![\"The](\"https:\/\/occamstypewriter.org\/mindthegap\/files\/2026\/08\/Proto-Scientist-smaller.jpg\")
One from the archives: writing up my scientific results as a summer intern at the US NIH<\/p><\/div>I have recently written about the perils of scientists relying on generative AI for writing<\/a>. But the topic isn\u2019t going away, and every day, more alarm bells seem to go off.<\/p>\n Two opposing opinion pieces on this topic were published in the same week this month, one in Science<\/a>, and the other in Nature Africa<\/a>. In the first, early-career clinician scientist Bastien Le Guellec recounts how all his drafts used to get massacred by his supervisor, and it took ages of iterative input to refine manuscripts. So he turned to ChatGTP for help. The more he relied on ChatGPT to draft his papers, the more his supervisor praised his outputs, and the more he felt his productivity and success increased. <\/p>\n However, he began to sense his joy in writing was ebbing, and he no longer experienced the pride and ownership that had made writing so special. What\u2019s more, he started to worry that in losing his direct engagement with the ideas, he was allowing the AI to uncritically tell the tale he wanted to believe<\/em>, whereas before, writing things himself tended to expose the weaknesses in his own arguments and force him to grapple with them. So he decided to ditch the chat and produce his work from scratch again, thereby restoring the role of the writing process in developing and refining ideas. The act of writing, he concludes, is not merely a way of expressing ideas, but a vital process to exert the critical focus that all scientists need to have.<\/p>\n In the second piece, which has incited a lot of ridicule on social media, senior professor Bertie Fielding recounts the fallout after a journal flagged up one of his pieces as being wholly AI-generated. The journal was not wrong: Fielding admits unapologetically that he created almost none of the prose, but remained the \u201cthought leader\u201d of the contents and ideas. To him, generative AI is a \u201cghostwriter\u201d in service to his human thoughts.<\/p>\n I personally think both authors have valid viewpoints, but as someone who finds writing one of her greatest joys in life, I am much more aligned with Le Guellec than with Fielding. As a bit of fun, I asked Copilot what it thought of the two pieces \u2013 as well as the blog post I wrote last time about the perils of LLM. It agreed with me that all of us had useful insights, but pushed back a bit on our all-or-nothing attitude. There was, perhaps, a compromise, whereby scientists could use AI like a tool without it completely replacing the critical thinking unearthed in the process of generating original text.<\/p>\n Not to put words in Copilot\u2019s mouth \u2013 here is exactly what it concluded:<\/p>\n You occasionally sound as though there is a binary choice between “human writing” and “AI writing.” In practice, I suspect many excellent scientists are already occupying a middle ground. A non-native English-speaking postdoc who uses AI to improve readability may not be surrendering any intellectual ownership at all. Likewise, using AI to shorten a paragraph is very different from asking it to devise the central hypothesis of a fellowship application.<\/em><\/p>\n Nevertheless, I still feel a key question here remains unresolved: how can a junior scientist learn the craft of critical thinking and argumentation if they no longer experience one of the main ways we all traditionally absorbed this essential training? <\/p>\n Like Le Guellec, I too learned how to write a scientific paper through the painstaking and iterative process of submitting drafts to my supervisor. The first manuscript I wrote during my PhD came back with a bloodbath of red ink (literally, as this was before online editing and tracked changes). It was embarrassing and exposing seeing my rookie errors underlined and frankly annotated with (a) exactly why it was so bad and (b) what word choices could fix the issue. Nonetheless, through this painstaking apprenticeship-like process, I slowly learned how to make cogent, succinct arguments; to overcome objections in advance; to apply narrative and flow and transition to explain, persuade and truly engage. And when I shared the first draft of my fourth paper, I was shocked when she returned it with only a few minor tweaks. I had, apparently, finally mastered this essential craft of my trade. I was a true writer \u2013 and scientist \u2013 at last. <\/p>\n Over the years, I continued to hone my skills, branching out into other forms of writing (popular science, news, and of course, my three novels). None of these additional works were ever completed in a vacuum \u2013 I was continually exposed to editors, some of whom were ruthless, most trying to improve my writing (appreciated), others seeming to want to replace my voice with theirs (resented). But I was building on that solid base of hard-won, learn-by-doing lessons obtained during my PhD years. And until recently, I was paying it all back by editing the writing of my trainees, every bit of feedback as uncompromising and meticulous as I\u2019d received at the same stage.<\/p>\n Of course, now there\u2019s a difference. Starting sometime last year, manuscripts increasingly come to me already beautifully polished by generative AI. Perfect language has the tendency to lull me into a false sense of security: I feel my own critical thinking skills struggling to crack the persuasive and confident tone, like fighting through a cloud of magic fairy dust. It somehow makes me work twice has hard to find the sloppy logic or the missing pieces. And that\u2019s me, a scientist with decades of experience.<\/p>\n But consider the early-career researchers. Instead of submitting an imperfect draft to me, and learning what\u2019s wrong with it, they are receiving perfectly crafted prose from their pet AI. Not only is the AI making it harder for me, the supervisor, to spot the errors, but it is depriving the researcher of that valuable back-and-forth training in critical writing skills during the ideas-crafting stage \u2013 and possibly, of the critical thinking skills themselves, if you buy the arguments of people like me and Le Guellec. <\/p>\n I feel that we, as a profession, are on the cusp of some great divide. Maybe the rise of LLM-generated scientific texts will continue and eventually take over, for better or worse. Alternatively, perhaps there will be a backlash: people might become increasingly sensitive to the too-good-to-be-true tone, and authentic writing might come back into fashion \u2013 just like sourdough during the Pandemic. <\/p>\n In the meantime, all I can do is be vigilant and try my utmost to make sure that those under my care truly understand the fire they\u2019re playing with.<\/p>\n","protected":false},"excerpt":{"rendered":" I have recently written about the perils of scientists relying on generative AI for writing. But the topic isn\u2019t going away, and every day, more alarm bells seem to go off. Two opposing opinion pieces on this topic were published … Continue reading