How can I check whether an AI answer is right? Open the source, verify the specifics, and match the effort to the stakes.
Short answer. Open the source it cites, read the passage, and confirm that the specific names, numbers and dates match the original. A citation is not proof: in a 2023 audit of generative search engines only 51.5% of generated sentences were fully supported by their citations, and a 2024 Stanford study of legal tools that cite sources still found hallucination rates of 17% to 33%. Then scale the checking to the stakes: skim for brainstorming, verify every fact for an internal note, and require a qualified person to verify every citation for anything a client, court or regulator will see. Record what you checked.
01
A seven-step routine
Ask for the source passage, not just the claim, and open it.
Read the passage: does it say what the answer says, or only something nearby?
Match every specific against the original: names, numbers, dates, section numbers and amounts.
Look for what is missing: ask what exceptions, conditions or counter-arguments the source contains.
Test the tool on cases you know. Run questions whose answers you already have, including one whose answer is not in the documents.
Cross-check high-stakes claims with a second, independent source or a second person.
Record what you checked and who signed off.
02
Why a citation is not enough
Evidence
Finding
Liu, Zhang and Liang, 2023 (Bing Chat, NeevaAI, Perplexity, YouChat)
Only 51.5% of generated sentences were fully supported by citations, and only 74.5% of citations supported their sentence
Magesh et al., Stanford, 2024
Legal research tools that retrieve and cite sources hallucinated between 17% and 33% of the time
EBU and BBC study, October 2025
More than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity on news: almost half had at least one significant issue, a third had serious sourcing problems and a fifth had major accuracy issues
NIST AI 600-1, 2024
Models can produce confabulated logic or citations that purport to justify an answer, and these may mislead people into trusting it
03
Match the checking to the stakes
Stakes
Example
Minimum check
Low
Brainstorming, a first outline
Skim for sense and do not reuse the facts
Medium
An internal memo, a meeting summary
Verify every figure, name and date against the source
High
Client advice, a court filing, a clinical note, a tender
A qualified person verifies every citation and quotation in the original and signs
04
Signs an answer needs a second look
Very specific numbers or quotations with no source.
A citation you cannot find, or one that exists but does not say what is claimed.
Text that reads more smoothly than the source it claims to quote.
An answer about something outside the documents you supplied.
Different answers when you ask the same question twice.
A confident explanation of how the answer was reached. NIST notes that models can give logical steps even when the answer is wrong.
05
How ANVE helps you check
ANVE shows the source passage next to answers over the office’s own documents. Each meeting action item must quote the transcript line it came from, and its owner must be confirmed by a person before it is approved, with the confirmation logged. In the agent loop, a claim made without a successful tool call is challenged. None of this removes the need for a person to check. ANVE is pre-release.
06
Questions people ask
Can I ask the AI to check itself?
It can help find inconsistencies, but it cannot reliably confirm facts: NIST notes that a model may give logical steps for an answer even when the answer is wrong. Use the source, not the model’s confidence.
How much checking is enough?
Enough that you could defend the work to a client or a regulator. For high-stakes work that means a qualified person verifies every citation in the original. The Supreme Court of India’s July 2026 order on AI-invented precedents shows the cost of skipping it.
Should I use two AI tools to cross-check?
A second tool can flag disagreements, but both may share the same mistake. Treat disagreement as a prompt to open the source, and agreement as no proof.
Does a local model need less checking?
No. Where the model runs changes who can see your data, not how often it is wrong.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
