Am I or AI? – A useful tool in (scientific) writing

A new article on the use of artificial intelligence (AI) in scientific writing caught my attention. It’s lurid headline says that a “Staggering 90% of biomedical papers now show signs of AI help”, suggesting widespread cheating by authors who let large language models (LLMs) do their work. However, in the article itself it then says that “LLM use was more frequent in abstracts, introductions and discussion sections than in methods and results sections”. This means AI was used to help with more general sections, less with the ones that contained the actual added-value of the papers and the original work of the authors – what was novel about the papers was written more often by the authors than what was known already (which was added with the help of AI). In my view, this makes a crucial difference. 

What is even more important, IMHO, is that the article concedes: “One problem with current AI-detection methods is that they can’t distinguish between a text generated by AI and one that was simply edited for grammar”, before acknowledging that “LLMs are commonly used by people who do not have English as their first language to translate text and polish grammar.” This means that in many cases AI simply levels the playing field between researchers who are lucky enough to be native English speakers and researchers who are not. So far, simply being a native English speaker gave researchers an unfair advantage: They could write faster and better and didn’t need to ask favours from others to proof-read their manuscripts (or even pay translators or scientific language editing services). Native English speakers just needed to be good in their field, others needed to be good in their field AND be good at English. Yet, I would say it is arguably better for science and society if now more researchers (who are good in their field but perhaps not in languages) manage to publish their work and reach a wider audience. 

As it is, it is my impression that some of the opposition to AI comes exactly from those people who – so far – benefitted from a language advantage by being native speakers of English (or having learnt English good enough that it gave them an edge over other non-native speakers). There may be many other issues with AI, ranging from invented references to concerns about data protection, but I think that helping people to write and communicate better is a benefit, not a problem. This is why I have a problem with headlines like the one above: It makes a real difference if a researcher uses AI to improve the language of a manuscript or to write an entire paper. After all, e.g. both SpringerNature itself and also Elsevier offer language editing services (and even “scientific” editing services). What’s the difference between outsourcing the improvement of grammar and style to the service that’s offered by the publisher itself or to an AI? In fact, the authors of the underlying preprint study themselves write that “LLM-assisted writing can be valuable by removing language barriers but at the same time causes concerns about misconduct and fraud”. 

And yes, I wrote this text myself, no LLM polished its grammar or style. I’m not a native speaker and I do make mistakes, but overall I manage fairly well, I think. Sometimes  I do use an LLM to improve the language of texts, though. It’s also a service to the reader to present a polished and easier-to-read text whose language is not tedious and in the appropriate register. In this respect I consider LLMs a tool like a dictionary or a spell-checker (both of which I used for the writing of this text), or like a thesaurus – I still remember how thrilled I was when I discovered the existence of thesauri many, many years ago, before the internet took off. All of a sudden I no longer needed to rely on my (limited) active vocabulary but could use more fitting synonyms and write texts that were more lively. I bought a small “Collins Gem Thesaurus in A-Z Form”, which I loved and still have in my book shelf.