Verify AI answers & generated code Check the evidence behind an answer before it becomes a business decision or a production change. Reviewed 15 November 2026 · Utilising AI Pty Ltd https://www.utilisingai.com/learn/verify-ai-answers This is a practical checking workflow, not a certification of any AI output. Sources can be stale or wrong; tests cover the cases they exercise. Use the original authoritative material and qualified advice when the decision requires it. Independent educational guide. Original examples; documentation-based guidance. Product availability varies. No provider integration or AI request is run by this download. Quick start =========== 1. Pick the claims that matter: names, dates, numbers, product behavior or code changes. 2. Find their original supporting sources or reproduce the behavior in a safe test. 3. Correct unsupported statements and record what remains unverified before sharing or deploying. Practical tips ============== Ask for uncertainty, not invented completeness Tell the model to mark missing evidence and separate facts from assumptions. For document work, request a supporting passage or location for important claims. Then compare it with the original document; another confident AI answer is not independent verification. https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations Open the actual source Check who published it, its date and the exact passage behind the claim. A Gemini Sources link can lead to related material without supporting every assertion. Search snippets and generated summaries are leads; use the original page for the decision. https://support.google.com/gemini/answer/14143489?hl=en Recalculate the numbers Our practical check: take a small row-level sample and recompute totals in a spreadsheet or calculator. Confirm currency, tax, units, rounding and the reporting period. Ask for the input rows behind a total so that an attractive chart cannot hide a bad calculation. Research changing facts again Use current source retrieval for changing information. ChatGPT web search can supply links, but retrieved pages are untrusted material to evaluate. If search is unavailable, use a supplied official source and mark the freshness limit. https://learn.chatgpt.com/docs/web-search Developer notes =============== Reproduce the original failure Use a representative input that demonstrates the bug. Check that the proposed fix changes the behavior you intended and that a relevant regression test catches a recurrence. Our standard is evidence from execution, not a generated statement that a test passed. https://docs.github.com/en/copilot/concepts/prompting-and-context/prompt-engineering Check beyond the happy path Inspect invalid input, missing permissions and failure states as well as the normal result. For a visual change, compare the page with its reference at relevant screen widths. A typecheck, browser check and ownership check answer different questions. https://code.claude.com/docs/en/common-workflows Review the change itself Read the diff for unrelated edits, new dependencies, secrets and changed access rules. Keep deployment review separate from a coding agent’s summary. Instruction files and model confidence cannot replace enforced access controls or a reviewer. https://code.claude.com/docs/en/permissions Original examples to adapt ========================== Audit a draft’s claims Review this draft against the supplied sources. Return a table: claim | supporting source and location | supported, contradicted or unverified | correction needed. Separate recommendations from factual statements. Do not invent missing citations. Draft: [paste] Sources: [attach or link] Original checking prompt; you must open and assess the supporting sources yourself. https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations Check a calculation Using this table, show which rows contribute to [metric], the formula, units and total. Flag missing or duplicate rows. Keep assumptions separate. I will independently recalculate a sample. Data: [paste redacted example] Original numeric audit prompt. Do not treat the model’s arithmetic as the independent check. Review code evidence Review this change against the requirement [requirement]. Identify behavior changes, potential regressions and missing checks. List checks actually executed separately from checks merely suggested. Do not claim a live flow works from a build result alone. Original review prompt for a tool with the relevant code or diff. https://docs.github.com/en/copilot/concepts/prompting-and-context/prompt-engineering Check the result ================ - Is every important claim supported by a source you opened? - Have totals been independently checked? - For code, what was actually tested, with which inputs and results? - Are remaining uncertainties visible to the person making the decision? Common mistakes =============== - Counting two AI answers as two independent sources. - Treating a citation, green build or confident tone as universal proof. - Publishing sensitive data in an attempt to make verification easier. Sources ======= Anthropic · Reduce hallucinations · Official documentation https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations Google · Sources in Gemini Apps · Official help https://support.google.com/gemini/answer/14143489?hl=en OpenAI · Web search · Official documentation https://learn.chatgpt.com/docs/web-search GitHub · Copilot prompt engineering · Official developer documentation https://docs.github.com/en/copilot/concepts/prompting-and-context/prompt-engineering Anthropic · Claude Code best practices · Official documentation https://code.claude.com/docs/en/best-practices Anthropic · Configure permissions · Official documentation https://code.claude.com/docs/en/permissions Anthropic · Common coding workflows · Official documentation https://code.claude.com/docs/en/common-workflows