You read it. But is it true?

A citation can make a sentence look as though it has passed a background check. Sometimes the sentence has barely met the source.

An AI answer may be useful, accurate and well supported. It may also sound equally polished when it is wrong. The challenge is to tell the difference without turning every simple question into a three-day investigation.

Start with the claim that matters to your decision. Then ask what evidence would actually support it. The confidence of the delivery can wait its turn.

Fluency does not verify a fact

Generative AI can produce confident statements that are false or inconsistent. NIST calls this risk confabulation, often described as hallucination. Its guidance also warns that invented citations and plausible explanations can make an incorrect answer more convincing. [1]

That matters because reading a smooth explanation is easier than checking it. An answer can connect ideas beautifully while getting a date, specification or condition wrong. Asking the same assistant “Are you sure?” may produce more explanation. What you need is evidence you can inspect.

Open the citation

Check two things: whether the source exists, and whether it supports the specific sentence. A real article about a subject may still fail to establish the claim attached to it.

A 2023 study of literature reviews generated by GPT-3.5 and GPT-4 found both fabricated references and errors in citations to real works. Those findings describe the versions and tasks tested, not an error rate for every AI tool today. They do establish why a convincing bibliography deserves a check. [2]

Look beyond the title. Read the relevant passage, including the qualifications. Notice whether the answer has quietly promoted “may help” to “proven to work,” or turned one narrowly defined result into a universal rule. If the original source is unavailable, keep that limitation visible.

Make the question smaller

Consider a hypothetical shopping question. An assistant says a travel mug is dishwasher-safe. A retailer’s description agrees. You open the manufacturer’s care instructions and find that the body can go in the dishwasher, but the lid must be washed by hand.

The answer needs a condition, not a debate about which website is generally more trustworthy. “Can I put the whole mug in the dishwasher?” is precise enough to resolve. Confirming the model and current instructions matters more than collecting five additional shopping summaries.

The same habit helps elsewhere: identify the exact product, date, location or population involved. Two sources may appear to disagree because they are answering different questions.

Look outside the first source

When you encounter an unfamiliar website, check who is behind it and what credible independent sources say about it. Researchers Sam Wineburg and Sarah McGrew observed professional fact-checkers doing this: leaving the original page to investigate elsewhere, rather than judging credibility from its appearance. They called the approach lateral reading. [3]

Also check whether supposedly separate sources are independent. Several articles repeating the same announcement still trace back to one announcement. A company’s manual may be the right source for its product instructions; an independent test may be more useful for evaluating its performance claims.

Disagreement deserves examination, but it does not automatically mean the truth lies halfway between two positions. Compare the methods, dates and strength of the evidence. Give a weak claim less weight, even when it has excellent typography.

Give the assistant a checking job

Try a request like this:

“Check this specific claim. Identify the strongest relevant sources and link to the passages that support it. Separate direct evidence from inference. Look for meaningful contradictory evidence and explain any differences in date, scope or method. Tell me what remains unverified. If you cannot access a source, say so.”

Then open the important links yourself. A second AI can offer another perspective, but agreement between tools is not independent confirmation of the underlying fact. Trace both answers back to their evidence.

Match the checking to the consequences

A caption idea and a medication instruction do not deserve the same level of scrutiny. For decisions involving health, legal rights or substantial money, verify against authoritative current information and get appropriate professional review. NIST highlights the particular risks of confabulation in consequential decisions. [1]

You do not need to distrust everything. You need to know what supports the answer you are about to use. Sometimes the most useful result is a narrower claim, a missing detail to confirm, or a clear “we do not know yet.”

A good answer makes the next check easier. The footnotes should earn their keep.

Sources

[1] NIST. Generative Artificial Intelligence Profile

[2] Walters and Wilder 2023. Fabrication and errors in the bibliographic citations generated by ChatGPT

[3] Wineburg and McGrew 2019. Lateral Reading and the Nature of Expertise