The first core controversy regarding the use of artificial intelligence in medical writing is the problem of epistemic reliability. Large language models can rapidly generate clinical background sections, pathophysiological explanations, mechanisms of rare diseases, and literature overviews; this looks attractive to clinician-authors who are under time pressure (Alkaissi & McFarlane, 2023; Safrai & Orwig, 2024; Pinto et al., 2024). However, this generation process can produce unverified claims, out-of-context generalizations, and, most critically, entirely fabricated or partially incorrect references. This creates a direct patient safety risk because clinical narrative that contains model-generated but evidence-free claims can then be presented as teaching material, a case report, or even a standard of care (Alkaissi & McFarlane, 2023; Safrai & Orwig, 2024). The unresolved core question is this: how can we quantify the probability that AI-originated scientific hallucination leaks into text in ways that may affect patient care, and under which verification thresholds should a text be declared clinically “unreliable”? This should be framed as a research question as follows: “In pathophysiological explanations and treatment suggestions produced by large language models that generate clinical content, what is the proportion of statements that are unverifiable or contain faulty attribution, and how does this proportion vary across specific clinical risk categories (e.g., rare metabolic disease, fertility preservation, post-invasive procedure follow-up)?” (Alkaissi & McFarlane, 2023; Safrai & Orwig, 2024; Thaichana et al., 2025). The reason this study must be conducted is clear. Clinical writing, especially case reports and review-type articles, is learning material for young physicians, and these texts are often used as secondary reference in clinical practice. If AI is generating content that is persuasive but wrong within such texts, that error does not only threaten academic integrity, it directly threatens patient safety. Without quantitatively showing this risk, the warning “use AI with caution” remains abstract and cannot yield a regulatory response (Safrai & Orwig, 2024; Daugirdas, 2025; Ramoni et al., 2024).
The second main controversy is accountability and responsibility. Journals and editorial boards are increasingly explicit that AI cannot be listed as an author, that the scientific and ethical responsibility for the text remains with human authors, and that AI use must be transparently disclosed (Tang et al., 2024; Yoo, 2025; Fettiplace et al., 2025). Yet this position does not resolve two critical deadlocks. First, what does “ultimate responsibility lies with the human” actually mean in practice? If a model generates a paragraph that may pose clinical risk, and a researcher merely polishes the language and signs off, who bears the legal and ethical liability when that content is published? Second, the problem of verifiability of disclosure persists. An author may declare “I only used it for language editing,” but in reality the entire discussion section could have been drafted by the model. This gray zone remains unresolved both ethically and in terms of insurance/liability (Fettiplace et al., 2025; Cohen & Moher, 2025; Miao et al., 2024). The research question that should be posed here is: “What is the degree of mismatch between declared AI use by authors and the actual use inferred from linguistic/statistical traces in the manuscript, and in which publication types (original research, case report, review) is this mismatch most pronounced?” (Cohen & Moher, 2025; Yoo, 2025; Fettiplace et al., 2025). This must be studied because editorial guidelines are built on an assumption of auditability. If the gap between disclosure and actual practice is systematic, then current guidelines merely create a symbolic sense of safety, which is unacceptable from a patient safety perspective, especially in manuscripts that contain clinical recommendations (Fettiplace et al., 2025; Ramoni et al., 2024).
