Introduction

Clinical studies published in journals of internal medicine rely largely on statistical analyses and quantitative inference, which makes the quality of statistical reporting and statistical peer review a key determinant of the credibility of this literature. Unfortunately, the quality of statistical reporting in medical journals remains low. Despite publication of methodological recommendations, the proportion of journals using specialist statistical peer review has not changed over the past 2 decades.¹ Difficulties in recruiting qualified reviewers, impractical approaches to teaching biostatistics, and failure to implement existing recommendations contribute to the fact that only a small proportion of accepted manuscripts contain statistical analyses that are assessed as correctly performed. Analyses of reporting practices have shown that even after the publication of editorial recommendations, no improvement in the quality of statistical reporting has been observed.² For example, evaluations of articles published in radiology journals indexed in the Science Citation Index showed that almost all analyzed papers contained at least 1 error, most often related to the summarization of data and reporting of P values. Importantly, the frequency of these errors was not associated with the journal impact factor, indicating that the problem also affects highly ranked specialist journals.³ A survey study among members of the World Association of Medical Editors showed that, according to 40% of the respondents, only 31%–50% of the manuscripts accepted for publication are statistically correct, and most respondents estimated the frequency of statistical peer review at 1%–10% of the submitted manuscripts.⁴ Taken together, these observations indicate that despite the availability of recommendations and increasing awareness of the problem, the quality of statistical analyses and reporting in the medical literature remains a substantial systemic challenge.

In parallel with the persistent problems related to the quality of statistical analyses in the medical literature, tools based on generative artificial intelligence (AI) are beginning to play an increasingly important role in clinical research and diagnostics. A systematic review and meta‑analysis published in 2025 in NPJ Digital Medicine showed that generative AI models achieve moderate diagnostic accuracy and, overall, do not differ significantly in effectiveness from physicians, particularly those without extensive clinical experience. At the same time, the significantly poorer performance of AI, as compared with clinical experts, indicates that the use of these tools in medicine and scientific research should be considered a support rather than a replacement for expert judgment, with due consideration of their limitations and potential consequences for research integrity.⁵ Similar conclusions arise from a systematic review and meta‑analysis of studies evaluating the performance of ChatGPT in medical licensing examinations, which demonstrated substantial variability in results depending on the model version, the type of questions, and the linguistic and examination context. Although the most recent versions of the model achieved high mean accuracy and, in many cases, obtained results comparable to or higher than those of medical students, the authors emphasize that instability of performance and susceptibility to contextual factors limit its safe application in formal medical education.⁶ An example of such limitations is provided by studies in internal medicine in which ChatGPT did not meet the passing criteria of the Polish specialization examination in internal medicine.⁷

The growing interest in the use of AI‑based tools in biomedical research further complicates the problem of the quality of statistical analyses and the credibility of scientific conclusions. Automation of method selection, data modeling, and result interpretation may lead to both improvements in analytical efficiency and the reinforcement or masking of methodological errors, particularly in the absence of appropriate expert oversight. The literature increasingly highlights that the use of AI and machine learning intensifies existing problems related to overinterpretation of statistical significance, limited reproducibility of results, and insufficient reporting of analytical assumptions. Consequently, there is a growing need for a critical assessment of the role of AI not only as an analytical tool but also as a factor influencing research integrity, reporting practices, and the scientific peer review process.⁸ Initiatives using AI tools, including large language models (LLMs), to screen scientific manuscripts are being reported with increasing frequency, accompanied by warnings from researchers regarding the risks associated with such approaches.⁹ At the same time, attention is drawn to the fact that the rapidly expanding use of generative AI in the creation and editing of medical publications may in the short term substantially change the way scientific literature is produced, raising serious concerns about transparency, authorship accountability, and the quality and integrity of the resulting content.¹⁰

The most recent article published in December 2025 in Journal of Korean Medical Science presents the role of AI in the detection of statistical errors as a component supporting research integrity.¹¹ According to the abstract, the main emphasis was placed on tools such as Statcheck and the GRIM test, which demonstrate moderate effectiveness in identifying statistical errors and may accelerate the peer review process, although they require high‑quality data and continuous human oversight. The authors also point to potential benefits and limitations of using AI, including the risk of incorrect interpretation of results, algorithmic bias, ethical concerns, and the need to preserve the decisive role of editors and reviewers. At the same time, as indicated in the abstract, the article focuses primarily on specialized tools for detecting statistical errors and does not include a review of studies evaluating the use of generative language models and other AI tools in conducting statistical analyses and statistical peer review. The absence of such a synthetic overview constituted one of the premises for writing this review.

