Recently, I attended a faculty meeting focused on artificial intelligence (AI) policy for our graduate students. After three years of creating and revising guidelines, we found ourselves debating whether to prohibit AI usage for thesis proposals and dissertations, even though enforcing such a ban would be practically impossible. One colleague referenced a popular MIT Media Lab study from the summer, which demonstrated a decline in neural connectivity among individuals who wrote using ChatGPT. Referred to as “cognitive debt” by the researchers, the study, despite its limitations, encapsulated a deep-seated concern that has emerged since the introduction of ChatGPT. If writing is a form of thought and the challenge of articulating an idea is essential for fully grasping it, then tools that simplify this process might undermine scientists’ cognitive abilities necessary for making significant discoveries.
For a considerable time, I’ve been contemplating the intersection of AI and scientific writing, and I find myself torn between two perspectives that I can’t wholly embrace. The concern about “cognitive debt” or skill erosion resonates with me, yet I also see merit in the counterargument that such concerns might be exaggerated. My quest for conclusive evidence on the matter has led me to believe that definitive answers may be elusive, especially regarding those actively engaged in scientific practice.
Anyone who engages seriously in writing recognizes its vital role in the thinking process. Consider a situation where you attempt to draft an aims page that originally seemed clear in your mind, but falters when put to paper. The logic that felt solid in your head begins to disintegrate on the page. After hours of wrestling with the text, you realize that what appeared to be a mere communication issue is, in fact, a deeper cognitive challenge that only becomes apparent through the writing process. Proponents of the notion that “writing is thinking” — whom we can refer to as “cognitive traditionalists” — echo this sentiment. As the writer Flannery O’Connor famously noted, “I write because I don’t know what I think until I read what I say.” Similarly, physicist Richard Feynman remarked that his notebooks were not just records of his thought process but constituted his actual thinking process.
I
n research backing these assertions, there exists a robust body of literature on writing as a tool for learning, spanning decades. Studies demonstrate that the act of composing text yields measurable learning gains that exceed what mere studying would produce. A theoretical framework based on connectionist cognitive models suggests that understanding often remains implicit and only becomes accessible through writing. This act helps convert vague intuitions into explicit propositions, allowing writers to evaluate their reasoning against specific goals, thus highlighting gaps in their logic. This mechanism underscores the notion that writing serves as an essential learning tool beyond formal educational contexts. Evidence from fields such as aviation and medicine indicates that reliance on automated assistance can accelerate skill degradation, even among seasoned professionals. My interpretation of the data leads me to believe that depending on AI for writing may hinder the development of writing skills in aspiring scientists and jeopardize those abilities in established experts.
The pivotal issue is whether the decline in unassisted writing is significant for the scientific community. Proponents of AI, often referred to as “AI apologists,” present arguments in their favor. Their first point is that the cognitive benefits attributed to traditional writing may not stem from the act of writing itself. Instead, it could be about externalization — the process of transforming internal representations into a visible format where they can be analyzed and revised. If this is indeed the case, then discussions with colleagues, defending ideas in lab meetings, or explaining concepts to peers in different fields could yield similar insights. Psychologists Daniel Kahneman and Amos Tversky famously took long walks together to generate ideas that contributed to their Nobel Prize-winning contributions in behavioral economics. If externalization is the key, some may argue that conversing with a suitable AI system could preserve cognitive advantages by articulating and clarifying implicit reasoning. Although the medium of communication changes, the underlying cognitive mechanisms might remain intact, suggesting that for scientific work, valuable cognitive processes can occur outside of traditional writing, therefore making the offloading of writing less detrimental than cognitive traditionalists propose.
AI proponents do offer empirical data in support of their position. A recent study published in Science revealed that researchers who utilized large language models increased their publication output by one-third to one-half, with the most significant gains seen among non-native English speakers. This finding is substantial, as the challenges faced by non-native speakers are well-documented; if AI can help mitigate these obstacles, it represents a meaningful contribution to equity within the scientific community. However, while these findings may indicate a rise in productivity among scientists, they do not necessarily reflect the quality of the underlying research. Concerns have been voiced by many scientists that the influx of AI might merely expedite paper production without addressing the substantive impediments to genuine scientific advancement.
S
o here we stand: cognitive traditionalists have educational research and expert opinion on their side, while AI supporters possess evidence of increased productivity and equity considerations. Yet neither side provides conclusive evidence about the repercussions of having working scientists — as opposed to students or white-collar professionals — outsourcing their scientific writing tasks to AI. Does this dependence affect their comprehension of their own research? Are their research initiatives less coherent? Do they lose the ability to identify reasoning gaps? Unfortunately, such studies are likely impossible to conduct effectively. It would be nearly unfeasible to randomize established scientists to “write your own grants” versus “delegate them to AI” conditions and track their careers over an extended period.
