More on the disturbing plausibility of ChatGPT

Prompted by a post on Mastodon (which, like all Mastodon posts, I can no longer find), I asked ChatGPT to tell me about my own papers. The response started out well but quickly got much worse. I will indent my comments on its response.

Q. What are some articles written by Michael P. Taylor?

A. Michael P. Taylor is a vertebrate paleontologist and open science advocate. He has written numerous articles and papers on topics such as the taxonomy and systematics of sauropod dinosaurs, open access publishing, and scholarly communication.

So far, so good. In general, LLMs seem to be good at synthesizing overviews. It’s specific facts that trip them up.

Here are a few examples of his articles:

Taylor, M. P. (2009). A re-evaluation of Brachiosaurus altithorax Riggs 1903 (Dinosauria, Sauropoda) and its generic separation from Giraffatitan brancai (Janensch 1914) (Dinosauria, Sauropoda). Journal of Vertebrate Paleontology, 29(3), 787-806.

The first one gives is pretty good. The only mistake is that it’s oddly appended “(Dinosauria, Sauropoda)” to the end of the article title. The journal, volume, issue and page-numbers are all correct.

Taylor, M. P. (2011). Sauropod dinosaur research: a historical review. In A. W. A. Kellner, & D. A. S. Silva (Eds.), Vertebrate paleontology and sedimentology of the Jurassic-Cretaceous boundary in the northern and central parts of the Lusitanian Basin, Portugal (pp. 11-43). Sociedade de Geologia de Portugal.

I did write a book-chapter titled “Sauropod dinosaur research: a historical review”. But it came out in 2010, not 2011. And it ran from pages 361–386, not 11-43. And it was in Dinosaurs and Other Extinct Saurians: a Historical Perspective, not Vertebrate paleontology and sedimentology of the Jurassic-Cretaceous boundary in the northern and central parts of the Lusitanian Basin, Portugal (which doesn’t exist). Which was edited by Richard T. J. Moody, Eric Buffetaut, Darren Naish and David M. Martill, not A. W. A. Kellner & D. A. S. Silva. And published by the Geological Society of London, not Sociedade de Geologia de Portugal.

Taylor, M. P., & Wedel, M. J. (2018). Why sauropods had long necks; and why giraffes have short necks. PeerJ Preprints, 6, e27537v1.

I and Matt did co-write a paper with a similar title in PeerJ. But it’s Why sauropods had long necks; and why giraffes have short necks, not the other way around. And it came out in 2013, not 2018. And it was in PeerJ proper, not PeerJ Preprints.

Taylor, M. P. (2020). The open-access wars: How to free science from academic paywalls. Wired.

I never wrote any such paper. There is an article out there whose HTML title is the one given here, though the visible title is completely different: “The war to free science”. It’s in Vox, not Wired; it came out in 2019, not 2020. And it’s by Brian Resnick and Julia Belluz.

Taylor, M. P., & Knope, M. L. (2021). The future of scholarly publishing is in our hands. The Conversation.

As far as I recall, I never wrote anything for The Conversation. There seems to be no article named “The future of scholarly publishing is in our hands”. And I have never met or corresponded with, far less co-authored with, anyone with the surname Knope.

All in all, a pretty horrible performance — but, and this is the scary part, one that even someone pretty familiar with my work could very easily have swallowed.

I made a very similar point a few months ago in These new “artificial intelligence” programs don’t know what they’re talking about. If the responses were obviously wrong, there wouldn’t be too much of a problem here. But this is close enough to fool most people most of the time.

As more and more of what we read is likely to be written by LLMs in the next years, I am concerned that we’re going to be dealing with an awful lot of this kind of thing. The ability to pan gold out of the stream of lies is going to become really important.