JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Worked example

How to check if a document supports an AI assistant's claim

This guide explains a method to test whether an AI assistant's answers match the original documents. It shows how to spot made-up quotes, false claims, and statements that lack evidence.

Original by TypeSafe AIRetrieval

Listen to this guide

JevMade’s plain-English explanation

0:00 /

AI narration

Credits

“Double-checking citations” by TypeSafe AI. Read the original source.

This expanded guide is an AI-narrated adaptation prepared by JevMade. It expands the source’s essential ideas, examples and caveats in JevMade’s own words and is not a word-for-word reading. The synthetic voice does not imitate the author or imply their endorsement.

Our summary

This guide tests a way to check if an AI assistant's answers are actually backed by the documents it read. People use this method to catch mistakes when software writes summaries or answers questions based on a specific text.

The process first looks for the exact quoted words in the original document. If found, an AI tool named Jev chooses whether the surrounding sentences support the claim, contradict it, or say nothing at all. Jev picks from options rather than writing a new answer.

This method is useful for fact-checkers, but it has strict limits. In these tests, the software required exact word matches and rejected quotes with minor typos. A supported quote does not prove the original document is true, and uncertain answers still need a person to check them.

Key takeaways

  1. Find the exact quoted text in the original document before judging if it supports the claim.
  2. Treat a document that says nothing about a claim differently than one that directly contradicts it.
  3. Ask a person to review the results if the software is not highly confident in its choice.

Finding a matching quote does not prove that the original source contains true information. A quote pulled from the wrong section can also lead to an incorrect judgment.

TypeSafe cookbook · Source reviewed

Read the original guide Opens the author’s site in a new tab.