Integrating artificial intelligence tools in health research Authors: Gemma Postill, Jana Sedlakova, Stefano Tancredi, Frerik Smit, Sintieh Ekongefeyin, Andreas M Baumer, Arnaud Chiolero, Jürgen Bernard, Laura C Rosella, Viktor von Wyl Published: 2026 ABSTRACT AbstractDisciplines conducting health research rely on discipline-specific workflows that are distinct from data science, the discipline which develops most artificial intelligence (AI)-enabled research tools. Here we reflect on the impact of such differences in workflows and provide practical advice for leveraging AI-enabled research tools in health research. Further, we provide recommendations for how to prepare the next generation of health researchers for the inevitable influx of research integrating AI-enabled tools.Subject terms: Business and industry, Computational biology and bioinformatics, Health care, Mathematics and computing, Scientific community Introduction With recent technological advancements, artificial intelligence (AI)-enabled research tools are increasingly being used, to support, and even conduct, clinical and population health research1,2. These tools include those that assist with research objective generation, data preparation, analyses execution, data visualization, and research output summarization3. In theory, humans plus AI-enabled research tools would be better than either alone4. Indeed, the quality of AI-enabled research tools and the collaborations produced with them are rapidly evolving, particularly for tools supporting quantitative analysis of tabular health data5. Key advancements in AI-enabled research tools include improvements in models (e.g., deeper chain-of-thought reasoning, multimodal architectures, and fewer hallucinations) and the user interfaces6–9. Yet, concerns still remain regarding the quality and future directions of AI-enabled health research10. So, with better models and user interfaces, why has AI-enabled research not yet surpassed human research and how can researchers use the existing tools? Disciplines conducting clinical and population health research, such as epidemiology, traditionally rely on discipline-specific workflows distinct from those of data science. Many AI-enabled research tools, in contrast, originate and are trained on software code bases from data science (e.g., the interdisciplinary field that draws on computational, statistical and mathematical fields, and subject expertise to generate knowledge from data) and, therefore, reflect the workflows and language of data science. Although there is potential to leverage these AI-enabled tools to accelerate or augment clinical and population health research, their application may change the research workflow in ways not apparent to researchers, resulting in poor-quality research outputs not aligned with current disciplinary standards. How should we leverage tools for research not structured around the typical research workflow? In an international collaborative effort, we sought to (1) reflect on how the research workflows of health research and data science contrast, (2) provide practical advice for leveraging AI-enabled research tools in clinical and population health research, and (3) reflect on the broader training implications for future clinical and population health researchers. To provide concrete, actionable advice, we focus on the discipline of quantitative epidemiology (a discipline of health research focused on distribution, occurrence, and determinants of health and disease in a population) as it is our shared expertise; however, these insights can be generalized to other disciplines that conduct health research. Similar challenges in structuring, governing, and operationalizing data science workflows in health research have been documented across a range of international contexts, and highlight common issues related to data access, reproducibility, governance, interdisciplinary coordination, and translation, which all intersect in complex ways with AI tools for analysis, suggesting these challenges are not context-specific but globally shared11–14. Contrasting the workflows of epidemiology and data science To understand the implications of embedding AI into health research workflows, the first step is to identify the commonalities and differences between a specific health research discipline (e.g., epidemiology) and data science. By having a clear picture of the differences in research workflow structure, one can better leverage the benefits of AI-enabled research tools, minimize their potential harms, and produce high-quality research. Epidemiology Epidemiologists are trained to generate or use existing data deliberately, guided by study design, and ideally from well-defined populations (Fig. 1). There is much emphasis on the method of collection and purpose of the data (e.g., “designed” versus “organic” data)15. Research questions typically fall into three categories: descriptive, predictive, or causal16. In answering these questions, the emphasis is placed on developing rigorous protocols that specify research questions and analytic plans in advance17. This includes clearly defining target and study populations, as generalizability and transportability of study findings are key elements of the epidemiology workflow18. These practices reflect the generally high stakes of clinical and population health research19, where poor scientific practices and biases can lead to erroneous findings that fail to inform improvements in health outcomes and may ultimately harm populations20. Comparison of the research workflows across the academic disciplines of health science and data science. The figure depicts what each discipline emphasizes at each step. The health research workflow presented is intended to be representative of the typical epidemiology workflow for tabular, quantitative health data analyses and the workflow of data science representative of computationally driven data science. There is heterogeneity within the fields of health research and data science not captured in this figure. Data science While there is a range of technical expertise within data science, the common understanding of data scientists is that they accept data from collaborators in some real-world setting and leverage their computational skills to generate data-driven insights and support decision-making (Fig. 1). As such, most data scientists work with data without an explicit step to conceptualize the study design21. Data scientists