AI Research Boom: More Papers, Less Progress?
▼ Summary
– High-profile retractions of AI-generated scientific papers have raised serious concerns about peer review quality in some journals.
– Researchers from Berkeley and Cornell analyzed millions of pre-publication papers to assess AI’s influence on scientific literature.
– They trained a model to distinguish between human-written and AI-generated text using pre- and post-ChatGPT abstracts.
– The study found that researchers produce far more papers after starting to use AI, and the language quality improves.
– However, the publication rate for these AI-assisted papers has actually dropped compared to the researchers’ prior work.
A recent study from researchers at Berkeley and Cornell provides the first large-scale data on how artificial intelligence is reshaping scientific publishing. The findings reveal a complex picture: while the adoption of large language models (LLMs) has led to a significant surge in the number of papers researchers produce, it has also coincided with a decline in their publication rates in peer-reviewed journals. This suggests that the sheer volume of AI-assisted research is increasing, but its ultimate impact and acceptance within the formal scientific community may be more nuanced.
To investigate this trend, the team analyzed over two million documents from three major preprint archives, arXiv, the Social Science Research Network (SSRN), and bioRxiv, spanning from 2018 to mid-2024. This vast dataset included work from before advanced LLMs were widely available, providing a crucial baseline for comparison. The researchers developed a method to detect AI-generated text by first training a model on human-written abstracts from the pre-ChatGPT era. They then had GPT-3.5 rewrite those same abstracts and trained the model to recognize the statistical differences. This tool allowed them to estimate the likelihood that any given abstract was authored by a human or an AI.
Using this detection method, the team identified a pivotal moment for individual authors: the point at which they first appeared to use an LLM for a submission. By comparing an author’s productivity before and after this adoption point, clear patterns emerged. The data shows that once researchers begin using AI tools, their scientific output, measured by preprint submissions, increases dramatically across all fields studied. However, this boost in quantity does not necessarily translate to success in the traditional publishing pipeline, where the rate of these preprints being accepted into journals has fallen. This disconnect raises important questions about whether AI is enabling more exploratory work or simply flooding the system with lower-quality submissions that struggle to pass rigorous peer review.
(Source: Ars Technica)
