Massive study detects AI fingerprints in millions of scientific papers - Other Sciences - Social Sciences - Education

Words showing increased frequency in 2024. (A) Frequencies in 2024 and frequency ratios (r). Both axes are on a log scale. Only a subset of points are labeled for visual clarity. the dashed line shows the threshold defining excess words (see text). Words with r > 90 are shown at r = 90. excess words were manually annotated into content words (blue) and style words (orange). (B) the same but with frequency gap (δ) as the vertical axis. Words with δ > 0.05 are shown at δ = 0.05. Credit: Science Advances (2025). DOI: 10.1126/sciadv.adt3813

Chances are that you have unknowingly encountered compelling online content that was created, either wholly or in part, by some version of a Large Language Model (LLM). As these AI resources, like ChatGPT and Google Gemini, become more proficient at generating near-human-quality writing, it has become more difficult to distinguish between purely human writing from content that was either modified or entirely generated by LLMs.

This spike in questionable authorship has raised concerns in the academic community that AI-generated content has been quietly creeping into peer-reviewed publications.

To shed light on just how widespread LLM content is in academic writing, a team of U.S. and German researchers analyzed more than 15 million biomedical abstracts on PubMed to determine if LLMs have had a detectable impact on specific word choices in journal articles.

Their investigation revealed that since the emergence of LLMs there has been a corresponding increase in the frequency of certain stylist word choices within the academic literature. These data suggest that at least 13.5% of the papers published in 2024 were written with some amount of LLM processing. The results appear in the open-access journal Science Advances.

Since the release of ChatGPT less than three years ago, the prevalence of Artificial Intelligence (AI) and LLM content on the web has exploded, raising concerns about the accuracy and integrity of some research.

Past efforts to quantify the rise in LLMs in academic writing, however, were limited by their reliance on sets of human- and LLM-generated text. This setup, the authors note, "…can introduce biases, as it requires assumptions on which models scientists use for their LLM- assisted writing, and how exactly they prompt them."

In an effort to avoid these limitations, the authors of the latest study instead examined changes in the excess use of certain words before and after the public release of ChatGPT to uncover any telltale trends.

The researchers modeled their investigation on prior COVID-19 public-health research, which was able to infer COVID-19's impact on mortality by comparing excess deaths before and after the pandemic.

By applying the same before-and-after approach, the new study analyzed patterns of excess word use prior to the emergence of LLMs and after. The researchers found that after the release of LLMs, there was a significant shift away from the excess use of "content words" to an excess use of "stylistic and flowery" word choices, such as "showcasing," "pivotal," and "grappling."

By manually assigning parts of speech to each excess word, the authors determined that before 2024, 79.2% of excess word choices were nouns. During 2024 there was a clearly identifiable shift. 66% of excess word choices were verbs and 14% were adjectives.

The team also identified notable differences in LLM usage between research fields, countries, and venues.

by Charles Blue, Phys.org

edited by Andrew Zinin 

Source: Massive study detects AI fingerprints in millions of scientific papers