identify problems relevant to available data and transform data into useful outputs for downstream tasks based on domain-specific needs. This process of iterative exploration and transformation is central in preparing for the analysis and often the most resource- and time-intensive phase22. During this phase, data scientists ensure that data is accessible, governed, curated, and structured before analyses begin. The actual data analysis phase can be multi-faceted, reflecting a range of technical depth and roles, as well as iterative. The development of AI-enabled research tools by AI engineers is one branch of data science analysis. The final step in the data science workflow involves integrating subject matter expertise to make data meaningful and credible for their tasks and needs in decision-making. Advice for how to integrate AI-enabled tools for epidemiological research Given the cultural and methodological differences between epidemiology and data science, the use of AI-enabled tools in epidemiological research raises concerns regarding how it may impact a study’s validity, rigor, transparency, and replicability. However, by recognizing such differences and striving for common understanding, we can begin to ground AI-enabled tools in rigorous epidemiological research design while also embracing the efficiency and power that these tools can bring. The advice provided below is thereby intended to serve as a guide for helping us achieve this wider objective. Never assume shared meaning; familiarize yourself with the terminology of data science and ensure appropriate interpretation of AI-enabled research tools Many commonly used terms and concepts (e.g., modeling, prediction, and validation) carry different meanings in epidemiology and other computational disciplines (Table 1). Likewise, different terms may be used across these disciplines to mean the same thing. These differences can hinder collaboration, lead to confusion, and cause misinterpretation when AI-enabled tools are applied to research tasks23–25. For example, in epidemiology, (statistical) significance is commonly tied to statistical hypothesis testing, with strict pre-specified thresholds (e.g., p < 0.05). In contrast, in data science, significance might refer to the weight or influence of a feature (corresponding to a variable in epidemiology) in a predictive model. In the context of AI-enabled research tools, terminology differences are even more problematic, as some tools allow the use of natural or lay language rather than code to run analyses, with the intention of making data analysis more efficient and accessible. [4] Table 1Examples of terminology where the meaning or application behind commonly used terminology may differ between disciplinesResearch workflow componentCommonly used terminologyUse and emphasis in clinical and population health scienceUse and emphasis in data scienceResearch question definition“Problem formulation”Emphasis is on clearly defining a research question structured as descriptive, predictive, or causal. Requires explicit definition of the study and target populations.Focuses on task specification for modeling. Often occurs after initial data exploration to inform the problem formulation based on available data and computational resources.“Model selection”Based largely on the research question and informed by the data available (e.g., whether data satisfies assumptions of model and the target population).Based largely on the data available (features and label). Common to select multiple models that have a relevant structure and compare the outputs.“Variable”/“Feature”Analytical inputs are typically referred to as variables. Variables are often differentiated based on the type of analysis and the role they play (e.g., estimate an outcome in a model, adjust casual associations, or assess causal relationship → predictor, confounder, exposure). Variables are commonly included based on known or suspected relationshipsAnalytical inputs are typically referred to as features. Selected based on availability and their inclusion in the final analysis, this is typically determined by the ability to optimize predictive accuracy regardless of potential for causality.Data collection, curation, and cleaning“Selection bias”A major concern for internal and external validity58. Minimizing selection bias is critical for ensuring representativeness of data, and for internal and external validity. Requires explicit definition of the study and target populations.Often overlooked as datasets are typically pre-existing and the specific study questions is usually defined after a preliminary analysis. Effects on model performance (e.g., through causing algorithmic bias) are more evident in external validation.“Feature engineering”Focus is on selection of variables (usually based on a priori knowledge), data cleaning, and variable transformation with the purpose of improving the quality of the data, satisfying model assumptions, or improving interpretability of outputs. Less emphasized with this framing (in other conceptualizations of this term, there is focus on functional forms and measurement characteristics).Encompasses a wide range of activities that include encoding, handling missing data, aggregating data, selecting features for modeling, and scaling data. Central step to enhance predictive power. Formalized into pipelines. Also called data preprocessing.Data analysis“Exploratory analysis”Typically called descriptive analyses and done with a specific focus. They serve to summarize distributions, identify trends, and inform model specification; often precede inferential analyses.Emphasis on pattern discovery, anomaly detection, and hypothesis generation rather than inference.“Model assumptions”Critical for the validity of model outputs. Model assumptions often guide the choice of model and the potential variable transformations.Less strict; focus on structured input suitability (e.g., images and an image-processing model).“Hyperparameter tuning”Not used explicitly, parallels in model specification choices. In epidemiology, data are usually optimized for the model (vs. the model optimized for the data).Involves modifying a subset of parameters (e.g., learning rate, number of layers) within a model for performance and fit optimization. Typically done iteratively with feature engineering.Results evaluation“Discrimination”Similar conceptually (how well a model gets the answer right) but different metrics are used (e.g., sensitivity, specificity, PPV, NPV) in addition to AUROC or C statistics.For predictive research questions, goal is to achieve sufficient discrimination for valid inference (vs. to optimize discrimination).Measured using AUROC and related statistics (e.g., accuracy, precision, recall, and F1 scores). Sensitivity is referred to as recall.For predictive research questions, goal is usually to optimize discrimination.