Biomedical research publication in the age of artificial intelligence: Current prospects for balancing integrity and innovation

Considerate use of AI tools for authors

Authors are permitted to use AI tools for limited, well-defined functions that do not influence scientific conclusions or compromise the originality of their work. These include grammar and spelling corrections, copy-editing, stylistic refinement, translation of non English text, English enhancement, and assistance in reference formatting. Such uses are considered acceptable provided that authors verify all AI-generated suggestions before submission. The authors remain fully accountable for the accuracy and validity of the final content, including factual data, analyses, and interpretations. Authors can use offline or enterprise-grade licensed software that guarantees data privacy and prohibits data reuse for algorithm training. ^[4,14-18]

Journals explicitly prohibit the use of AI systems for generating primary data, performing analyses, interpreting results, or drafting substantial portions of scientific arguments without human oversight. Authors must not upload confidential or unpublished manuscripts to public AI interfaces that store or reuse inputs for training purposes. Any attempt to employ AI for authorship attribution or content fabrication constitutes serious academic misconduct. Authors are reminded that any inclusion of patient data or sensitive institutional material in AI systems without anonymization and prior approval may constitute a breach of legal confidentiality obligations. ^[4,14,16-18]

Generative AI systems (like ChatGPT, SciSpace, Thesis AI, Paperpal, Wordvice, Wordtune etc.) can not only produce extensive text, figures, or datasets, but also write research paper, review articles, and theses, with remarkable fluency and accuracy. While this ability enhances writing efficiency, it also introduces serious risks to research integrity and scientific concerns over authorship, accountability and bias. AI can generate text that mimics authentic scientific writing, but it is often based on incomplete, incorrect, or entirely fabricated information. Because LLMs remix existing content from their training data, they may inadvertently reproduce others’ ideas or language without attribution, resulting in unintentional plagiarism. Furthermore, when AI-generated outputs are presented as genuine experimental results or interpretations, it constitutes a misrepresentation of scientific work. ^[2,16,19,22]

“Paper mills” are illicit commercial entities that sell fabricated or manipulated research papers to authors seeking publication. With generative AI, these operations can now produce fake manuscripts at scale, using AI to write convincing text, generate false data, and even create images or tables that appear authentic. Such fraudulent submissions flood journals and threaten the credibility of peer-reviewed literature. Furthermore, advanced AI algorithms can create synthetic images, audio, or videos, known as deepfakes, that imitate real experimental data, patient scans, or even researchers themselves in scientific contexts. Deepfakes may falsify clinical images, microscopy data, or interview recordings, leading to false evidence that can mislead readers, distort scientific conclusions, and endanger patient safety when used in clinical research. ^[2,22,23] Therefore, maintaining the sanctity of academic authorship and ensuring that technological assistance complements, rather than compromises, scholarly rigoris essential. AI-based editing tools can enhance readability, reduce linguistic barriers for non-native English authors, and streamline administrative workflows. Nevertheless, without clear ethical boundaries, these same tools can distort authorship contributions, violate confidentiality agreements, and erode trust in the scientific literature.

Limitations of AI

AI tools are trained on vast datasets created by humans, which means they can inherit the same cultural or linguistic biases, such as those related to gender, race, or inclusivity, present in the source material. These biases may appear overtly or subtly, for instance, by favoring research in the majority language or mainstream academic perspectives.^[2]

Although LLMs can simulate reasoning, they lack the human intuition, emotional intelligence, and contextual awareness of human thought, often missing subtle cues such as humor, irony, or metaphor. This limitation may restrict their capacity to contribute creatively or critically to complex scientific discussions, which can influence clinical decision-making because of limited possibilities for higher-level thinking.^[2,24]

