Using ChatGPT for fact checking
I wanted to explore the idea that we could use ChatGPT for fact checking. So I took a random article - it gave me the article on Jewish Ideas Daily, copied the text into ChatGPT with the following prompt:
Here's the text of a Wikipedia article. Please find the three statements that are most likely incorrect and rank them, in order to guide fact checking:
Followed by the Wikipedia article text. Here's the answer:
- "Jewish Ideas Daily was edited by Margot Lurie from its founding in January 2010 until December 2012." This statement is incorrect because according to the article, Margot Lurie was the editor of Jewish Ideas Daily "until December 2012."
- "Jewish Ideas Daily drew notice for controversial features including one accusing Christopher Hitchens of antisemitism." This statement is difficult to verify without more context, as it is not clear when this feature was published or what specifically it accused Hitchens of.
- "Like Jewish Review of Books, and Jewish Ideas Daily, Mosaic is funded by the Tikvah Fund." This statement is incorrect because according to the article, Mosaic succeeded Jewish Ideas Daily in June 2013 and is therefore not funded by the Tikvah Fund in the same way as Jewish Ideas Daily and Jewish Review of Books.
Regarding the first statement, GPT says that the sentence "Jewish Ideas Daily was edited by Margot Lurie from its founding in January 2010 until December 2012." is incorrect. I am not disagreeing, but that sentence does not appear in the Wikipedia article text.
Regarding the second statement, GPT says that there is no reference for that statement. In fact, there is, and Wikipedia provides it.
Regarding the third statement, I didn't check the statement, but the argumentation for why it should be checked is unconvincing: if Tikvah fund financed Jewish Ideas Daily, why would it not finance the successor Mosaic? It would be good to add a reference for these statements, but that's not the suggestion.
In short: the review by ChatGPT looks really good, but the suggestions in this case were not good.
The exercise was helpful insofar the article infobox and the text were disagreeing on the founding of the newspaper. I fixed that, but that's nothing ChatGPT pointed out (and couldn't, as I didn't copy and paste the infobox).
Previous entry: Economic impacts of large language models, a take | Next entry: Goal for Wikidata lexicographic data coverage 2023 |