Can AI make things up? Use this 5-step check before you trust an answer

One rule: treat AI as a lead, not proof

AI can help you frame a problem, list search terms, simplify dense text or turn a messy topic into a checklist. But if an answer involves names, dates, statistics, laws, research papers, news events, medical guidance, legal interpretation or money decisions, do not stop at the chatbot’s reply.

NIST — the U.S. National Institute of Standards and Technology — includes data provenance, auditing and assessment, monitoring and risk-based controls among the plans and actions associated with managing generative AI risk. For everyday users, that translates into a simple habit: look past the confident tone and follow the source trail.

The 5-step check

1. Ask for sources you can verify

Do not just ask for an answer. Ask the AI to connect each major claim to a source you can check:

List the sources that support this answer. Prioritize official documents, original research, government websites, company announcements or databases, and explain which claim each source supports.

If the answer relies on vague phrases such as studies show, experts say or many reports mention without a document title, publisher, date, link or searchable detail, treat that part as unverified. Source traceability matters more than a polished paragraph; NIST’s generative AI risk material includes data provenance and auditing or assessment among relevant risk-management actions.

2. Open the source and read the original

An AI-supplied citation is not a stamp of truth. Check at least three things:

• Does the link exist, and can you open it?
• Does the source actually contain the information the AI mentioned?
• Did the AI overstate, oversimplify or turn a limited point into a broader conclusion?

A common risk is that the answer appears to have a citation, but the cited material does not support the conclusion. Verification means moving from the AI says it has evidence to I have checked the evidence myself.

3. Check the details most likely to expose an error

You do not always need to read an entire report first. Start with details that are easy to verify and likely to reveal problems:

• People’s names and organization names
• Dates, years and version numbers
• Statistics, percentages and rankings
• Laws, policies and document titles
• Paper titles, authors and journal names
• Direct quotations

If the AI says a study found something, a company announced something or a regulation requires something, search for the original study, announcement or legal text. If you cannot find the primary material, do not treat the claim as confirmed.

4. Cross-check with independent sources

One source may be incomplete, and an AI summary may leave out important limitations. For higher-risk topics, look for at least one independent source that was not simply copied from the first one. This is especially important for:

• Medical and health information
• Legal, tax or contract questions
• Investment, finance and insurance decisions
• Breaking news and public events
• Technical specifications, security risks or policy requirements

If sources disagree, do not simply choose the version you prefer. Go back to the issuing organization, the original document, a professional database or a qualified expert with responsibility for the topic.

5. Do not let AI be the final decision-maker for high-risk questions

NIST’s generative AI risk-management approach emphasizes risk-based controls, monitoring and assessment rather than treating every output as equally trustworthy. You can use the same logic as an individual user: the higher the stakes, the stronger the verification needs to be.

If an answer could affect your health, legal rights, financial security, work decisions or public safety, use AI for background and question-building only. The final check should come from original documents or an appropriate professional, such as a doctor, lawyer, accountant, financial professional or subject-matter expert.

Red flags that an AI answer needs extra scrutiny

AI hallucinations are risky because they may not look like obvious mistakes. The answer can be fluent and well-structured while still being inaccurate. Pause before trusting an AI response if you see any of these signs:

• The tone is very confident, but no sources are provided.
• A citation looks complete, but you cannot find the original source.
• The source exists, but it does not support the AI’s conclusion.
• Numbers, dates or version details appear without attribution.
• The answer contradicts itself.
• The AI presents guesses, common assumptions or opinions as facts.

Copy-and-paste prompts for better verification

Use prompts like these to make an AI answer easier to check:

Turn your answer into a four-column table: factual claim / source / exact supporting passage / uncertainty.

Mark which statements are supported by sources, which are inferences, and which need manual checking.

Answer only from the documents I provide. If the documents do not say it, say that the information is not available.

List the five details in this answer most worth checking, such as dates, figures, quotations, policy names or people.

FAQ

If an AI answer includes citations, is it trustworthy?

Not automatically. A source can be nonexistent, irrelevant, outdated or misread. A proper check means opening the source, reading the relevant passage and confirming that the evidence supports the claim.

Do I have to verify every AI answer?

No. Match the amount of checking to the risk. A low-stakes brainstorming question may need only a quick sanity check; advice that could affect health, rights, money, safety or major work decisions deserves stronger verification. NIST’s generative AI risk materials use a risk-based approach that includes controls, monitoring and assessment.

What should I do if I cannot find the source?

Mark the claim as unverified. Do not cite it, forward it or make decisions on it as if it were fact. Ask the AI to answer again using only checkable sources, or search directly for official documents, original research, company announcements or trusted databases.

Bottom line: trust evidence, not confidence

AI can make things up. That does not mean you should never use it; it means you should not mistake fluent writing for proof. The practical workflow is straightforward: ask for sources, read the original, check the details, cross-check important claims and bring in human expertise for high-risk decisions. Used that way, AI becomes a faster research assistant — not an unchecked answer machine.