIST identifies confabulation, including errors involving citations, as a generative AI risk. [1] Treat a citation as a route to evidence, not a stamp of approval.

02 / Tech Ka Asli Sach

Pull one claim out of the paragraph

A sentence such as “This laptop is cheaper, supports more memory and will last longer” contains three claims. They need different evidence. A manufacturer specification can help with supported memory. It cannot establish today’s checkout price or predict how long your particular machine will remain useful.

Copy each decision-critical claim into its own row. Leave out decorative facts you do not need. If you are choosing a storage drive, you may need compatibility and warranty details; a paragraph about the history of the manufacturer is unlikely to change the purchase.

Use plain verdicts: supported, partly supported, unsupported, or not checked. “Unsupported” means the inspected source does not establish the claim. It does not automatically mean the claim is false.

This distinction keeps the exercise honest. You are testing the evidence you have, rather than manufacturing certainty in either direction.

03 / Tech Ka Asli Sach

Open the page and find the passage

Start with the actual document, not the search snippet or the assistant’s quotation. Find the relevant words, then read the sentences around them. Check the publisher, date, version, country and product model. If the page summarizes a study, follow the reference to the study when its findings matter to your decision.

Suppose a quoted policy applies to a business plan. The answer applies it to the free plan you use. Your ledger should record that mismatch, even if the quotation itself is accurate. A true statement about the wrong account is poor buying advice.

For a specification or a service policy, prefer the organization responsible for publishing it. For a judgment about value, seek evidence beyond the seller’s description. Being authoritative about a product’s ports is different from being an independent judge of whether it is worth your money.

Google Research describes work on grounding generated answers in sources and discusses limitations of adding citations after generation. [2] For a reader, the practical lesson is to inspect the connection between the claim and the passage. The presence of a reference is only the beginning.

04 / Tech Ka Asli Sach

Keep a ledger you could hand to someone else

Your worksheet needs four columns: the exact claim, the supporting passage or section, the scope, and your verdict. Add a checked date for facts that change. Write enough context for another person to understand why you accepted or rejected the claim without repeating the whole search.

For the laptop example, a useful row would say: “Editing suitability; source establishes an OS installation minimum; scope does not cover our editing workload; not established.” That row tells you what to research next. A saved URL alone does not.

When a page cannot be opened, mark it not checked. Ask for an accessible primary source or remove the claim. Do not let the assistant paraphrase the same inaccessible page more confidently and call that verification.

Save the ledger beside your draft. If you later correct a claim, the note helps you find the article, script or carousel that repeated it.

05 / Tech Ka Asli Sach

Give the assistant a narrower job

Try this instruction: “Separate factual claims from suggestions. For each factual claim, give the source URL, the passage that supports it and any limits on its scope. Mark missing evidence. Do not invent a quotation or fill a gap with a guess.”

Then do the opening-and-reading step yourself. A better prompt can make the output easier to check; it cannot make that check unnecessary.

If you already have reliable documents, give the assistant that small source set and ask it to stay within them. Require a list of questions those documents cannot answer. This is particularly useful when the task is summarizing a brief rather than researching the entire web.

Before publishing, check the facts that change quickly again. Prices, eligibility, service limits and model-specific specifications deserve more care than an evergreen explanation. For consequential medical, legal or financial decisions, source checking is only one part of getting appropriate professional guidance.

Finish with one question: can you point to the passage that supports the claim you are about to repeat? If not, the answer is not ready to leave your notes.

Claim-checking worksheet

Sources & revision note

Primary sources checked 2026-11-14. Practical frameworks are editorial guidance; illustrative examples are not measured results.

1. NIST: Generative AI Profile
2. Google Research: grounding and attribution

Revision 2 · 2026-11-14: Rewritten with clearer examples, a practical worksheet and a stronger decision checklist.