We conducted a cross-sectional analysis of the manuscript submission guidelines for the top 100 journals, as ranked by the 2023 SCImago Journal Rankings. Data were collected from the Instructions for Authors webpages of these 100 journals. This study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.

The selection of eligible journals was carried out through consultation between the two investigators (PC, AVT) and the medical research librarian. The 2023 SCImago Journal Rankings were used to gather 396 journal listings for screening. SCImago is an online platform that ranks journals using bibliometric indicators, including the h-index and the SCImago Journal Rank Indicator (SJR). The SJR is considered a more robust metric than the traditional impact factor, as it is open-access and evaluates both the quantity and quality of a publication’s citations. The rankings are generated annually using Elsevier’s subscription-based Scopus database, which encompasses a vast collection of scientific journals, providing a comprehensive dataset for this study. All peer-reviewed journals categorized under subject area ‘medicine’ and subject category ‘neurology’ of the 2023 SCImago journal listings were included in this study.

Journals were excluded from our study if they met any of the following criteria: (i) were discontinued, (ii) lacked editorial office contact information on their website, as we sought to minimize bias by providing editors an opportunity to clarify their publication policies, or (iii) were published in a language other than English without offering a translation option.

Data was independently extracted by two investigators (KK, NC) from the Instructions for Authors pages of included journals in a masked, duplicate manner. The data were recorded using a standardized Google Form, which was pilot-tested and designed in advance by investigators PC and AVT. During this process, we carefully reviewed publishing policies, authorship criteria, and any editorials or updates from journals and publishing companies related to the use of AI, chatbots, and large language models (LLMs). Data extraction occurred between November 1, 2024 and July 31, 2025. Once the extraction was completed, investigators KK and NC reconciled their data, with a third investigator (PC) available to resolve any discrepancies. All journals were re-reviewed prior to final data reconciliation to ensure that policy information was accurate as of July 31, 2025.

For journals without a statement on AI usage in their Instructions for Authors webpage, a standardized email was sent to the Editor-in-Chief or a member of the editorial office to inquire about the development of policies regarding AI use in their publication process. To enhance response rates, emails were sent weekly for three consecutive weeks. All responses, including non-responses, were documented.

The primary outcome of this study evaluated AI usage statements in journal guidelines. The secondary outcome involved reviewing and summarizing the number of journals that permit or prohibit the use of AI-generated content, images, writing, and the inclusion of AI authorship.

Data summaries of descriptive statistics for journal AI policies were created using R (version 4.2.1) and RStudio. These descriptive statistics consisted of statements regarding: (1) AI generated images (2) AI generated content (3) AI authorship inclusion (4) AI assisted manuscript writing. Bias analysis was not necessary, as the data collected involved a direct evaluation of the Instructions for Authors as opposed to an assessment of each individual study.

In addition, R (version 4.2.1) and RStudio were used to conduct correlational analyses. The strength of correlation was analyzed between the AI usage policies in research and SJR score, journal rank, impact factor, and publishing country. Point-biserial correlations were used because AI policy variables were dichotomous (e.g., permitted vs. not permitted), whereas journal metrics were continuous, making point-biserial correlation an appropriate effect-size measure equivalent to Pearson correlation in this setting. Given the exploratory nature of these analyses, we considered multiple-comparison adjustment but did not apply formal correction; exact p-values are reported for transparency.

To ensure transparency and reproducibility, analysis scripts, standardized emails, extraction forms, and all raw data collected are publicly available on Open Science Framework (OSF) (21). Generative AI (ChatGPT, OpenAI; GPT-5.1) was used solely to assist with grammar and wording refinement during manuscript preparation; it was not used for data extraction, analysis, interpretation, or content generation, and all edits were reviewed for accuracy by the authors.

Most journals (97/100, 97.0%) referenced artificial intelligence within their Instructions for Authors (Table 2). Fewer endorsed established editorial frameworks: 14.0% (14/100) cited International Committee of Medical Journal Editors (ICMJE) recommendations, 26.0% (26/100) cited Committee on Publication Ethics (COPE) guidance, and 10.0% (10/100) referenced World Association of Medical Editors (WAME). Only one journal (1/100, 1.0%) endorsed an AI-specific reporting guideline. The three journals without an AI statement were contacted via editorial outreach; no responses were received and no policy classifications changed.

AI authorship was universally prohibited among journals with clear policies, with 97.0% (97/100) explicitly disallowing AI tools as authors (Table 3). Disclosure of AI use during manuscript submission was required by 96.0% (96/100) of journals. AI-assisted writing was widely permitted (93/100, 93.0%). Policies regarding AI-generated outputs were more variable: 77.0% (77/100) allowed AI-generated content, whereas 37.0% (37/100) allowed AI-generated images.