AI models are also prone to generating inaccuracies or hallucinations, producing authoritative-sounding outputs, statements, or references that appear plausible but are entirely false. Experiences with tools such as ChatGPT have shown that it may fabricate citations or alter publication dates, which can mislead researchers, raise ethical or legal concerns, and ultimately erode public confidence in scientific communication.^[25] Because LLMs are designed to predict text patterns, they unintentionally reproduce phrases or ideas from existing literature without acknowledging the creator of the intellectual idea, resulting in inadvertent plagiarism or duplication of scholarly content.^[26] The datasets used to train AI systems are derived from a mix of reliable and unreliable internet sources, which means that some generated information may be misleading, particularly in sensitive fields such as healthcare or clinical research, where errors could impact patient outcomes. Predictive algorithms are trained to identify recurring patterns, which can bias them towards conventional or widely accepted ideas while overlooking novel hypotheses. They may also rely on mainstream ideas that are based on incomplete or outdated data, thereby limiting innovation and perpetuating past biases.^[2] Moreover, many AI tools function as “black-boxes”, with their datasets, architectures, and training processes kept proprietary by developers; this opacity hinders academic scrutiny and prevents users from evaluating potential bias or data quality, sometimes leading to unreliable or context-inappropriate outputs.^[27]

Although AI can make publishing more inclusive in theory, without equitable access and localized training initiatives, it risks amplifying the very inequalities it aims to solve, by creating a digital divide within the global research ecosystem due to unequal access to training and infrastructure. ^[28-30]

Another critical limitation of frequently using AI in scientific writing is that convenient AI-driven content generation can diminish researchers’ engagement with primary and original literature, thereby weakening their skills in reviewing literature, verifying sources, and interpreting context. Core academic practices, such as lateral and vertical reading, which involve critically cross-checking information across diverse sources, are increasingly being replaced by superficial text synthesis. This shift ultimately affects essential cognitive abilities, including argument analysis, critical thinking, creativity, and scientific integrity.^[31,32] Furthermore, the uncritical reliance on AI-generated text or data has contributed to the rise of low-quality biomedical research papers, characterized by repetitive phrasing, factual inaccuracies, and poorly substantiated claims.^[33,34] This not only dilutes the credibility of scientific literature but also poses a challenge to editors and reviewers to distinguish genuine data from algorithmic output. To protect the intellectual rigor of biomedical science, researchers must therefore use AI only as an augmentative tool, rather than a surrogate for analytical reading, conceptual understanding, and scholarly authorship.

Responsible use and transparency

A systematic review conducted by Khalifa & Albadawy^[35] across PubMed, Embase, and Google Scholar, selecting 24 studies published after 2019 that directly examined AI’s role in academic writing, editing, research design, and publishing, suggested six core domains in which AI can assist in manuscript writing. AI helps in brainstorming, identifying research gaps, and generating hypotheses by analyzingexisting data and literature to inform these processes. It can propose suitable methodologies, sample sizes, and statistical models, supporting more structured research planning. However, over-reliance on AI may divert research objectives or reduce human creativity.^[35-37] AI tools, such as ChatGPT, improve writing fluency, expand text, predict terminology, and ensure logical flow through structuring and tone analysis. They also enhance visual and multimedia presentations via auto-generated figures, infographics, and presentations. However, ethical vigilance is needed to prevent misuse, such as fabricating or falsifying content. ^[38,39] The AI revolutionized literature reviews by extracting, summarizing, and synthesizing massive volumes of data through semantic analysis (e.g.^, www.scispace.com). It creates summary tables and comparative insights, offering more comprehensive overviews while saving time. However, researchers must validate accuracy and context to avoid misleading interpretations.^[40] AI enhances data interpretation, visualization, and duration. Tools like Tableau, NVivo, and RapidMiner can analyze large datasets and generate accessible insights. However, maintaining data integrity and ethical standards during automation is of utmost importance.^[41] AI tools, such as Grammarly, Paperpal, and ChatGPT, refine manuscripts by proofreading, summarizing abstracts, and assisting in drafting peer-review responses. They streamline submission tracking and correspondence, improving workflow efficiency while requiring ethical disclosure of AI use in manuscripts.^[42] AI aids in dissemination and accessibility, including multilingual translation, social media outreach, and chatbot-based engagement, while also ensuring ethical compliance through plagiarism detection, integrity checks, and adherence to research ethics frameworks. However, transparency and disclosure of AI usage are strongly recommended.^[43]

