Medical Literature in the Age of Artificial Intelligence

The utilization of artificial intelligence (AI)-based clinical decision-support tools, including large language model chatbots (e.g., ChatGPT) and retrieval-augmented search engines (e.g., OpenEvidence), is rapidly increasing across medical specialties and care settings with the potential for greater clinician uptake than traditional clinical decision-making references 1-3. Clinicians increasingly employ these tools, which query the published medical literature to generate differential diagnoses for common and rare conditions or to identify optimal treatments 4, 5. Several AI tools may approach or match licensed physician accuracy in some pre-selected clinical vignettes 6-8. Others have already demonstrated utility in clinical prognostication 9-12. As these AI-based platforms increasingly become a destination for health care professionals to work through complex clinical decisions, a trend reflected in public search data, changes in how clinicians access medical information will undoubtedly lead to transformations in medical practice and literature 2. These facts prompt important questions: How will the medical literature adapt, and how will relevant stakeholders respond? Among the most consequential implications for the medical literature are those for review articles. Narrative reviews offer an accessible synthesis of a topic to answer specific questions with the researcher's perspective influencing the work, whereas systematic reviews use standardized search and reporting procedures to enhance methodological rigor. Both have a well-defined place in the current medical literature landscape. Systematic reviews and meta-analyses provide valuable tools for assessing evidence strength and evaluating risk of bias using predefined, often stringent, inclusion and search criteria 13. Grading of evidence quality and evaluation of bias are likely the capabilities that most AI tools are currently unlikely to replicate due to the inherently difficult nature of this process—even with standardized rubrics 14. Nevertheless, review articles face a fundamental limitation: They are fixed in time. Individual literature reviews are unable to adapt to the thousands of new articles published weekly (some 80 new systematic reviews are published per day alone), leaving a large potential knowledge gap 15. AI clinical decision-support tools offer the ability to answer specific questions and maintain rigor in search based on the end user's query, though they are frequently prone to bias 5. Compared with a review, these platforms can be updated in near real time, which in the realm of rapidly changing clinical guidelines and new drug approvals offers a significant benefit for clinicians. The case report and case series, once the backbone of medical literature, have fallen out of favor due to the development of advanced methodologies and large patient cohorts 16. Many medical journals no longer accept cases due to concerns regarding perceived low evidence quality, and even if published, they often receive little attention 17. However, they have the potential to regain prominence if reconceptualized as peer-reviewed data points to be used by AI-based clinical decision-making tools. Case reports, similar to other article types, are already indexed by many large language models. In this paradigm, each peer-reviewed case report contributes to a collective disease phenotype aggregated on demand when a relevant query is entered into the platform. In particular, this may result in diagnostic and therapeutic advances for rare, orphan, and long-tail diseases (e.g., IgG4-related disease or Gaucher disease) where large randomized controlled trials are not feasible 18. Case reports of these diseases could be indexed and linked by AI to similar reports, with the synthesized patterns returned to clinicians, enabling better diagnosis and treatment of rare conditions. The pharmacovigilance literature offers one illustrative precedent where individual adverse drug reaction reports have demonstrated value prior to the widespread use of AI 19. In a hypothetical example, an adverse effect from a newly approved drug is documented in several case reports across different journals and languages. This information is then synthesized by an AI clinical decision-support platform, which communicates the emerging signal to a treating clinician in a matter of seconds. Returning multiple references alongside the model's clinical reasoning could improve clinician understanding, trust, and interpretability 20. Despite considerable promise, AI clinical decision-support tools have many recognized limitations 21. They currently lack “an expert's touch” from a seasoned physician, which may be gained from a narrative review. Further limitations include the user's phrasing of the question, the model design, and the data made available for inclusion 22, 23. It has already been shown in a large clinical vignette study that currently available platforms are highly biased by how a clinician phrases a specific query, which may paradoxically reduce diagnostic accuracy, leading to misdiagnosis and clinician overconfidence 24, 25. Beyond query sensitivity, broader concerns regarding model transparency persist. How these AI models interpret and select data for inclusion when responding to a query is poorly understood 4. Further guardrails against hallucination or fabrication need to be implemented, such as providing the references used by the model for independent verification. Reporting by medical AI platforms to clinicians of a model's confidence (i.e., perceived level of answer certainty) is needed so that clinicians may adapt their expectations accordingly. This change could reduce automation bias, the tendency to uncritically accept AI advice, thereby preventing downstream patient harm 26. The quality of model inputs (i.e., the medical literature itself) will likely increase in importance as stakeholders eventually attempt to achieve higher accuracy 5. Moreover, enhancing the volume of literature accessible to the model is crucial for the success of AI clinical support tools. The demonstrated increase in the publication of medical data in open access journals and associated increases in accessibility will benefit clinicians by enabling better access by AI models to scientific articles 27. Authors selecting this format for publication will undoubtedly increase the visibility of their research to AI models as copyright concerns resurface as a major long-term issue 28, 29. Clinicians are likely to see scientific publishers continue to form partnerships with corporations developing AI clinical decision support tools, which could improve health-care professional uptake and article accessibility for inclusion in the model. Publishers, editors, and academic institutions must act proactively to ensure the medical literature is optimized for an AI-supported future; thus, changes in academic departments and among the editors and publishers of medical literature will need to be made. The format of case reports, and all medical literature, would benefit from standardization to promote more accurate indexing and analysis by AI models 30. Although authors' use of currently available case report consensus guidelines, which seek to improve clinical rigor (i.e., CARE), is likely to make an immediate impact, a larger community undertaking to improve case report quality (and the broader literature entirely) is needed 31. Potential targets for change include improved use of standardized keywords (e.g., MeSH terms) and shifts toward universal abstract formatting, which could enhance machine readability. For medical trainees and seasoned faculty, adaptation to this increasingly AI-augmented profession likely includes enhanced training in how to form an AI query for the best result, awareness of how individual queries influence model bias, and techniques for critically assessing AI recommendations 22, 25. A greater emphasis should be placed on developing health-care professional AI literacy, which is likely low 32. Proactive changes, such as standardized reporting of study findings, enhanced education programs, and reconsideration of the value of medical case reports, have the potential to strengthen the clinician–AI–patient relationship and improve clinical decision-making. Hunter Scott: conceptualization, investigation, writing – original draft, project administration, software, data curation, supervision, resources, methodology, visualization, writing – review and editing. Claude 4.5 was used for the refinement of spelling and grammar. The author has nothing to report. The author has nothing to report. The author declares no conflicts of interest. The author has nothing to report.