Weak scholarship was once easy to spot at a glance: The signs included bad writing, a lack of relevant citations, and little engagement with previous scholarship, such as a paper that promised a “novel” conceptual framework without acknowledging the many other papers making similar claims. As an associate editor of a bioethics journal, I’d make a quick “desk reject” decision and not send these articles for peer review, which effectively killed any hope of publication. Now I can say: AI is affecting every part of academic publishing. Readers have started to notice that AI-assisted articles are slipping into journals. But AI is creating problems behind the scenes as well.
Weak articles now arrive on my desk with fluent prose, plausible-sounding arguments, neatly organized sections, and citations that pass initial AI checks. Only with a deeper look does the AI use become apparent. There are sentences that have all of the trappings of real arguments, but that rarely make sense—such aslike “Algorithmic triage functions less as a clinical tool than as an infrastructure of deferred accountability.” There are citations that refer only to open-access or highly cited articles, because the most popular LLMs can’t readily access paywalled journal articles, so they rely on what’s either free or described at length elsewhere.
Unfortunately, the scholars submitting papers aren’t the only ones using AI. So are many of their peer reviewers. One recent peer review that my journal received—from a senior scholar—read like an unedited cut-and-paste job from an LLM. The giveaway was that it criticized the authors for failing to cite a particular text—in this case, the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders—when the paper did indeed cite it.
[Read: The rise of parasite authors]
Some of my own papers have received similarly suspicious peer reviews. A recent one of a paper that I had co-authored came in the telltale header-bullet style of ChatGPT, requesting 22 major and 14 minor edits on matters spanning ethics, statistics, experimental design, and machine learning. A review by a human typically comes from somewhere; I can usually tell, without knowing the name, whether a reviewer is a philosopher or a social scientist. But this review came from both everywhere and nowhere. The prose sparkled (human reviewers don’t take the time to polish text that no one but the authors and editors will see); the review quibbled with an inessential word and offered three synonyms to use instead (reviewers will object to a word or phrase, but they don’t usually supply a thesaurus-like menu of alternatives); and it questioned the year a reference was published (reviewers regularly check sources, but not the accuracy of small bibliographic details).
But as every academic knows, peer review has always been a highly imperfect process. One might imagine experts with deep knowledge of a paper’s topic or methods carefully reviewing each manuscript prior to publication. The reality is that those with relevant expertise are overwhelmed with review requests and almost never have the time to write one, particularly if the request comes from a lesser-known journal. (I learned this in my first years as a journal editor, when my review requests to top scholars were mostly ignored or declined.)
Although many reviews are thoughtful, others come from people with tangential expertise. (Entire corners of the internet are dedicated to memes from “Reviewer 2,” a caricature of a pedantic scholar who has clearly misunderstood the paper.) When retractions or fabrications make the news, such as the 2024 Frontiers in Cell and Developmental Biology article that included an AI-generated image of a rat with oversize genitals, the public often reacts with outrage at the failure of the peer-review process. But as a journal editor, I’m surprised it doesn’t happen more often.
AI has pushed this already flawed system of peer review to its breaking point. Journals are being inundated with manuscript submissions. Some journals have reported their submissions doubling or even tripling since ChatGPT was released, a few years ago. Across the Elsevier portfolio of journals, submissions were about twice as high in 2025 than in 2024. One journal reported having to double its editors—all unpaid volunteers—to handle the boom in submissions.
Reviewing a growing number of submissions with a critical eye now takes time that minimally compensated or unpaid journal editors, including me, do not have. As a result, even as journal editors desk-reject more manuscripts, they are also sending more and more AI-assisted papers out for review, which in turn off-loads the work of picking apart AI-assisted articles to two or three other overworked scholars. These scholars are, in turn, receiving more and more review requests. In some cases, the cycle has become absurd: AI-assisted papers land with overworked editors, who send them to rushed reviewers, who in some cases use AI to write the review, which the authors revise and resubmit using AI.
A recent study of 7.3 million journal articles published from 2020 to 2025 found that by 2025, more than half showed evidence of LLM assistance, up from 12 percent in 2023. These problems are most likely to afflict the humanities and social sciences, since conceptual argumentation is easier to mimic than experimental results. But AI can generate much of an article even in the hard sciences.
[Steven Strogatz: A mathematician reads the AI future]
The journal article has long been the currency of academic social capital. Most early-career academics can tell you offhand how many they have published, like a salesperson rattling off their most recent quarter’s numbers. This currency has remained valuable because, until recently, a good article was difficult to produce. Such scholarship required time, sometimes funding, and the expertise to design a study or work through an argument. Currencies collapse when people can no longer trust them. Even if a handful of top-tier journals can preserve the value of their articles, they cannot survive in an academic economy in which tens of thousands of other outlets cannot distinguish serious scholarship from work that just looks like it.
Each week, I meet with my trainees to read and critique a journal article in our field. I recently assigned an article with a promising title that had been published in an established journal weeks earlier, which had looked fine after a skimmed read. But as we dissected it, we began to see all of the hallmarks of AI: the sheen without the substance, the citations that were slightly off, the odd absence of conceptual coherence. I’d been on a conference panel with one of the authors earlier that year. Ironically, the topic was the challenges of AI in academic publishing.
I can usually find a teachable moment in any situation. But if I can no longer tell my trainees—postdocs and graduate students—that rigorous scholarship has a value that an academic journal can recognize, then I’m not sure what kind of future, exactly, I am preparing them for.