[AISWorld] GPT for Reviews

[Walden, Eric]

I recently received a review and two of the reviewer comments were very similar.  They had the same number of bullet points and similar content in the bullet points.  Moreover, the bullet points had similar numbers of words.  Finally, the bullet points almost all says X was not discussed, even though X was clearly discussed latter in the paper.  X was mentioned in the introduction and then a deep dive was discussed in the methods section, because that is how papers are organized.

For example, bullet point 4 in the two reviews was:

* The particular modifications that were made to the Alexnet, GoogleNet, and VGG16 architectures to adapt them for 3D data, as well as the various optimization techniques available for applying to the initial layers of the CNNs to maintain computational efficiency, have been skipped.

4.      Why are the initial layers of a CNN optimizing the most for 3D counterparts not given? As well, an account of how optimizing the initial layers helps in managing the computation time and complexity of the CNN is not clarified.

Same bullet point number, same basic comment, same sentence structure, similar size, and they both complain that something that is discussed in the methods section is not fully accounted for in the first few thousand words.  This is what would happen if you cut a paper in the GPT acceptable chunks of a few thousand words and asked for a review.

Anyway, I was wondering if people had been getting reviews that they suspect were generated by a large language model?



--
Eric Walden
Director of the Texas Tech Neuroimaging Institute
Rawls Endowed Chair of Information Systems and Quantitative Sciences
Rawls College of Business
703 Flint Avenue
Texas Tech University
806-834-1925
eric.walden at ttu.edu<mailto:eric.walden at ttu.edu>
http://ericwalden.net<http://ericwalden.net/>