Everyday practical guides

Separate an AI answer from its evidence

Treat an AI answer as text to examine, then make a short list of checkable claims. A convincing sentence and accessible evidence are different things. The cited UNESCO guidance places human judgment at the centre; the grid below is our editorial reading method.

Published by World Web Certified · about 3 min to read

Notes compared with books and a computer
AI-generated illustration · Fictional scene.

The method

  1. Choose a narrow question and copy only statements answering it directly. Number claims about identity, dates, quantities, rules or causal relationships. Distinguish drafting suggestions from assertions about the world; they require different kinds of checking.
  2. Find a document establishing each precise point, rather than stopping at an organization’s name. Open the relevant passage and record what it actually supports. A source discussing the topic does not establish every sentence in the answer.
  3. Mark rows as supported, contradicted, unestablished or outside scope. Rewrite your synthesis using the table and retain visible uncertainty. Do not upload a confidential personal file to the service merely to ask it to validate its own response.

Fictional worked example

A fictional assistant supplies three dates and an explanation. Your documents establish one date, contradict another and say nothing about the third. The table preserves these separate states; the explanation remains open until relevant material is found. Finding one source does not validate the answer as a whole.

The record to keep

Keep the question, initial answer, documentary passages and row-by-row conclusion. This record lets you repeat the checks when the text or sources change.

Your record, step by step

For your own situation, record the following information. Check a field when it is documented; missing information still needs verification. The boxes are a reminder within this page.

This checklist does not validate your project. No entries are sent or saved.

Fictional worked case. A fictional AI answer mixes a number, an attribution and a proposed method. Examine each separately.

Your record, step by step : Separate an AI answer from its evidence
ElementObservation in the caseInterpretation or next check
Number“42%” with no period or populationDo not reuse the value while scope is missing
AttributionOrganisation named without a precise documentFind the text actually published by that organisation
LinkOpened source lacks the claimed sentenceRecord the mismatch rather than treating a link as proof
MethodSequence of steps suggested by the assistantSeparate it from an observed or measured result

What this method can establish

The grid records checks rather than guaranteeing truth. An unavailable reference remains unestablished. Medical, legal or financial decisions require suitable context and an appropriate professional.

Useful questions before you start

Does a working link validate the nearby claim?

No. Check that the passage supports the exact fact with the same date, unit and scope. An active link can be irrelevant.

Does an unavailable reference mean that the claim is false?

No. Mark it as unestablished and look for the title, author, publisher or an identifiable archived version. While the document and passage remain inaccessible, do not describe them as consulted. A documented contradiction and missing evidence are different conclusions.

Do two assistants giving the same answer provide two independent pieces of evidence?

Not by themselves. They may repeat the same sources or the same error. Compare the original documents and the passages supporting each claim. Agreement between answers can suggest a research lead; it does not replace documentary checking.

Supporting documentation

These references support the method. Each source shows its language and consultation date; an available directory record explains its purpose.

An editorial method, a fictional example and references for further reading.

Within our AR network

These websites belong to our network. They appear separately from this topic’s external resources; inclusion does not verify their offers.

Back to the topic →

Published on · Updated on