Authors, reviewers, editors… smitten and bitten by the artificial intelligence (AI) bug!

Generative artificial intelligence (AI) and large language models (LLMs) have transformed almost every sector in real life including healthcare, scientific writing, and publications. The universal popularity of LLMs among academic researchers is clearly evident. Wiley ran a survey in 2025 which found that the use of AI for research had jumped from 57% in 2024 to 84% in 2025.[1] Another group of researchers ran a vocabulary analysis on 15.1 million biomedical PubMed abstracts published between 2010 and 2024 and found that at least 13.5% of them reflected the distinctive style of LLM-assisted processing, with usage going as high as 40% in some disciplines.[2] To write AI in medical writing off as a passing gimmick would be akin to Dr. Robert Metcalfe, the inventor of Ethernet, predicting the demise of the Internet in 1995! So, it is a Hobson’s choice, when this editorial chooses its stance, ‘generative AI and LLMs form a part of the core infrastructure of academic writing today.’

AUTHORSHIP IN ARTIFICIAL INTELLIGENCE TIMES

AI models assist researchers in otherwise tedious tasks such as extensive literature search and analysis, language translation, grammar correction, and even content generation. The collection and organisation of sources for research was a task that took days if not weeks in the past and now gets done quickly with the right prompt and the click of a button. Authors, who are non-native English speakers, are often faced with the challenge of fluency in English when writing for journals which use the English language. Poor language and grammatical errors are very often the cause of rejection of articles in Indian Journal of Anaesthesia (IJA) as mentioned in previous editorials of the IJA. However, nowadays, authors take help of AI tools such as ChatGPT, Google Gemini, Claude, and Perplexity, and the manuscripts are better presented, thus reducing the agony and work load of the editors and reviewers.

Nonetheless, it is important to point out at this juncture that since LLMs train on historical text patterns, AI-generated text has a tendency to sound fluent but often uses repetitive linguistic patterns, as if ‘making its claim to authorship’ in the manuscript.[3] It is indeed easy to be lulled by a sense of complacence by looking at the fluently worded and grammatically correct output. These models have information about medical literature but no understanding of disease biology and hence no clinical context or judgement to compare what they generate with. As a result, they do not possess the capacity to know when their output is incorrect.

A study of ChatGPT-generated literature reviews found that 18% of GPT-4 citations and 55% of GPT-3.5 citations were entirely fabricated, invented papers in real journals that simply do not exist.[4] A separate evaluation of ChatGPT answering medical questions found that the fabrication rate for references climbed to 69%.[5] When AI fabricates data or sources it from retracted articles, it raises questions regarding our scientific rigour too. Potentially such factitious data could become the source and platform for future studies, introducing the possibility of introduction and further perpetuation of flawed data into clinical practice.[4] This leads us to an important question that as authors, if the LLM chooses to use blogs or advertisements or even completely fake references as sources, and we choose to include them in our manuscript, would we in fact be publishing medical fiction?

If and when an author resorts to plagiarism or falsification of data, it is possible to hold them accountable and award punitive actions, but when LLMs become authors, this accountability is lost. The International Committee of Medical Journal Editors (ICMJE) also categorically states that LLMs cannot be held accountable for the accuracy, integrity, or originality of the work published and hence cannot be considered as authors.[6] The IJA saw this coming early. A 2023 commentary in this journal on ChatGPT in scientific publication warned that as language models spread through the research community, trust in science would hinge on authors being transparent about how the tools were used, and noted that journals were already moving to clarify the role of AI in authorship.[7] The IJA has also taken steps to ensure that the warning does not remain theoretical. IJA now requires disclosure of AI assistance, filed alongside trial registration, ORCID, and data-sharing statements, as part of a wider push towards Medline indexing standards.[8] Further, the journal will keep in tune with the changing scenes in the fast-progressing AI world and AI-related policies of the journal will be updated from time to time. AI-related misconduct including misuse by authors and reviewers will be curbed in time to protect research integrity. AI detection tools will soon be integrated into the online manuscript management system, keeping in mind that the current times are not of just ‘plagiarism check’, but of ‘AI check’ as well.

WHEN REVIEWERS AND EDITORS GET SMITTEN BY ARTIFICIAL INTELLIGENCE

For reviewers and editors too, the temptation is not unwarranted. Just as Eve was tempted to eat the forbidden fruit, peer reviewers and editors are increasingly taking the help of AI for generating their review comments. Nevertheless, the painstaking vigilance and attention to detail needed to ensure that the right manuscripts get published is made easy by the LLMs’ ability to meticulously comb through large volumes of data to find errors. It is indeed inevitable that humans, especially busy reviewers and editors, will assign these repetitive and arduous tasks to AI tools to increase efficiency, improve the quality of their review comments, and make them appear impressive. A study published in Nature in 2025 flagged 21% of the reviews as being fully AI-generated.[9] If reviewers use LLMs to generate reviews, it means that sensitive patient data including but not limited to clinical details and unpublished manuscripts are getting uploaded onto commercial freely available LLM models, which not only breaches patient confidentiality and Health Insurance Portability and Accountability Act (HIPAA) norms but also provides data that can be harvested for LLM training.

