Like any crappy human writer, AI chatbots have a tendency to overuse specific words — and now, scientists are using that propensity to catch their colleagues when they secretly use it in their work.
As the New York Times reports, scientists estimate, based on an analysis of those overused terms, that there could already be hundreds of thousands of academic papers written with the assistance of AI.
In a new study published in the journal Science Advances, researchers from Germany’s University of Tübingen identified some 454 words that large language models (LLMs) use frequently — terms like “garnered,” “encompassing,” and “burgeoning” — and found that anywhere from 13.5 to 40 percent of biomedical article abstracts were written entirely or with assistance from AI.
With roughly 1.5 million papers indexed each year on the academic journal database PubMed, that means that at least 200,000 of those papers could have been written with the help of LLMs. That whopping figure may, as the NYT notes, be conservative when accounting for any intentional editing of AI-generated text.
While some deceptive journal writers take pains to conceal their AI use, others didn’t seem to care who knew. In an example posted on X by Arizona State University computer scientist Subbarao Kambhampati, for instance, the “writers” of a low-quality radiology journal left in an acknowledgement that it had been penned by a chatbot.
“I’m very sorry,” the paper quote reads, “but I don’t have access to real time-information or patient-specific data as I am an AI language model.”
Not all mistakes are going to be that obvious. Unless you’re familiar with the term “regenerate response” — an option on ChatGPT that forces the chatbot to rework a shoddy answer — you could easily miss it being sprinkled throughout respected journals, the blog Retraction Watch found back in 2013.
That same blog also flagged a frustrating debacle involving a paper about millipedes with completely made-up references, which was initially withdrawn from a pre-print server, only to reappear online on a different academic database with those same hallucinated sources.
And let’s not forget the journal that was forced to retract a paper after observers noticed it was filled with nonsense, including an AI-generated image of a rat with comically gigantic genitals.
Complicating matters, academics are taking pains to hide their AI use. As Kambhampati told the NYT, academics have even started changing the way they write so their work won’t be mistaken for AI, removing terms like “delve” that are overused by LLMs.
In the Science Advances paper, the Tübingen researchers proffer that all this AI use in academic journals, if corroborated, could be an “unprecedented impact on scientific writing in biomedical research, surpassing the effect of major world events such as the COVID pandemic.”
To the mind of coauthor Dmitry Kobak, however, the whole thing seems counterintuitive.
“I would think for something as important as writing an abstract of your paper,” he told the NYT, “you would not do that.”
More on AI: OpenAI Says It’s Hired a Forensic Psychiatrist as Its Users Keep Sliding Into Mental Health Crises
Source link
#Bombshell #Research #Finds #Staggering #Number #Scientific #Papers #AIGenerated
