Analysis of AI usage in scientific writing by looking at word usage

I saw this scientific american article [1], which covers an analysis of AI usage in scientific writing. Since this has been a hot topic on RH, I thought others might be interested as well. There is also a paper on arxiv [2] by Andrew Gray, which the sciam article is based on, which more details. The claim at the beginning is that 1% of articles in 2023 had AI generated content.

What I found interesting is how the analysis was done. Essentially they took a list of words which are associated to output from LLMs and compared the usage of those words against a list of control words.
From [2]

In the sciam article, there is a nicer illustration of the word frequencies over the past 10ish years than in the arxiv article. You can see that the usage of control words scales with the number of publications and dips in 2023 due to fewer publications. However, the AI-words stay the same or go up.

It's also noted that part of this could be to collective change in language usage, but the change is pretty striking, especially when combining multiple words into groups:
From [2]

It's not gradual like you might expect from changes in how people use language. It seems to me that this kind of analysis will get harder to do as LLMs get better at mimicking human text, and/or people who use it become aware of the odd features in the output and either remove it themselves or ask the LLM to.

[1] https://www.scientificamerican.com/article/chatbots-have-thoroughly-infiltrated-scientific-publishing/

[2] https://arxiv.org/abs/2403.16887

Preprint

Preprint

Analysis of AI usage in scientific writing by looking at word usage