The aim of this narrative review was to comprehensively summarize the available evidence regarding the use of AI tools in statistical analysis and statistical peer review in medical research, to discuss the limitations of these approaches, and to assess the implications of their use for the integrity of medical research and the scientific peer review process, with particular emphasis on clinical studies published in journals of internal medicine. Although the available evidence is cross‑sectional in nature and not limited to a single clinical specialty, it relates to universal analytical and peer review practices that are also present in studies published in journals of internal medicine. Therefore, the conclusions of this review are of direct relevance to authors, reviewers, and editors of internal medicine journals, in which the reliability of statistical analyses constitutes one of the key conditions for the credibility of clinical conclusions.

Methods

This narrative review encompasses literature published between 2023 and 2025. The temporal scope of the review was restricted to the years 2023–2025, as this period corresponds to the widespread public availability and rapid development of LLMs, enabling the use of natural‑language interfaces for statistical analysis, result interpretation, and methodological review. Earlier applications of AI in data analysis were predominantly algorithmic or computation‑automation tools and did not encompass generative capabilities or language‑based evaluation of statistical analyses. A comprehensive literature search was conducted in PubMed, Web of Science, and Scopus databases. The search strategy employed combinations of key words related to AI and statistical practice, including but not limited to: “artificial intelligence,” “generative AI,” “large language models,” “statistical analysis,” “statistical review,” “statistical reporting,” “statistical errors,” “methodological errors,” and “research integrity,” as well as the names of specific AI tools evaluated in the literature. Eligible studies were English‑language publications addressing the use of AI tools in statistical analysis, statistical review, data extraction, methodological assessment, or detection of statistical errors within medical research. The review included original research articles, validation studies, methodological studies, systematic reviews and systematic review–related studies, editorials, perspectives, and tutorials, reflecting the heterogeneous and rapidly evolving nature of this research area. No restrictions were imposed on study design, provided that the publication directly evaluated or discussed AI‑assisted statistical analysis or statistical review in a medical research context. Study selection was based on relevance to predefined thematic areas, including AI‑assisted data analysis, statistical test selection, statistical reporting and guideline adherence, statistical peer review, detection of methodological or statistical errors, and implications for research integrity.

For clarity of synthesis and to reflect the rapid evolution of AI tools, the included studies were grouped according to the year of publication (2023, 2024, and 2025). This temporal stratification was adopted to capture changes in the scope, maturity, and methodological focus of AI‑assisted statistical analysis and statistical review over time. Within each yearly stratum, evidence was synthesized narratively, emphasizing study objectives, types of AI tools evaluated, analytical tasks addressed, and key findings, rather than presenting isolated study summaries. This approach allowed for identification of emerging trends, shifts in application domains, and recurring limitations across successive stages of AI development.

Artificial intelligence–assisted statistical analysis and review: empirical evidence (2023–2025)

An overview of the evolution of AI‑assisted statistical analysis and statistical review in medical research between 2023 and 2025 is summarized in Table 1.

Implications of artificial intelligence for statistical analysis and statistical review in internal medicine

The literature reviewed above focuses on the use of AI in statistical analysis and statistical review within medical research. These methodological processes are common across clinical disciplines and become particularly consequential in areas of medicine, where clinical evidence is predominantly derived from heterogeneous observational studies and pragmatic clinical trials. Within this broader methodological context, internal medicine serves as a clinically relevant setting in which the implications of AI‑assisted statistical analysis and statistical review can be discussed, without implying that the reviewed evidence is specific to this field.

Key implications of AI‑assisted statistical analysis and statistical review for internal medicine are summarized in Table 2.

