Before you trust an AI answer: a verification checklist.
A confident tone is not evidence. Before you use an AI-generated answer in an assignment, a work decision, or a personal project, take a few minutes to check what supports it and where it might fail.
Step 1: Separate the answer into checkable claims
Do not try to verify a long response as one block. Break it into individual claims: dates, names, definitions, numerical values, quotations, instructions, and cause-and-effect statements. Mark which claims are essential to the decision you need to make.
For a short answer, a simple list is enough. For a longer answer, make a table with the claim, the evidence needed, the source checked, and the result.
Step 2: Find the right kind of source
Choose evidence appropriate to the claim. For a course topic, start with your prescribed textbook, lecture material, or instructor guidance. For a product feature, check the manufacturer's current documentation. For a policy or deadline, use the official organization responsible for it. A search result snippet or another AI answer is not a substitute for reading the underlying source.
- Prefer primary or official sources when they are available.
- Check the publication or update date for changing information.
- Open the source and confirm that it actually supports the claim.
- Use more than one independent reliable source when the stakes justify it.
Step 3: Verify numbers, quotes, and references separately
Numerical answers should be recalculated when possible. Check units, assumptions, rounding, and the inputs used. For quotations, search the original source and verify the exact wording and context. For citations, make sure the cited page exists and supports the sentence attached to it.
Never treat a citation as proof just because it looks formal. AI systems can produce plausible-looking but inaccurate titles, links, page numbers, or quotations.
Step 4: Ask what is missing or overstated
Even a mostly correct response can be misleading if it leaves out a key condition. Look for absolute language such as “always,” “never,” or “guaranteed.” Check whether the answer explains assumptions, exceptions, uncertainty, and the limits of the evidence.
Try asking: “What assumptions does this answer rely on?”, “Which claims are not directly supported by the sources?”, and “What would change this conclusion?” These questions can reveal gaps, but their answers still need checking.
Step 5: Match the level of checking to the risk
A low-stakes brainstorming suggestion may need only a quick sanity check. A claim that affects health, legal rights, money, safety, or an important academic submission deserves much stronger verification with qualified or authoritative sources. Do not rely on an AI response as the sole basis for a high-stakes decision.
| Risk level | Reasonable next step |
|---|---|
| Low | Check for obvious errors and test whether the suggestion works. |
| Moderate | Verify important claims against reliable sources and document corrections. |
| High | Consult authoritative sources or a qualified professional; do not rely on AI alone. |
Copyable claim-check table
| AI claim | Evidence checked | Status |
|---|---|---|
| Write one specific claim | Official source, textbook, calculation, or original record | Supported / Incorrect / Unclear |
| Record any number or quote | Recalculate or compare with original wording | Supported / Incorrect / Unclear |
| Record a key conclusion | Check assumptions and missing conditions | Supported / Needs qualification |
Before using the answer
- I identified the important claims instead of trusting the overall tone.
- I checked the underlying sources, not just snippets or AI-generated citations.
- I recalculated important numbers and checked quotations in context.
- I noted uncertainty, assumptions, and missing information.
- I used stronger independent verification for higher-risk decisions.
For a deeper workflow, see How to verify an AI answer and How to verify AI-generated study notes. To compare different AI systems fairly, start with the reusable comparison template.