Global organizations, such as the COPE, WAME, and ICMJE, are actively developing guidelines to promote transparency and accountability in the responsible use of AI within research and publishing. LLMs and other AI tools do not qualify for authorship and must not be listed as authors. When used for writing, language editing, or data handling, their role should be explicitly declared in the acknowledgements section, stating the name, version, and nature of use.^[4,14,17,18] If AI tools assist in data analysis or figure generation, details of the model, prompt, and date of use should be clearly mentioned in the methods section. Researchers remain fully accountable for all AI-generated content. They must keep records of the system used, including the date, verify the accuracy of references, validate concepts and interpretations, ensure that outputs and language are unbiased and correctly attributed, and ensure that independent human oversight is essential to safeguard the reliability and integrity of scientific communication.^[2,19]

Current consensus and policy on disclosure of AI use

Transparency is central to ethical scholarly communication. The author must declare any AI involvement in manuscript preparation in either the acknowledgementor method section. A suitable statement example of disclosure is: the author used [Tool Name, Version] for grammar and language refinement. No AI tool was used for data generation or interpretation. Failure to disclose the use of AI within the editorial workflow adheres to privacy and ethical standards. Disclosure aligns with fundamental norms of scientific integrity, including honesty, transparency, accountability, reproducibility, and fairness.^[44] Resnik and Hosseini^[45] enumerated five ethical rationales for disclosing the use of AI in research. First, it ensures fair credit assignment, so that human authors are not improperly credited for work produced or significantly shaped by AI systems. Transparency about AI use also enhances accountability, allowing any issues, such as plagiarism, data fabrication, or algorithmic bias, to be traced back to the responsible individuals. Disclosure supports reproducibility by enabling other researchers to understand and replicate methods or analyses that involve AI tools. Furthermore, it establishes a clear boundary between human and machine contributions, clarifying the extent to which the work was guided by human judgment versus automated processes. Ultimately, openness about AI involvement reinforces trustworthiness, strengthening both public and academic confidence in the credibility and integrity of the scientific record.^[45] They further introduced a two-part criterion for determining whether disclosure is mandatory and suggested that AI use must be disclosed if it is intentional and substantial. Intentional use refers to the researcher's deliberate application of an AI tool for a specific scientific purpose. Substantial use means the tool directly affects the researcher's evidence, analysis, or conclusions. Three practical indicators of substantial AI use proposed by them are summarized in Table 2.^[45.46]

AI use in peer review and editorial processes

Peer reviewing is the cornerstone of scholarly publishing. It depends on several essential principles, including confidentiality, accountability, transparency, data minimization, fairness, and bias, as well as expert human judgment. In medicine and healthcare, where evidence-based accuracy determines patient outcomes, the shortage of qualified reviewers is acute. A small fraction of dedicated academics shoulders the majority of review tasks, leading to fatigue and variability in the quality of reviews.^[47] AI tools are now powerful enough to reduce administrative burden, automate quality checks, and support editors in identifying suitable reviewers. Further, it assists reviewers in organizing, verifying, and critically appraising manuscripts.^[48,49]AI can assist reviewers in rephrasing comments, leading to kinder and more professional reviews, thus helping clarify and improve the tone of feedback. AI has the potential to accelerate the peer review process by automating tasks such as pre-submission checks and managing large volumes of submissions, as well as quickly identifying issues related to plagiarism, statistical errors, or data inconsistencies. AI can help match manuscripts with established experts by analyzing keywords and prior publications, thereby streamlining the reviewer selection process.^[50]

