Use it safely
AI can sound confident and still be wrong: how do I check it?
A lightweight verification habit for dates, facts, recommendations, calculations, and anything else that matters.
The short version
Do not verify every harmless sentence. Verify the claims that would change what you do: dates, prices, rules, quotes, locations, numbers, safety advice, and other consequential facts.
Optional explanationRead the full guide
Fluent is not the same as factual
Modern AI is optimized to produce a useful response, not to visibly hesitate every time its information is uncertain. That means a wrong answer can sound just as polished as a right one.
The goal is not to distrust everything. It is to notice which claims deserve a check before you act on them.
Use the consequence test
- Would a wrong date make me miss something? Check it.
- Would a wrong price or fee change a purchase? Check it.
- Would a wrong rule affect legal, medical, financial, or workplace action? Use an authoritative source or qualified professional.
- Would a stale opening hour waste a drive? Check the current source.
- Is this just a brainstorming suggestion for dinner? You probably do not need forensic verification.
Ask the AI to make verification easier
You can structure an answer so facts and assumptions are easier to inspect.
Try this
Separate confirmed information from assumptions. For each date, price, rule, or factual claim that matters, tell me what I should verify in the original or current source before acting.
Go back to the source
If the AI summarized a document, compare important claims with that document. If it gave current information, check a current authoritative source. If it made a calculation that matters, inspect the inputs and arithmetic. Asking the same AI “Are you sure?” is not independent verification.
Use three levels of checking instead of verifying everything
Verification does not need to become a research project. Match the effort to the consequence of being wrong.
- Low consequence: brainstorming dinner or rewriting your own sentence — use ordinary judgment.
- Medium consequence: current prices, hours, dates, product specs, school requirements — check the original or current authoritative source.
- High consequence: medical, legal, financial, safety, contractual, or employment decisions — use qualified or authoritative sources and treat AI as preparation, not the final authority.
Watch for four answer types that deserve extra attention
- Current facts: opening hours, prices, availability, schedules, policies, software versions, or recent events.
- Tiny details with big consequences: dates, amounts, addresses, deadlines, dosages, model numbers, and quoted language.
- Extracted text: scanned bills, receipts, screenshots, tables, and handwriting can be misread.
- Calculations: a polished explanation can still be built on the wrong input or arithmetic.
Useful verification prompt
“List the claims in your answer that could change with time or would materially change what I do. For each one, tell me the best original source to check.”
Asking the same AI “Are you sure?” is not a second source
The assistant may confidently restate the same mistake. Independent verification means going back to the document, current official source, calculator, provider, or qualified professional that can actually confirm the claim.
Optional depthGo deeper when you want to
Verification becomes manageable when you do not treat every sentence equally. The deeper method is to identify which claims can change your behavior, trace them to the best source, and preserve uncertainty instead of forcing a yes/no answer.
Build a claim ledger for consequential answers
When an answer contains several facts, extract the small subset that would materially change what you do. For each claim, record whether it came from your supplied material, a current external source, calculation, or model interpretation.
This turns ‘double-check the AI’ into a finite review task instead of rereading everything with equal suspicion.
- Claim: what exactly would change your action?
- Source: where should that claim be confirmed?
- Freshness: can the fact change over time?
- Status: confirmed, contradicted, uncertain, or not checked.
Useful verification request
‘Extract only the claims in your answer that would change a deadline, payment, booking, safety decision, or legal/medical/financial action. For each, tell me the best original source to verify.’
Preserve the difference between absent, negative, and unknown
Comparisons often fail because the assistant converts ‘I did not find it’ into ‘it is not included’ or ‘false.’ Those are different evidence states.
Use explicit labels such as stated, explicitly excluded, not stated, unclear, and conflicting. That one habit dramatically improves contract, estimate, policy, and product comparisons because missing evidence stays visible.
- Not stated ≠ not included.
- No source found ≠ false.
- Two conflicting sources ≠ permission to choose the one that sounds newer unless the source actually resolves it.
Use independent tools for independent failure modes
Asking the same model to critique itself can catch mistakes, but it is not independent verification. Match the verifier to the failure mode.
Use the original document for extraction claims, a calculator for arithmetic, an official current source for policies/hours/prices, and a qualified professional when the decision requires expertise or accountability the assistant does not have.
Verification stack
AI extracts the payment deadline → you open the original notice to confirm the date. AI calculates the monthly difference → calculator/spreadsheet confirms the arithmetic. AI explains the policy → official policy confirms the current rule.