AI citation accuracy leaderboard
The same prompts for each chatbot, with every reference checked against scholarly databases.
Real: a published work matches the reference in a scholarly database or on the publisher's site. Some details may still be wrong.
Usable as written: real, and every element APA 7 asks for is present and matches the record, so you can cite it as is.
A range means some references cite ACM article numbers we could not confirm (5 from ChatGPT, 4 from Claude). The low end counts them as needing a fix, the high end as usable.
How we tested
We gave each chatbot three prompts on two topics: working memory and anxiety in adolescents, and knowledge graph construction. One prompt was casual, one set requirements, and one asked for real, peer-reviewed articles with a DOI for each. Every prompt went into a new chat on the free version with default settings, and we kept the first answer. Then we looked up each reference in scholarly databases and on the publisher's site and compared it with APA 7.
How we judged each reference
- Real: a published work matches the reference.
- Needs a fix: the work is real, but at least one element is missing or differs from the record, such as the DOI, article number, authors, year, volume, issue or pages.
- Usable as written: real, and nothing needs fixing.
The three prompts
The second topic used the same three prompts with "knowledge graph construction" in place of the topic.