When AI-padded manuscripts reach the editors’ desks, the problem unfortunately worsens rather than getting better. The new age editors and reviewers are tasked with ensuring not only that the authors have not plagiarised content from other humans but also that the manuscript is written only by humans. The situation is grim and cannot be taken lightly. In a large study evaluating authenticity of articles, more than 32,000 were surmised to be ‘paper mill’ articles where data, references, and even authors were found to be non-existent.[10] In the year 2023 alone, a publishing house has had to retract more than 8000 such articles.[1] It is distressing indeed that such an alarmingly large volume of papers escaped the vigilance of reviewers and editors. Torielli et al.[9] explored this very question. They demonstrated how the architecture of the LLMs is vulnerable in being unable to distinguish system instructions, inputs from the users, and data retrieved from uploaded manuscripts, leading to authors being able to pad the feedback in their own manuscript reviews. They published a companion piece alongside, which exposed how commonly used AI detection tools are failing to consistently catch AI-written review reports, thus revealing a bigger problem than our earlier estimates.[11] The system can reverse too as sometimes, the authors may receive AI-generated peer review comments. This is an embarrassing situation for all: the author and editorial team. The editorial boards across the world today often find themselves exposed to this particular source of embarrassment, in being able to recognise the cadence and input of a chatbot from a peer reviewer and realising that the watchdogs are the ones that need to be watched. Nothing is more agonising for an editor than to be faced with an author who questions the efficiency of the editorial team after receiving AI-generated peer review comments. Evidently, editing in this AI-driven world is not an easy task, as was expressed by our editors in a previous editorial of IJA.[12]

NAVIGATING THIS BRAVE NEW WORLD

As the machines get better and better at generating, reviewing, and editing plausible and accurate manuscripts, it is but natural that our reliance on them will only increase in the near future. The true threat to human authorship is near as we get closer to witnessing ‘writing singularity’, transcending which machine-generated text will no longer be distinguishable from human efforts by pattern alone.[13] As men and women of science, we understand that the manuscript is not just a mouthpiece for our process. It represents our cognitive process, the recognition of our shortcomings, and contradictions. When the LLMs take over this process entirely, we risk deskilling future generations of researchers who may become experts in generating prompts to ensure a flawless manuscript output but may lose their hard-earned ability to critically appraise evidence. The universal use of the imitative, repetitive catchphrases would erode into the bold and diverse styles employed by various authors that often leave lasting impressions on the readers.

That being said, the way forward cannot possibly be to enforce an outright ban on the use of AI in academic writing. Seeking a strategic balance by ensuring accountability is the more viable option. Sound structural governance that is based in ethical practices is mandatory to ensure that AI can be harnessed as the powerful tool that it is, without over-extending its reach. Assessment of the ‘ethical readiness and trust in artificial intelligence among anaesthesiologists’ may be the perfect way to assess baseline readiness which has been justifiably assessed by Jamir T et al. in a survey study published in this issue of IJA. In a cross-sectional survey conducted among anaesthesiologists, the authors found moderate scores across the population with factors such as ethical readiness and transparency and control as independent predictive factors.[14] Though their small sample size of participants may not be representative of the entire Indian anaesthesiology community, it is a start and will definitely spark more research along these lines.

Where the ‘no authorship to the non-human authors’ principle remains non-negotiable as of now, it is essential to recognise that AI disclosures need to be detailed and truthful. Most editorial policies including the most recent one from the Journal of the American Medical Association (JAMA) agree that the use of AI for grammar and spelling correction is the only use that may be accepted without disclosures and represents a fairly low-risk use of AI.[15] Its use for literature search, manuscript preparation, formatting data, and graphical representation may be acceptable, provided the authors provide a complete disclosure and take ownership of any errors that may arise.

For reviewers, the policy is unwavering; no AI use is permitted for reviewing as of now. Firstly, current AI tools can assess the fluency of the prose, but without the clinical context, its judgement about clinical relevance, design validity, and ethical integrity would remain questionable. Although AI writing tools are meant to help non-native English speakers write better, the currently available AI detection software shows an unfairly high rate of false positive detection (61.3%) among the very same authors increasing this very bias.[1] Secondly, reviewers should not upload confidential manuscripts or unpublished data to external generative AI systems or use such systems to generate substantive peer-review assessments unless explicitly permitted by the journal and appropriate safeguards are in place.

The use of AI-powered tools as co-editors is where LLMs show great promise. Using secure LLMs could assist with editorial tasks like reference formatting, identification of duplicate text, consistency checks between tables and text, and preliminary technical screening, while final scientific and editorial decisions remain with human editors. The catch, however, would be to ensure that the human control be present at all times and that these secure LLMs be open-source editorial tools to equitable availability even in low- and middle-income countries and smaller journal societies.[1]

Like we said when we began, the use of AI in medical writing is now an inevitable reality. It would be unwise to assume that we have control over its infiltration into medical writing. To put it bluntly, our strongest bet to safeguard creativity, scientific rigour, and patient interests is to have AI act as a co-pilot, by assisting in academic publishing, but have governance in place, to ensure that human authors, reviewers, and editors must remain responsible for the accuracy, integrity, and confidentiality of the work.

Author contributions

MK: Concept, drafting of the manuscript, manuscript review. GRG, SS: Literature search, manuscript writing and revision. All authors contributed to critical reviews and final approval of the manuscript.

Disclosure of use of artificial intelligence (AI)-assistive or generative tools

AI tools were not used.

Declaration of use of permitted tools

No permitted tools were used.