The Irony of Clarity

Colleagues recently told me that my blogs have started to sound like they are written by artificial intelligence (AI). To be honest, it might have hurt less if I was sure that they actually read my blogs (I have been writing them since 2012).

Nevertheless, these days it’s a question that cuts to the heart of what it means to be an author of any form. Have I spent four decades honing a style, learning the craft of scientific communication, only to arrive at a voice indistinguishable from an AI? It is also a particularly painful experience for me. As I was accused (incorrectly) of plagiarism early in my career (my first ever student essay) [1], I have been somewhat ‘sensitive’ ever since.

The accusation of electronic enhancement isn’t unique to me. In a recent study, 78.6% of reviewers evaluating an AI-generated manuscript were unable to identify its origin, highlighting the profound challenge we now face in distinguishing human from machine in scientific prose [2].

The Training of a Scientific Voice

As scientists we are trained to be impersonal. From our first research proposal to our hundredth journal article, the principles are drummed into us: avoid emotional language, structure arguments logically, explain complex ideas with clarity, and eliminate ambiguity.

Scientific writing is the art of making the complicated accessible and the subjective seem objective. It prizes precision and consistency [3][4]. The gold standard is a text where the argument flows so logically that the reader’s attention is held solely by the data and its interpretation, never by the writing itself [5][6].

When I started in my first job, I was desperate to get my name on publications. However, in our laboratory a junior scientist was required to serve their apprenticeship and earn their stripes. In a very real sense, I succeeded (eventually). I achieved my goal of crafting prose that is clear, concise, and focused through the repeated correction of colleagues, my posy graduate professor and my first House Editor. At times the experience was brutal. The problem, as I now see it, is that these lessons on creating clear structure, consistent grammar and logical flow, are precisely the hallmarks of the Large Language Models (LLMs) that power tools like ChatGPT [5].

Studies are beginning to show that AI-generated scientific content exhibits clear and measurable differences from human-written text in features like readability and perplexity (a little pun for AI aficionados). But these are aggregate differences. At the level of a single, well-crafted 1200-word blog, the features that make writing ‘good’ by scientific standards are becoming the very features that trigger both human and electronic AI detectors who, in the current climate, expect to see the fingerprints pf AI on almost every written word.

The criticism reveals a fundamental misdirection. My colleague’s comment suggests they are looking in the wrong direction. It isn’t that AI writes like me. It’s that an AI, trained on billions of words generated by generations of writers, scientists, and editors [7][8], learned to emulate the same professional standards of clarity and objectivity that I spent my career mastering [3][4][9]. The irony is that some of the training materials were written by me. AI didn’t teach me to write. I, and countless others like me, unwittingly taught it.

The Rise of the Algorithmic Aesthetic

The accusation of being ‘AI-like’ stems from a broader cultural shift. People are increasingly judging authorship based on superficial characteristics. Good grammar? That’s AI. Polished prose? AI. Structured, perfectly punctuated paragraphs and regular posts? Definitely AI [10][11]. We have arrived at a bizarre new point where the traditional marks of a professional writer are now viewed as signs of automation.

This is the algorithmic aesthetic: the assumption that only a machine can be flawless, and that human writing is, by its very nature, messy and idiosyncratic. This perspective overlooks the decades of deliberate effort behind the final product. The ‘flawless’ prose of an experienced writer is the result of countless drafts, rejections, and revisions. It’s the product of a lifetime spent learning to distrust words until they have been interrogated. Research on expertise distinguishes between routine and adaptive expertise. Routine experts develop efficient, stable patterns of performance; adaptive experts retain flexibility and the capacity to innovate [5][12]. In scientific publishing, routine expertise is often rewarded: editors ask authors to remove personality, shorten sentences, simplify language, avoid opinion and follow reporting guidelines [4][13]. Over time, this can push writers toward a narrow band of acceptable style.

I know I have become formulaic. Decades of writing produces habits. I regularly introduce arguments in a certain way. I use similar transitions. I favour balanced, evidence-based conclusions. For me this predictability is the essence of expertise, it also helps me construct articles relatively quickly – in my head it comes together like Lego. It’s also the essence of the system that is now being used to question me. The criticism forces a deeper reflection: Did I sacrifice too much of my individual voice on the altar of scientific objectivity? I regularly suppress the little devil inside me that wants to make rash conclusions [13]. It’s becoming a question that scientific publishing itself must contend with [7][15][16].

In a twist of Leonardo da Vinci’s "Poor is the pupil who does not surpass his master." I am creating a new euphemism, “Arrogant is the master who cannot learn from the student.” And I have. I have virtually eliminated en and em dashes from my writing – despite having warm memories of my first editor, Flick (RIP), teaching me how to use them correctly. I squeeze as many of my own experiences into each new article. I speculate as wildly as 40+ years of scientific indoctrination will let me. And occasionally I misuse apostrophes.

Beyond Imitation

And this is where the true argument lies. Large language models excel at reproducing this routine expertise. They generate objective, well‑structured prose that conforms to editorial expectations [3][9][17]. The machine can assemble information, but it cannot genuinely express 40+ years of failures. It cannot recall the wise words of the PhD supervisor who shaped my research. It hasn’t navigated the murky ethical debates with editors or felt the sting of a rejection letter [6][11]. It has no lessons learned from interactions with patients, from difficult conversations with regulatory bodies, or from the collective wisdom of mentors and collaborators [5][11]. Those experiences don’t just provide facts. They shape judgement, and they form the bedrock of my stories [4][8].

