Shall we give up writing scientific papers to AI?

Abstract

Artificial Intelligence (AI) tools for language processing are currently being used in many fields where the write up of a summary or a commentary is required. Most internet content is, as a matter of fact, AI-generated ones, sometimes supervised by a human Editor. Scientific writing makes no exception, with AI tools being increasingly adopted for text generation. Should we be alarmed by AI text generation tools or let AI write our papers? Writing a scientific paper is just the last task in a logical scientific workflow, which I tried to represent in Scheme 1. Typically, such a workflow is initiated by curiosity, possibly linked to unanswered questions in a given scientific field that generate the idea of researching and investigating the underlying phenomenon. We typically associate this stage with the creativity of a researcher. This initiating phase is then followed by a likewise creative stage where researchers start planning experiments (or simulations) to address the questions raised by point 1. This stage requires a combination of global field knowledge, technical skills and inventive capabilities. Researchers in the group then execute the experiments (or calculations) planned at stage 2. This is when deep technical skills come into play. The quality (accuracy) of the data produced at stage 3 many times equally contribute to the research success as it is planned in stage 2. Analysis of the results produced at stage 3 and their interpretation according to commonly accepted physical or chemical laws is also a creative stage, including knowledge of the literature of the field. Although apparently less exciting than 1 and 2, this is, in my opinion, what really makes the difference between having an interesting result, possibly of technical relevance, and having a scientific result worth being published to address a broader audience. Last but not least, we have to write the paper according to some commonly accepted structure (e.g., Introduction, Methods, Results and Discussion, and Conclusions), resembling somehow in parallel the structure to the workflow of Scheme 1. Without a well composed write up, the efforts of stages 1–4 go unnoticed. Writing a good scientific paper is usually a cooperative effort involving almost the entire research team, with a different degree of contribution depending on the role, skills, and personal vocation. In my experience as a researcher, I have to admit that sometimes the last stage can be tedious, especially when one needs to start writing from scratch the “Introduction” section. Those first two or three paragraphs, embodying literature references, may represent an obstacle which significantly delays the paper write up, leading to coauthors emailing “about the status of our paper”. AI tools can be of great help in overcoming such an “activation barrier”, (1) although one cannot expect much depth or insight being achieved by a typical chat bot, with a similar output to that achieved by performing an Internet search on the subject and looking at a summary of the first two or three results. One may be tempted to extend the use of AI tools to the write up of the entire paper, but most of the time (with the probable exception of uncritical reviewing of a given field) AI tools would likely generate stylistically good but scientifically trivial text. We now have to face reality and use our own creativity to discuss the main paper outcome and present them in a catchy way. As a matter of fact, one can probably recognize a scientific breakthrough even if its description is not stylistically outstanding. Most notably, we cannot expect AI tools to make connections beyond what they already know. It is interesting in this respect to look at the paper probably representing one of the major scientific breakthroughs of last century, i.e., Watson and Crick’s DNA structure. (2) After critically reviewing previous DNA models, these authors proposed the double helical structure with adenine-thymine/cytosine-guanine pairing on the basis of known bond lengths and angles and van der Waals radii. What I think is impressive in this work is the conclusion paragraph, copied here: “It has not escaped to our notice that the specific pairing we have postulated immediately suggest a possible copying mechanism for the genetic material”. This kind of broader view and connection capability is what still makes human intelligence superior to AI and what, in my opinion, makes a scientific breakthrough recognizable even in the absence of fancy text being generated to describe it. In this sense I am not personally scared by the use of AI tools in assisting scientific writing, because science is a complex process that still requires human intelligence and supervision. As a consequence, I think researchers should focus more on the breadth and depth of their results rather than on the quality of text associated with their discussions. Probably AI has already changed the way scientific papers are written, but I hope it will not change the way researchers think and perceive reality. I conclude this editorial by recalling the AI Best Practices and Policies at ACS Publications: (3) “The use of AI tools for text or image generation should be disclosed in the manuscript within the Acknowledgment section with a description of when and how the tools were used. For more substantial use cases or descriptions of AI tool use, authors should provide full details within the Methods or other appropriate section of the manuscript.” This article references 3 other publications. This article has not yet been cited by other publications.