Is this AI answer right?
Four steps for checking the one sentence you would act on, taken from how people who check facts for a living work, and from what studies found when they checked AI answers the same way.
How often the sources hold up
In 2023 three researchers checked the answers of four AI search engines (Bing Chat, NeevaAI, Perplexity.ai and YouChat) against the pages they cited, sentence by sentence.
Liu, Zhang and Liang, Findings of EMNLP 2023. The engines have changed since; the audit shows what to look for.
Four steps
Pick the sentence you would act on
An answer can run to many sentences, and most of them will not change what you do. Find the one that would: the dose, the date, the rule, the number. That is the sentence to check, and the rest can wait.
Mike Caulfield’s four moves for checking anything online start with the same one: stop, and remember what you set out to check before you read further. Caulfield, SIFT (The Four Moves), 2019
Open its citation and find the words that support it
Click the source the answer gives for that sentence and look for the passage that says the same thing. If the page loads but says something else, or says nothing about it, the answer has not been backed. If there is no source, finding one is the job.
In the 2023 audit drawn above, about one citation in four did not support the sentence it sat beside. Liu, Zhang and Liang, 2023
Leave the page to find out who wrote it
Before trusting the page, open a new tab and search for whoever published it. What do other people say about them? A page can look official and belong to a group with a stake in the answer.
Asked which of two pediatrics websites to trust, every professional fact checker in Wineburg and McGrew’s study chose the 64,000-member American Academy of Pediatrics, because they left each site to read about it. Most of the Stanford students chose a splinter group with a similar name, judging it from its own pages. Wineburg and McGrew, Teachers College Record, 2019
Find a source written by people that settles it
Look for the original: the study, the law, the maker’s instructions, the official figure. A second AI answer, or a page that copies the first, does not count, because it can repeat the same mistake.
This is the first finding of my own research. One local AI checker agreed with whatever reference it was handed. Given the correct value, it told right answers from wrong ones perfectly; with nothing to compare against, it almost could not; told that a wrong value was correct, it passed almost every wrong answer. That was measured on an AI. For a person checking, the lesson of this step is the same: check against the original, and make sure it is the original. Checking AI, what to hand a checker
Then decide whether to accept the sentence, revise it or reject it. Try that on one answer.
What the steps cannot do
- A question nobody has written about cannot be settled this way. Then the honest result is that the answer is unchecked, and it should be treated as a guess.
- It takes time. In the study the fact checkers reached better conclusions in a fraction of the students’ time, but not in no time, and the steps are worth it only for a sentence you would act on.
- The audit is from 2023 and the engines in it have changed since. It shows the kind of failure to look for. It does not say how often today’s systems fail.