Among journals that explicitly referenced AI in their guidelines (n = 97), all prohibited AI authorship (97/97, 100%) and nearly all required disclosure of AI use (96/97, 99.0%; Table 4). AI-assisted writing was permitted by 95.9% (93/97). Policies on AI-generated content (77/97, 79.4%) and AI-generated images (37/97, 38.1%) remained heterogeneous.

No significant associations were observed between AI-related policies and journal characteristics, including Impact Factor, SCImago Rank, or publisher. Prespecified correlation analyses are reported in Supplementary Additional File 1.

The majority of neurology journals in our study prohibited AI authorship and required disclosure of AI use, aligning with the recommendations of the ICMJE (23). However, notable heterogeneity emerged regarding which forms of AI assistance were permissible. While nearly all journals (95.9%) allowed AI for language refinement or writing assistance, only 79.4% permitted AI in content generation and less than half (38.1%) endorsed AI-generated images. This variability reflects an ongoing tension between encouraging technological innovation and maintaining the integrity of scientific communication (24). Such inconsistencies are not unique to neurology. In ophthalmology, for example, Almobayed et al. reported that only two-thirds of journals had implemented AI policies, a finding that mirrors the uneven adoption we observed (25). Similar discrepancies across disciplines indicate that policy development is lagging behind technological progress, creating uncertainty for authors and editors alike (26).

Despite the ubiquity of AI in contemporary research, adherence to AI-specific reporting guidelines was almost nonexistent in our sample, only one journal referenced the Consolidated Standards of Reporting Trials–Artificial Intelligence (CONSORT-AI) and Standard Protocol Items: Recommendations for Interventional Trials–Artificial Intelligence (SPIRIT-AI) extensions (27, 28). These guidelines were designed to enhance transparency, reproducibility, and interpretability in studies that develop or evaluate AI interventions (29). Their absence in neurology journals raises concerns about the reporting quality of AI-driven studies. Even when broader editorial and ethical frameworks were considered, adherence remained limited. Only 14% of journals referenced ICMJE, 26% cited COPE, and 10% mentioned WAME in regards to their AI policies (23, 30, 31). These figures indicate that, while some recognition of established ethical guidance exists, consistent enforcement across neurology journals is lacking. Each of these organizations has recently begun addressing AI’s role in scientific writing, uniformly prohibiting AI authorship, mandating disclosure of AI assistance, and affirming that human authors bear full responsibility for verifying the accuracy and integrity of all AI-generated content (32). Integrating these limited ethical frameworks with robust, AI-specific reporting guidelines would provide a more cohesive and enforceable foundation for responsible AI use across both research conduct and manuscript preparation (22).

While the current lack of standardization presents challenges, AI also offers substantial benefits to neurological science. AI-assisted data analysis can accelerate hypothesis generation, detect subtle imaging biomarkers, and uncover complex patterns in high-dimensional datasets that exceed human analytic capacity (33–36). Natural language processing tools have demonstrated efficiency in screening abstracts and extracting data for systematic reviews, thereby reducing researcher workload and enhancing reproducibility (37, 38). However, these advantages are contingent upon transparent documentation of AI involvement. Unregulated or undisclosed use risks propagating bias or producing misleading conclusions, which could ultimately compromise clinical translation (39). Therefore, enforcing standardized policies is not merely procedural, it is essential for maintaining scientific rigor as neurology increasingly embraces computational methodologies.

This study possesses several strengths, especially in its design. To minimize bias and improve accuracy, data extraction was performed in a masked, duplicate manner with two trained authors. Additionally, we adhered to a pre-established protocol, ensuring a structured and systematic approach. While this study has many strengths, it is not without limitations. Human error in data extraction and analysis remains a possibility; however, we aimed to minimize this risk through reconciliation, discussion, and the involvement of a third party to resolve any discrepancies. Additionally, some neurology journals lacked standardized Instructions for Authors, which may have led to missed information if AI policies were only accessible through external links. To address this, we contacted editorial teams to verify the existence of an AI policy when it was initially coded as absent.

The ethical and methodological stakes of AI integration are particularly high in neurology, where diagnostic accuracy and interpretive precision are paramount (40). To meet these demands, AI systems require sufficient transparency to allow clinicians and reviewers to evaluate how predictions are derived and whether they align with neurobiological principles. Standardized reporting guidelines, such as CONSORT-AI and SPIRIT-AI, could be implemented to help mitigate these risks by enforcing transparency in dataset composition, model validation, and interpretability (27, 28). Furthermore, providing clear guidance on the permissible use of generative AI in writing and figure creation would ensure that human oversight remains central to scientific authorship (41). As AI continues to shape the landscape of neurological research, journals must act proactively to establish enforceable, transparent, and ethically sound frameworks.