AI and academic publishing

Some thoughts from June 2026 about the state and prospects of transportation research

Some notes I drew up for a June 15 panel discussion of AI in Academic Publishing at a conference I attended at Southwest Jiaotong University in Chengdu.

1. Authors should use paid AI (see 2) before submitting. They should use it not just to clean up English but as a hostile referee. Ask your LLM of choice (mine at this stage of history is ChatGPT and Codex) what is wrong with the paper. Ask it to review the paper like a peer reviewers for the journal you are planning to submit to, ask what is unclear, undefined, multiply-defined, unsupported, unreplicable, under-cited, overclaimed, or likely to irritate Reviewer 2. Once you satisfy your LLM of choice with revisions, use a second LLM and repeat, it will find different things. Of course ask it how it would revise the paper as well, but don’t just accept it at face value. This does not guarantee acceptance, but not doing this will probably soon guarantee rejection. As discussed below, reviewers and editors are likely to be using AI too, and perhaps better (more tokens, more advanced, more expensive) AI than the authors used. Even then, reviewers will find something “wrong”, because that is how reviewers are. They are not paid. Their reward is anonymous criticism. (For-profit journals should pay reviewers.)

2. Newer paid AI is much better than a year ago. This is not just a smoother autocomplete. Free AI and paid AI are different tools now. Paid AI can read the paper, follow the method, look at the code, question the equations, find unsupported claims, and write something that looks a lot like a referee report. For technical work, the Codex (or Claude Code) version is better than the ChatBot version, because transport research is dominated by code, data, equations, maps, tables, and claims about the world. Make the investment in yourself.

3. Reviews are already being done with the help of AI, or entirely by AI. This is happening whether or not journal policies have caught up. This is done by some editors as well, I believe but cannot prove, who are having trouble finding reviewers.

4. Editors should use AI as a screener before sending papers to human review, especially at the desk-reject stage. This should be fully disclosed to authors. This would be a first pass: Is the paper in scope? Is there a research question? Do the claims follow from the evidence? Is the method described? Are the data and code available? Is the literature review just a pile of citations? Is this paper wasting everyone’s time? This would help editors. It would help authors. It would save scarce human reviewer time.

5. AI should be used after human review too. Did the authors actually answer the referee reports, or just write “we thank the reviewer for their insightful comments” 37 times? Are the data and code available? Do the tables support the claims? Are they internally consistent? Are the variables in the equations defined properly, do variable definitions slip across the paper? Are the citations real? Are the limitations stated? These are basic scientific checks. Humans need not do all of this by hand.

6. AI is now, in June 2026, good enough to mostly write, under supervision, an acceptable paper in many transport journals, if there is a good research question, available data, a real method, and a human who knows what they are doing. It can also draft a theory paper. This is no longer hypothetical. It can write, iterate, and revise.

7. AI still does not yet generally generate good questions. That remains, so far, the human job. The scarce skill is knowing what is worth asking, what evidence matters, what claims are justified, and why anyone should care.

FIN