“Something is up in academic research,” write the members of an AI Task Force convened by the journal Organization Science. As they go on to elaborate:
“If you are an editor or reviewer at a journal these days, you probably already know this. The manuscripts are arriving in greater volume, with a particular feel that is hard to pin down. On the surface, the papers look the same as ever, but the writing feels weightless in a way that rarely describes academic writing…you find yourself scratching your head at the meaning the words are trying to convey.”
The culprit? The task force crunched the numbers and produced a clear answer. Starting in 2023, after ChatGPT became available, the number of submissions to Organization Science rapidly increased. At the same time, the percentage of submissions classified as using minimal AI has plummeted from near 100% down closer to 30%.
The impact of this shift on readability has been marked, with scores on a standard “reading ease” metric falling by 1.28 standard deviations between January 2021 and January 2026:
“Submissions have become far harder to read,” the Task Force reports. “This is counterintuitive. Most people assume that AI produces cleaner, more polished text. And in some narrow dimensions, it does…but on the measures that capture whether a reader can actually parse and absorb the prose, AI writing is worse…[using] longer words, more complex sentence structures, more jargon, and more nominalizations.”
Papers that are more difficult to read might be worth it if AI increased the amount of good science being produced. But this doesn’t seem to be the case. Organization Science is desk-rejecting (e.g., rejecting a paper before even sending it to peer reviewers) nearly 70% of manuscripts that made heavy use of AI. This number drops to 44% for papers written without AI.
Similarly, only 3.2% of high-AI papers are ultimately accepted compared to 12% of low-AI papers.
(It’s important to note here that the editors making these decisions do not themselves know the role of AI in the paper construction. These are retrospective analyses.)
All of this points to a distressing conclusion: generative AI tools are leading to many more poor paper submissions, which are taxing the time and patience of the community tasked with reviewing this research.
These tools make individual researchers’ lives easier in the moment (writing is hard!), but they are leading to worse outcomes for the field as a whole.
I tell this story because I think it’s a useful cautionary tale about AI. As I’ve been trying to argue from many different angles in recent weeks (e.g., 1 2), making things faster or easier is not the same as making things better.
Sometimes there really is no shortcut to taking your time.
I was so relieved when my tenure as an academic journal editor ended a couple months ago. The situation is getting quite bad and is taxing our already stretched (volunteer!) resources. Mostly, the AI-produced papers are still very obvious: they look great at a glance, but when you read them, there is nothing there. Nevertheless, it’s a great waste of everybody’s time. I’m hoping the situation will hit a breaking point and then calm down before I take up any more editorial roles!
I also see these AI assisted papers in peer-reviewed reports. Deciphering the gist of the recommendation and commentary is difficult because of the frequency of nonsense synonyms and vapid vocabulary. I would much rather have poor grammar and medical jargon to sort through!
This is my experience exactly and is deterring me from reading papers at all. The minute I get a whiff of A.I. generated writing, I instinctively repel it and can’t read on. Do you have any thoughts on how we could change the incentives to encourage human generated content? Maybe some journals which only accepted no AI writing assistance use.
I disagree with it being “counterintuitive” that papers written with AI are harder to read. People who are even moderately well read have said for a while that AI writes badly.
There is a similar trend going on for HR teams in charge of hiring. We are receiving hundreds more applications and resumes, and they are increasingly hard to read. They’re often filled with large filler words that don’t add value, and leave me feeling like I don’t know what they’re trying to convey. In hiring team conversations, it repeatedly comes up how the submissions that were clearly written without AI (or minimal use of AI) have a different feel. Even ‘mistakes’ are seen in a different light because we know the person created it (largely) on their own. This has changed the landscape of hiring and interviewing, and in my personal experience, not for the better.
I think it’s clear that Cal’s longtime comment that “If you want to get strong you have to lift the heavy things” applies to mental work as well. The only way to develop and maintain the ability to do the work is to do the work.
If you let a machine summarize papers, you lose the ability to read and understand them. If you let a machine do the writing, which requires thinking and processing, you lose that ability.
It’s quite frustrating because writing the papers is not the bottleneck; designing good experiments and collecting good data is.
I find myself more and more leaning towards the idea that LLMs are only good for allowing incompetent people to produce substandard work without effort. Their chief role in efficiency, therefore, may be to identify those people who can safely be dumped because they’re useless.
“All of this points to a distressing conclusion: generative AI tools are leading to many more poor paper submissions…”
Too bad you hide the actual subject of the sentence rather than stating “a distressing conclusion that researchers are choosing to look for paper-writing shortcuts rather than putting in the work themselves…”
Not focusing on human researchers actively making choices plays into unhelpful stereotypes of inevitability (e.g., “The manuscripts are arriving in greater volume”) or magic or other such.
Interesting post, especially against the backdrop of “country of geniuses AI replacing everyone in the workforce”. Why, then, are so many of the publications rejected? Perhaps this rejection rate should be one AI progress metric. Then of course someone would think to fine tune the models against this metric (would be interesting to know if it’d work).
Having spent 10 years total in higher education (4 undergrad, and 6 in grad school), I can attest that learning to write well sharpened my thinking skills and improved my learning. If you use AI as a writing crutch then you are depriving yourself of valuable learning, and the development of thinking skills. There was a definite reason why in the traditional educational model at Oxford and Cambridge the students were (still are?) required to write an essay each week. It forces you to think.
Yes students here at Cambridge are (depending on the Tripos) still required to write (usually multiple) essays per week. I’ve been reassured by the fact that the majority of my students are repulsed by the idea of using AI to help them. They’re far too competitive and enjoy thinking and writing too much. One of the concerns I have had is that because the Oxbridge system is notoriously pressured, and students find it stressful, that they will feel it is easier to turn to AI to cope with the workload. However so far that’s not been the case, thankfully.
Not sure how to judge causality.
Is AI making submissions worse?
Or, are good submitters using AI in a limited but effective way, while bad/low quality/lazy submitters can use AI now to merely churn out more submissions?
There might be a lot more work to narrow down the bigger number of submissions, but for the question of whether it improves science, is there any way to measure whether the final group of papers that the journal publishes have been better than previous years?
Is it impossible to think that there could be BOTH more AI slop, AND more “good science”?