The third unresolved issue is that AI-assisted texts produce content that sounds like clinical guideline language, but is not actually a guideline. Large language models, in high-risk and fast-evolving domains such as hematology, can summarize treatment algorithms, follow-up strategies, and clinical decision points in a tone that sounds fluent, authoritative, and standardized (Ahn, 2024; Yang & Hwang, 2025). Similarly, in areas that are critical for patient safety such as lower extremity wound care, the model can present principles of care in a homogenized template form (Thaichana et al., 2025). On the surface, this looks beneficial because it seems to speed up inter-team communication, enforce linguistic consistency, and improve continuity of care. But most of these model-generated outputs contain no explicit link to an official clinical guideline, no timestamp, no grading of evidence level, and no declaration of the responsible authority. In other words, the model speaks as if it were a guideline, but it is not a guideline. This produces a de facto but invisible “shadow guideline” in the health system. These shadow guidelines can become normative for junior clinicians, and can shift local decision-making authority into the hands of the single model output (Ahn, 2024; Ramoni et al., 2024; Weidmann, 2024). The research question here is: “To what extent are AI-generated clinical care recommendations aligned in content and timing with current evidence-based guidelines, and in which subdomains (diagnosis, treatment selection, follow-up frequency, patient counseling) does this alignment break down?” (Ahn, 2024; Yang & Hwang, 2025; Thaichana et al., 2025). This work is necessary because without a systematic measurement of that alignment gap, we cannot define under which conditions AI-assisted clinical text is acceptable for patient safety. The absence of quantitative risk thresholds means that regulatory actors (hospital ethics boards, quality units, professional societies) are not given an intervention standard and the burden is shifted to the individual clinician’s intuitive judgment (Ramoni et al., 2024; Weidmann, 2024).
The fourth core controversy concerns the reliability of detection technologies. Editors and journals want to make AI use transparent and, at the same time, want to detect AI-written content that manipulates references or paraphrases source material to the edge of plagiarism. The AI detection tools being deployed for this purpose are not yet stable. Findings from the clinical and behavioral health literature show that these tools sometimes flag human-written text as AI-generated (false positives) and sometimes miss AI-generated text and treat it as human-written (false negatives) (Flitcroft et al., 2024; Popkov & Barrett, 2025; Khera et al., 2025). Both error types matter for patient safety. In the false positive case, a legitimate clinical warning text may be labeled “written by AI, therefore unreliable,” and pushed out of the system, suppressing a potential safety signal. In the false negative case, a clinically risky, unverified recommendation generated by a model may be accepted as human-written, pass through without extra scrutiny, and get published (Flitcroft et al., 2024; Khera et al., 2025). The research question that must be asked is: “How do false positive and false negative rates of AI detection tools vary across clinical domains, and in which content types (treatment recommendation, pathophysiology explanation, patient information material) do these misclassifications pose the highest patient safety risk?” (Popkov & Barrett, 2025; Flitcroft et al., 2024; Khera et al., 2025). This study is necessary because the editorial process is already inclined to trust these tools. If that trust is blind, journals’ claims to protect patient safety are technically baseless. Unless these error profiles are demonstrated meta-analytically, a policy of accepting or rejecting manuscripts based on detector output cannot be defended on patient safety grounds (Fettiplace et al., 2025; Yoo, 2025).
The fifth and still unresolved issue is the economy of human voice, professional subjectivity, and responsibility. AI is presented as a powerful equalizer particularly for researchers who are not native speakers of English, because it can rapidly produce the fluent scientific English expected by high-impact journals. This lowers the publication barrier for researchers, broadens access to scientific communication, and has the potential to increase the visibility of geographically and historically marginalized regions in the literature (Hwang et al., 2023; Johnson & Rubo, 2025; Jembu & Balang, 2025; Al Salti et al., 2025). Yet the same tool also homogenizes the academic text by erasing the clinician’s own experiential reasoning, for instance the justified intuitive decision-making behind management of a rare complication. In other words, the critical micro-decisions in patient care, and the clinical reasoning behind them, get flattened by AI into “standard academic discourse.” This is dangerous in two ways. First, it reduces transparency of clinical reasoning: an external reader can no longer distinguish which parts of the decision were based on human observation and which parts were grounded in formal evidence. Second, when an error occurs, the trace of the specific human author becomes blurred; the language is perfect, but it is no longer clear whose decision it actually was (Matsubara, 2025; Cohen & Moher, 2025; Weidmann, 2024). The research question in this space is: “In AI-assisted medical manuscripts, by how much does the proportion of passages that explicitly present the clinician’s individual clinical reasoning decrease when compared to fully human-written, comparable texts, and in which clinical areas does this reduction most weaken decision transparency relevant to patient safety?” (Matsubara, 2025; Johnson & Rubo, 2025; Weidmann, 2024). This work is imperative because clinical safety culture depends on the traceability of error. If AI, by producing seemingly flawless language, erases the traces of individual reasoning, then root cause analysis after an adverse event becomes harder, and this directly threatens patient safety (Matsubara, 2025; Ramoni et al., 2024).