However, AI-generated comments may be overly generic, lack specificity, and fail to engage with the manuscript's actual content. AI-generated data/reference outputs could produce inaccurate and fabricated results, and if passed to authors without proper verification, undermine the reviewer’s credibility. Without consent, unauthorizedAI use risks compromising confidentiality and intellectual property, which may further impact the quality of reviews when confidential manuscripts are uploaded to external servers. Algorithm-based AI decisions may bias research selection and blur the boundaries between human and AI, thereby reducing originality and credibility. Over-reliance on AI may erode human judgment, leading reviewers to skip critical analysis and provide less thoughtful and detailed evaluations.^[50,51]

Therefore, reviewers require training to use AI responsibly, recognize its limitations, and maintain ongoing professional development. They must understand that most AI tools operate through cloud-based or public servers, which inherently compromise confidentiality, and their unregulated use can compromise the ethical foundation of peer review. When a reviewer uploads even a portion of a manuscript to an online AI platform, it may be temporarily stored, used for model training, or expose sensitive data. Reviewers must adopt workflows that ensure data never leaves their secure environment while still harnessing AI’s cognitive strength. Practical workflow for ethical AI assisted reviewing without uploading substantial part of manuscript commenced with reading of full manuscript manually; followed by use of AI to search recent studies and summarize published literature, cross-check citations or verify contextual accuracy of references; followed by data/ method verification for statistical claims, detect p-value errors using offline AI calculators or statcheckers and thus AI can help explain complex concepts or refresh knowledge about statistics, methods or terminology mentioned in the paper without disclosing the manuscript text; then write/ draft original review and use AI to refine the grammar and tone or summaries own notes to ensure professionalism and neutrality; and final check ensuring originality to remove any AI artifacts while taking full accountability [Table 3]. In any case, reviewers and editors are discouraged from submitting any manuscript text, figure, or table to third-party AI systems without explicit consent from the corresponding author or publisher. Human oversight must remain central, with AI used to complement, not replace, the judgment of reviewers or editors.^[50] In the end, if AI tools are used in any part of the review preparation, phrasing or literature clarification, reviewers should disclose this to the editor. A disclosure statement by reviewers affirming that “AI tools were used for linguistic or conceptual assistance, not for content generation or manuscript analysis, and that confidentiality was maintained throughout” can be added for transparency and accountability.

AI misconduct detection and publication ethics

Adele da Veiga ^[52] analyzed the AI-related publication unified policies of the world’s top 10 academic publishers to propose ethical standards for the use of AI in research and publishing. Using a literature-based thematic analysis, the author identified significant inconsistencies among publishers regarding the acceptable use of AI. Therefore, a coherent, transparent, and globally harmonized set of ethical guidelines is needed to distinguish the responsible use of AI-assisted tools from the more cautious use of generative AI in scholarly publishing.^[52] Research integrity communities, journals, and publishers have been utilizing numerous AI tools to detect research misconduct, including plagiarism, image manipulation, and paper mill activities. While AI tools like ImaCheck, FigCheck, Papermil Alarm, Proofig, and ImageTwin can enhance research integrity, they also introduce ethical, procedural, and fairness challenges.^[46,53,54]

AI screening systems are not error-free and can generate false positives, wrongly flagging legitimate work, which may lead to unsubstantiated or harmful allegations against researchers who have been vindicated. This is because an AI tool cannot determine intent, which is essential in distinguishing misconduct from honest error under established definitions by US federal policy and COPE standards. Hossenini & Resnik^[46] highlighted that AI detection tools have technical limitations, such as the inability to recognize subtle image alterations or conceptual plagiarism. Furthermore, misuse of AI screening could introduce bias, especially if applied unevenly across submissions from different regions or institutions. They emphasize that any concerns identified by AI systems must undergo human verification before action is taken, ensuring that false positives do not unfairly harm researchers. Journals should rely only on validated and reliable AI tools, while staying alert to emerging forms of misconduct that may evade current detection methods. Authors flagged by these systems must be informed and allowed to respond or clarify before conclusions are drawn.^[46] Moreover, AI screening should be applied consistently across all submissions to avoid regional or institutional bias. Publishers should publicly disclose their policies on how AI tools are implemented and how potential cases of misconduct are addressed.