In a recent study, researchers designed the ‘Project Rachel’ experiment, creating an AI academic identity that successfully published papers and even received a peer review invite [18]. This underscores the system's ability to mimic the formalities of scholarship. However, the AI lacked the genuine, formative intellectual and emotional journey that shapes a human scholar’s perspective. It doesn’t understand the drive that possesses someone to write and write some more.

The voice of a machine is a composite of its training data; my voice is the culmination of my life’s work and aspirations. I will reiterate (as I think I have already implied it), good scientific writing is not simply correct language; it is language infused with accumulated wisdom, uncertainty, and the willingness to acknowledge limits [4][8].

Should I change how I write? Can I? Should I deliberately inject more personality, more anecdotes, into my papers to ‘prove’ I’m human? After more than 300 blogs I am not sure I have many stories left and risk undermining the very standards that make scientific communication accessible and trustworthy [3][13]. Perhaps I need to stop writing and go and get some new experiences to pack into the next 300.

There is also a social dimension. In some cases, the accusation of ‘AI‑like’ writing comes from people who do not themselves write regularly, or who feel threatened by the idea that someone can publish frequently without outsourcing their voice to a machine [10][16]. Organisational behaviour research suggests that professionals may react defensively when automation appears to encroach on domains they have long considered markers of expertise and identity [11][14]. Labelling another’s work as ‘AI‑generated’ can become a way of dismissing its legitimacy without engaging with its substance.

Conclusion

After four decades of trying to write with clarity, precision and objectivity, I find myself being told I write like AI. The irony is difficult to ignore. I feel this issue extends beyond AI detection. It speaks to a much broader anxiety among experienced professionals: that decades spent mastering a craft are suddenly being viewed through the lens of automation [10][11][14]. It also sidesteps the ultimately unproductive debate over whether AI detectors ‘work.’ Accuracy studies already show that detectors can misclassify both human and AI texts, raising serious concerns about fairness and trust in academic evaluation [10][16]. Will new statistical word ‘tagging’ – using specific word combinations - remove all doubt on ownership? Editors might feel so.

The more constructive question is this: what should journal editors and scientific communities value in writing? If we continue to demand soulless reporting of numbers, we will invite further convergence between human and machine prose. If, instead, we allow authors to speculate responsibly about their findings, to situate results within lived experience and disciplinary history, we may deepen our articles in ways that no algorithm can replicate [4][8][16].

Authorship, in this new era, cannot be defined solely by style. It must be grounded in accountability, intellectual contribution, and the irreducible complexity of a human life spent thinking, failing, revising and, occasionally, being told that one sounds like a machine.

References

1. Hardman TC. (2018). Plagiarism in Scientific Writing: Lessons Learned.
2. Öztürk, A., et al. Artificial intelligence as author: Can scientific reviewers recognize GPT-4o-generated manuscripts? Am J Emerg Med 97 (2025): 216-219.
3. Khera R, et al. Scientific Writing in the Era of Large Language Models: A Computational Analysis of AI- Versus Human-Created Content. Stroke. 2025 Oct;56(10):3078-3083.
4. Coberley D, Dux Speltz E. Modeling scientific uncertainty in language: Applied linguistic insights from human and artificial intelligence texts. Research Methods in Applied Linguistics. 2026.
5. Weng Z. Investigating routine and adaptive expertise of experienced teachers in English for academic purposes writing. J English Acad Purposes. 2025.
6. Rogers B. To be… an author. Ann R Coll Surg Engl. 2025 Nov;107(8):539.
7. Kim D‑ Peer Review Integrity Framework (PRIF‑2026‑01): Comprehensive Editorial Protocols for Detecting AI‑Generated Manuscripts in Scientific Journals. Zenodo. 2026.
8. Mehta N, et al. A call for clarity: a unified checklist for reporting use of large language models in writing scientific manuscripts. Res Integr Peer Rev. 2026 May 10;11(1):24.
9. Jain A, Gupta V. Readability and quality assessment of human versus artificial intelligence‑generated plain language summaries across six large language models. Curr Med Res Pract. 2026.
10. Erol G, et al. Can we trust academic AI detective? Accuracy and limitations of AI-output detectors. Acta Neurochir (Wien). 2025 Aug 7;167(1):214MTSU Writing Center. SWC 7: Writing in the Sciences.
11. Freitas G. Challenges and ethical issues of maintaining publication integrity and accuracy with the use of artificial intelligence. JAAD Rev. 2026;8:63‑68.
12. Kimchi, I., Reshef, N., Vasileva, A., & Arnon, I. (2026). Enhancing second language learning through orthographical changes. Language Learning and Development, 1–24.
13. SWC 7: Writing in the Sciences.” MTSU Writing Center
14. Hardman TC (2026). Why Chaos Tempts Us, and When We Should Listen.
15. Responsible and transparent use of AI in scientific publishing. Nat Comput Sci. 2026;6:803.
16. Kock B, et al. Ensuring peer review integrity in the era of large language models: A critical stocktaking of challenges, red flags, and recommendations. Eur J Radiol Artif Intell. 2025;2:100018.
17. Barrot JS. Trinka: Facilitating academic writing through an intelligent writing evaluation system. Computers and Composition. 2025.
18. Monperrus M, Baudry B, Vidal C. Project Rachel: Can an AI Become a Scholarly Author? arXiv. 2025.