In conclusion, the dataset points to three critical gaps. First, the quantitative assessment of evidence reliability is not yet systematic; we do not know how often and with what clinical severity hallucination, fabricated references, and out-of-context clinical recommendations appear (Alkaissi & McFarlane, 2023; Safrai & Orwig, 2024; Daugirdas, 2025). Second, the responsibility chain has been normatively defined but not empirically tested; because the divergence between declared AI use and actual AI use has not been measured, the auditability of accountability remains an assumption (Tang et al., 2024; Fettiplace et al., 2025; Yoo, 2025). Third, the way AI output generates guideline-like language may be indirectly weakening clinical quality assurance by delegating clinical decision safety to shadow guidelines, and we still do not know the magnitude, distribution, or which patient groups are most exposed to this risk (Ahn, 2024; Ramoni et al., 2024; Weidmann, 2024; Thaichana et al., 2025). Work along these three axes is not just of academic interest. It builds the empirical infrastructure that regulators, medical malpractice and liability frameworks, patient safety committees, and journal editorial boards need for decision-making. This is why these questions are urgent and why they are necessary for defining safe boundaries for AI-assisted writing in medicine (Fettiplace et al., 2025; Miao et al., 2024; Cohen & Moher, 2025).
References:
Ahn, S. (2024). The transformative impact of large language models on medical writing and publishing: Current applications, challenges and future directions. Korean Journal of Physiology & Pharmacology, 28(5), 393–401. https://doi.org/10.4196/kjpp.2024.28.5.393
Al Salti, M., Al Yahayei, A., Al Shamsi, M., Al Dughaishi, E., Al Naseebi, B., Alhadhrami, A., & Al Maktoumi, T. (2025). Perceptions of nursing faculty on utilizing AI tools in academic writing and publication productivity: A cross-sectional study. Nursing Forum, 2025(1), Article 7447348. https://doi.org/10.1155/nuf/7447348
Alkaissi, H., & McFarlane, S. I. (2023). Artificial hallucinations in ChatGPT: Implications in scientific writing. Cureus, 15(2), e35179. https://doi.org/10.7759/cureus.35179
Cohen, J. F., & Moher, D. (2025). Generative artificial intelligence and academic writing: Friend or foe? Journal of Clinical Epidemiology, 179, Article 111646. https://doi.org/10.1016/j.jclinepi.2024.111646
Daugirdas, J. T. (2025). Use of artificial intelligence in scientific writing: The danger of trying too hard to please. Hemodialysis International, 29(4), 430–433. https://doi.org/10.1111/hdi.13270
Fettiplace, M. R., Bhatia, A., Chen, Y., Orebaugh, S. L., Gofeld, M., Gabriel, R. A., Sessler, D. I., Lonsdale, H., Bungart, B., Cheng, C. P., Burnett, G. W., Han, L., Wiles, M., Coppens, S., Joseph, T., Schreiber, K. L., Volk, T., Urman, R. D., Kovacheva, V. P., Wu, C. L., Mariano, E. R., & Ip, V. H. Y. (2025). Recommendations for disclosure of artificial intelligence in scientific writing and publishing: A regional anesthesia and pain medicine modified Delphi study. Regional Anesthesia and Pain Medicine. https://doi.org/10.1136/rapm-2025-106852
Flitcroft, M. A., Sheriff, S. A., Wolfrath, N., Maddula, R., McConnell, L., Xing, Y., Haines, K. L., Wong, S. L., & Kothari, A. N. (2024). Performance of artificial intelligence content detectors using human and artificial intelligence-generated scientific writing. Annals of Surgical Oncology, 31(10), 6387–6393. https://doi.org/10.1245/s10434-024-15549-6
Hwang, S. I., Lim, J. S., Lee, R. W., Matsui, Y., Iguchi, T., Hiraki, T., & Ahn, H. (2023). Is ChatGPT a fire of Prometheus for non-native English-speaking researchers in academic writing? Korean Journal of Radiology, 24(10), 952–959. https://doi.org/10.3348/kjr.2023.0773
