Publishing in the age of AI - a view from the trenches

No posts for a while! I have been overwhelmed with teaching, grant applications, and editorial work. For all the claims that AI is incredible and can automate so much for us, everyone seems busier than ever.

The Physical Review journals are facing a massive increase in submissions - double compared to last year.

A substantial fraction of this increase is due to LLM-assisted (or entirely LLM-generated manuscripts).

To the surprise of some, most of these submissions to Physical Review A are not from "independent researchers," but rather from university-affiliated students and faculty. People who should know better.

The pool of willing referees (already a minority of authors who publish with us) cannot cope with this increase. So editors need to be more selective and reject more papers without external review. Unfortunately for us, LLMs are trained to create content that seems convincing at first glance. And we can't use LLM detectors (which have reliability issues), because the issue is not with the text itself (LLMs are great for polishing, and are now explicitly allowed), but with the (lack of) new and correct physics.

Physical Review A is an archival journal, which means its role is to document substantial research progress, not only the very best results. A researcher may spend months reproducing and then extending a prior result, learning much in the process and laying the foundation for future research breakthroughs.

An LLM model doing the same in a day? No. It will give a result, but you need to understand the result enough to trust it. Which takes time. Publishing unverified material corrupts the scientific record. It is the author's responsibility to be sure what they are publishing is correct.

To be a good writer, read what great authors have written.

To be a good researcher, read about good research. The original material, not the AI summary.

The smoking gun of an AI paper is that the writing is off. Out of place jargon. Unusual article structure. Too many tables when a figure would be much clearer. An overly authoritative tone. Dozens of propositions and theorems in a physics paper where the deepest maths is just diagonalising a small matrix. Seeing these papers, it becomes apparent that the supposed author has not read a single paper in the field in which they are trying to publish.