First, the reviewed evidence consistently indicates that generative AI tools should not be treated as fully autonomous instruments for statistical analysis in clinical research. Although recent agentic systems demonstrate increasing operational autonomy, their performance remains dependent on human‑defined objectives, prompt design, and expert oversight. Across studies published between 2023 and 2025, models such as ChatGPT demonstrated substantial variability in responses, incomplete consideration of statistical assumptions, and a tendency to recommend inappropriate analytical methods, particularly in more complex or nonstandard scenarios.^12,15,19,28 In light of these findings, the use of such tools in the context of research published in internal medicine journals raises specific concerns, as studies in this field commonly rely on heterogeneous observational datasets, multiple clinically‑driven end points, and analytical decisions that require careful consideration of confounding, model assumptions, and clinical plausibility. Under these conditions, reliance on AI‑generated recommendations without expert verification may compound, rather than mitigate, the risk of methodological error identified in the reviewed literature.

Second, the findings suggest that the most robust role of AI lies in augmenting, rather than replacing, statistical expertise. Studies comparing AI outputs with classical statistical software showed high concordance for simple analyses, while discrepancies emerged in advanced procedures requiring contextual judgment and methodological nuance.^18,20 Similarly, evidence from consultation‑based and educational settings indicates that AI can support routine tasks, such as code generation or preliminary method selection, but its effectiveness remains strongly dependent on the user’s statistical competence.^22,33 In the context of research published in internal medicine journals, where analyses frequently involve complex clinical variables, multiple interrelated outcomes, and nonrandomized study designs, these findings imply that AI‑assisted tools may facilitate preliminary or repetitive analytical tasks, while responsibility for methodological choices, model validation, and interpretation of results must remain with experienced statisticians and clinically informed reviewers.

Third, the expansion of AI applications into statistical review and quality control has direct implications for editorial and peer review practices in internal medicine journals. Empirical studies published in 2025 show that AI‑assisted tools are capable of identifying selected reporting deficiencies, inconsistencies with statistical reporting guidelines, and certain formal statistical errors with moderate‑to‑high agreement, as compared with human reviewers.^25,30,32 However, as highlighted across the reviewed evidence, these tools have limited capacity to evaluate context‑dependent methodological decisions that are common in internal medicine research, including analyses involving multimorbidity, complex exposure definitions, composite or competing outcomes, and clinically motivated subgroup analyses. Consequently, while AI may support initial screening or standardization of statistical review, the evaluation of methodological adequacy and clinical plausibility of statistical choices in internal medicine manuscripts must remain the responsibility of experienced human reviewers.

Fourth, the reviewed literature highlights potential risks associated with AI‑assisted statistical analyses that are relevant when such tools are applied to the interpretation of clinical research in internal medicine. These risks include masking of methodological flaws through fluent narrative explanations, reinforcement of inappropriate default analytical choices, and overconfidence in statistically significant findings without adequate consideration of assumptions, uncertainty, or study limitations.^14,27,31 When transferred to the context of internal medicine, where statistical results often inform broad diagnostic and therapeutic decision‑making across diverse patient populations, these risks may affect not only methodological rigor but also the clinical interpretability and credibility of published evidence.

Finally, the reviewed evidence supports a model in which AI contributes primarily to scalability and efficiency in statistical workflows, while responsibility for analytical validity remains with human experts. Across the analyzed studies, hybrid approaches in which AI assists with preliminary screening, data extraction, or generation of analytical syntax, and statisticians retain responsibility for validation and interpretation, appear most consistent with the strengths and limitations identified for current AI tools.^16,26,36 When such models are considered in relation to research published in internal medicine journals, this implies that any integration of AI into statistical analysis or statistical review should be explicitly framed as supportive rather than substitutive, and accompanied by clear editorial policies regarding disclosure, acceptable use, and the continued primacy of expert statistical judgment. In the editorial phase, AI‑assisted tools may play an increasingly prominent role in automated screening for reporting guideline compliance (eg, CONSORT or STROBE), verification of citation formatting, and detection of incomplete methodological reporting. Such applications are well‑suited to structured, rule‑based checks, and may enhance consistency and efficiency in manuscript processing. However, their integration should complement, rather than replace, expert editorial and statistical oversight.⁴⁵

Given the rapid evolution of LLMs, including the emergence of reasoning‑oriented (“thinking”) systems and agentic workflows, findings from earlier studies should be interpreted within the temporal context of the specific model version evaluated. Continuous reassessment will therefore be necessary as AI capabilities further develop.