Today in grading undergraduate writing assignments;

mycotropic@beige.party

"the journal name is real, the title is the real name of an article, the DOI is real, the authors are real… but nothing matches. For example, the DOI is not from the same article as the journal name, or title, or authors. Same with all other elements of the citation.

I think this is likely a Claude hallucination (rather than Chat GPT); Claude is known for doing a better job at pulling real citations or pulling real information to create a fake citation. "

So we assigned students a straightforward writing assignment; pick one of five proposed policy changes and write 500 words in support or in opposition to it. Two or three citations were required. No ChatGPT without sharing your original 500 words, your AI prompt , the resulting AI output and your final revision. 90% awesome, reasonable arguments for and against and then 10% that were essentially perfect.

We went into the citations via PubMed and other sources and found the above. So, obviously 10 points for that section and we indicated that we'd carefully evaluate all future work from the students with hallucinations, any further indication of AI use outside the syllabus will be zero on the assignment and a referral to student affairs for cheating.

So does anyone have any idea why these citations are f-ed up in this reproducibly awful way? Like, what is Claude doing wrong, exactly? It's like Claude is ALMOST identifying relevant work but it can't quite produce actual citations, just things that look like citations. It's like they missed the step where you actually have to READ THE DAMN PAPER YOU WANT TO CITE. Which it could do of course, it could ingest the whole paper, summarize it, simply pattern match versus the students writing and insert a valid doi link. Why is it making this specific error? Anyone?

#Enshittification
#Science #AI #Citations #Hallucination