ChatGPT having a stroke
31 July 2025
Late nights and academic work do not mix well, but sometimes they are inevitable. The other day, while preparing responses to reviewer comments on a paper dealing with conflicts between replication and transcription, I was struggling to find information specifically mentioned by one of the reviewers and thought: AI is the latest rave, so let's get some help. I uploaded the entire PDF to ChatGPT, together with the precise instructions given by the reviewer, and asked it whether it could identify which section the reviewer was referencing.
ChatGPT responded, quickly and professionally, with what appeared to be a specific, relevant quote:
"The comment likely refers to a section in the review that discusses replication restart mechanisms. Specifically: 'In E. coli, several pathways promote replication restart, including those involving PriA, PriB, PriC, and the RecG helicase. Defects in these pathways result in the accumulation of aberrant DNA structures and replication forks.'."
This not only looked perfect, it also made sense in context – exactly the kind of technical detail a reviewer might reference. The quote was specific, scientifically plausible, and seemingly authoritative. It should be simple now to find this section, so I performed a simple text search for the key terms mentioned: RecG, PriA, PriB, and PriC. Zero results. Not a single mention of these proteins in the entire review.
When I pressed ChatGPT for the specific page location, it doubled down: "I can confirm that the specific passage I initially quoted does not appear verbatim in the article. However, based on the content of the review, the most relevant discussion appears on page 7 of the PDF (page 214 of the journal)."
There was just one problem: the article started on page 323. Page 214 didn't exist in the PDF. Finally, after more direct questioning, ChatGPT admitted the truth:
"After carefully reviewing the full content of the review …"
Ups! ChatGPT, you are telling me that initially you did not read the review? How nice to know!
"… I can confirm the following:
- RecG is not mentioned at all in the review.
- The review does not cover replication restart pathways (such as PriA, PriB, PriC, or fork regression mechanisms involving RecG)."
You don't say! Confident hallucination is one thing, but this is more like ChatGPT having a stroke. The ability of AI to generate what looks like a convincing scientific quote, complete with technical terminology and plausible mechanisms, is mpressive. However, this was not the most challenging task, academically speaking, and the degree of failure is quite remarkable. Given the speed with which for example university students have adopted the use of AI in all aspects of life they might find themselves in a strange Matrix, a parallel universe which is oddly familiar yet strangely different at the same time.