In this article
To check an AI answer, verify its important claims against sources you open yourself. A confident tone, a plausible link or agreement between two chatbots is not proof. Check both whether the source exists and whether it supports the precise statement.
This guide gives you a repeatable verification process for research, work and everyday technical questions. It is a documentation-based editorial workflow, not a benchmark of individual AI products.
Start with the claims that could change your decision#
Do not spend equal effort on every sentence. Pull out the claims that affect what you will buy, install, submit or believe:
- Numbers, dates and eligibility requirements.
- Product features, prices and compatibility.
- Quotations and research findings.
- Commands that modify data or permissions.
Suppose an answer says: “The app is free, works offline and exports encrypted backups.” Those are three claims. A page confirming the free plan does not establish the other two.
Create a short evidence ledger. For each claim, record the source URL, the relevant passage, the date you checked it and your decision: supported, contradicted or unresolved. That final category prevents uncertainty from quietly becoming a fact.
Find the source independently#
Search for the product, document or paper yourself. Prefer the current official manual for a feature, the provider's pricing page for a price and the original paper for a research finding.
When an AI supplies an academic reference, check its title, authors, publication venue and identifier. Crossref Metadata Search lets you look up bibliographic metadata. A matching record establishes identity; it does not prove that the paper supports the AI's conclusion.
For example, an answer might attribute a battery improvement to a paper whose actual subject is charging efficiency. Even if the authors and DOI are real, the conclusion still needs checking.
Research on citation hallucination detection addresses the problem of plausible references that do not correspond to real works. Treat a reference as a lead to investigate, not evidence you have already verified.
Read beyond the headline#
Check who or what the source describes. A study of one device does not automatically apply to all laptops. A feature available in an enterprise plan may be absent from a personal account.
Ask four questions:
- Does the passage address the same claim?
- Does it apply to the same version, market or audience?
- Is it current enough for this decision?
- Does the answer preserve the source's limitations?
A number without its denominator can be particularly misleading. “Half of respondents” is not necessarily “half of all users.” Likewise, a forecast is not a measured result.
Resolve conflicting answers without taking a vote#
Several AI tools may repeat the same outdated or mistaken source. Instead of counting agreement, compare the underlying evidence.
If a recent changelog conflicts with an older tutorial, check whether the behavior changed. If official pages disagree, keep the claim unresolved and contact the provider before relying on it for a consequential action.
You can ask an assistant to list uncertainty or separate facts from assumptions. That helps organize the work, but it does not replace verification.
Use a stopping rule#
For a low-stakes definition, one clear authoritative explanation may suffice. For a purchase, check the exact product configuration and current terms. For a command that could delete files, inspect the command's official documentation and understand the effect before running it.
Stop when the important claims are supported and the remaining uncertainty is acceptable for your decision. If a crucial claim has no accessible evidence, leave it out or label it unverified.
Common questions#
Does a citation make an AI answer trustworthy?#
It makes the answer easier to check. The citation can still be invented, outdated or unrelated.
Can I ask another AI to fact-check it?#
Use another tool to find possible errors, then inspect the sources. A second generated answer is not independent documentary evidence.
What should I do with generated code?#
Code needs behavioral verification as well as factual checking. Use our AI-generated code review checklist. If you want to keep inference on your own machine, read the local AI laptop checklist.