Jembu, J. P., & Balang, R. V. (2025). Artificial intelligence adoption in nursing students’ academic writing: A qualitative study. Teaching and Learning in Nursing, 20(4), e1012–e1020. https://doi.org/10.1016/j.teln.2025.04.015
Johnson, J. E., & Rubo, K. (2025). Academic writing in nursing graduate programs: The use of artificial intelligence. Journal of Nursing Education, 64(10), 613–619. https://doi.org/10.3928/01484834-20250516-03
Khera, R., Pedroso, A. F., Keloth, V. K., Xu, H., Silva, G. S., & Schwamm, L. H. (2025). Scientific writing in the era of large language models: A computational analysis of AI- versus human-created content. Stroke, 56(10), 3078–3083. https://doi.org/10.1161/STROKEAHA.125.051913
Matsubara, S. (2025). Artificial intelligence in medical writing: Addressing untouched threats. JMA Journal, 8(1), 273–275. https://doi.org/10.31662/jmaj.2024-0268
Miao, J., Thongprayoon, C., Suppadungsuk, S., Garcia Valencia, O. A., Qureshi, F., & Cheungpasitporn, W. (2024). Ethical dilemmas in using AI for academic writing and an example framework for peer review in nephrology academia: A narrative review. Clinics and Practice, 14(1), 89–105. https://doi.org/10.3390/clinpract14010008
Pinto, D. S., Noronha, S. M., Saigal, G., & Quencer, R. M. (2024). Comparison of an AI-generated case report with a human-written case report: Practical considerations for AI-assisted medical writing. Cureus, 16(5), e60461. https://doi.org/10.7759/cureus.60461
Popkov, A. A., & Barrett, T. S. (2025). AI vs academia: Experimental study on AI text detectors’ accuracy in behavioral health academic writing. Accountability in Research: Ethics, Integrity and Policy, 32(7), 1072–1088. https://doi.org/10.1080/08989621.2024.2331757
Ramoni, D., Sgura, C., Liberale, L., Montecucco, F., Ioannidis, J. P. A., & Carbone, F. (2024). Artificial intelligence in scientific medical writing: Legitimate and deceptive uses and ethical concerns. European Journal of Internal Medicine, 127, 31–35. https://doi.org/10.1016/j.ejim.2024.07.012
Safrai, M., & Orwig, K. E. (2024). Utilizing artificial intelligence in academic writing: An in-depth evaluation of a scientific review on fertility preservation written by ChatGPT-4. Journal of Assisted Reproduction and Genetics, 41(7), 1871–1880. https://doi.org/10.1007/s10815-024-03089-7
Tang, A., Li, K.-K., Kwok, K. O., Cao, L., Luong, S., & Tam, W. (2024). The importance of transparency: Declaring the use of generative artificial intelligence (AI) in academic writing. Journal of Nursing Scholarship, 56(2), 314–318. https://doi.org/10.1111/jnu.12938
Thaichana, P., Oo, M. Z., Thorup, G. L., Chansakaow, C., Arworn, S., & Rerkasem, K. (2025). Integrating artificial intelligence in medical writing: Balancing technological innovation and human expertise, with practical applications in lower extremity wounds care. International Journal of Lower Extremity Wounds. https://doi.org/10.1177/15347346241312814
Weidmann, A. E. (2024). Artificial intelligence in academic writing and clinical pharmacy education: Consequences and opportunities. International Journal of Clinical Pharmacy, 46(3), 751–754. https://doi.org/10.1007/s11096-024-01705-1
Yang, J. J., & Hwang, S.-H. (2025). Transforming hematological research documentation with large language models: An approach to scientific writing and data analysis. Blood Research, 60(1), Article 00062-w. https://doi.org/10.1007/s44313-025-00062-w
Yoo, J.-H. (2025). Defining the boundaries of AI use in scientific writing: A comparative review of editorial policies. Journal of Korean Medical Science, 40(23), Article e187. https://doi.org/10.3346/jkms.2025.40.e187