“Calibration”Evaluates agreement between predicted and observed outcomes, essential for predictive model validity. Poor calibration would highlight need to revise model selection, variable selection, or variable transformation.Used similarly (e.g., reliability plots, Brier score) though its less emphasized in reporting results and often omitted. Increasingly applied to assess model reliability across subgroups for fairness evaluation.“Algorithmic bias”Not traditionally discussed with this terminology. Greater emphasis is placed on selection biases and information biases (which can cause biased behavior of algorithms).Used as a term that refers to equity and misclassification concerns. Concept is central to AI ethics and fairness assessments. In results evaluation and reporting, often overlooked if model performance metrics are satisfactory.“Validation”Often means comparing results to a gold standard (e.g., diagnostic code validation).Used as a term that refers to performance in another dataset or context. Often distinguished between internal (same data) and external (different data) validation.Communication“Transferability”Typically, discussed with (and distinguished by) concepts of:Generalizability –applying the model to situations where the study sample is derived from the target population. Transportability –application of findings to a population that is external from which the study sample was derived.Used as a term that refers to the ability of a model or its learned patterns to work when applied to new data environments. Successful transferability ensures models are not overfitted to specific training data and can be integrated into different institutional or geographic contexts.“Data drift”In traditional health science methods, typically discussed through related concepts (e.g., temporal bias, measurement changes, or changes in clinical practice patterns). These concepts are recognized as sources of variability that may undermine longitudinal validity and relevance of results.Central concept in the post-deployment lifecycle of ML systems. Reflects changes in the statistical properties of input or outcome variables over time, which can degrade model accuracy and fairness. Monitoring involves continuous auditing, recalibration, and performance tracking to ensure sustained validity in dynamic data environments.“Knowledge mobilization”Refers to the process of translating research findings into actionable knowledge, policies, or interventions that improve population health and healthcare delivery. Involves dissemination, stakeholder engagement, and integration of evidence into decision-making, with an emphasis on transparency, reproducibility, and contextualization within public health or clinical systems.Analogous to the model deployment and implementation phase in AI workflows, where models or analytic outputs are embedded into operational systems or clinical workflows. Includes user interface design, interpretability tools, and human-in-the-loop mechanisms that support responsible adoption, as well as governance structures for maintaining accountability, transparency, and ethical oversight.This table includes selected terminology and is not intended to be a comprehensive list of terminology that may vary across discipline. Likewise, the table cells do not represent dictionary definitions, but rather the author’s reflection of the differing use and emphasis of the terminology in the clinical and population health science as well as data science. To do so, the authors draw on their own expertise and the selected resources15,59–61. For further detail on AI and epidemiology terminology and their definition, see T-CAIREM’s International AI in Medicine Glossary59. For further terminology related to public health surveillance, refer to the “Glossary for Public Health Surveillance in the Age of Data Science.”15ML machine learning, SHAP: shapely additive explanations, LIME Local Interpretable Model-Agnostic Explanations, PPV positive predictive value, NPV negative predictive value, AUROC area-under-the-receiver-operating-curve. Examples of terminology where the meaning or application behind commonly used terminology may differ between disciplines Similar conceptually (how well a model gets the answer right) but different metrics are used (e.g., sensitivity, specificity, PPV, NPV) in addition to AUROC or C statistics. For predictive research questions, goal is to achieve sufficient discrimination for valid inference (vs. to optimize discrimination). Measured using AUROC and related statistics (e.g., accuracy, precision, recall, and F1 scores). Sensitivity is referred to as recall. For predictive research questions, goal is usually to optimize discrimination. Typically, discussed with (and distinguished by) concepts of: Generalizability –applying the model to situations where the study sample is derived from the target population. Transportability –application of findings to a population that is external from which the study sample was derived. This table includes selected terminology and is not intended to be a comprehensive list of terminology that may vary across discipline. Likewise, the table cells do not represent dictionary definitions, but rather the author’s reflection of the differing use and emphasis of the terminology in the clinical and population health science as well as data science. To do so, the authors draw on their own expertise and the selected resources15,59–61. For further detail on AI and epidemiology terminology and their definition, see T-CAIREM’s International AI in Medicine Glossary59. For further terminology related to public health surveillance, refer to the “Glossary for Public Health Surveillance in the Age of Data Science.”15 ML machine learning, SHAP: shapely additive explanations, LIME Local Interpretable Model-Agnostic Explanations, PPV positive predictive value, NPV negative predictive value, AUROC area-under-the-receiver-operating-curve. Consideration of ontological frameworks and metadata structures within health research is also foundational for the integration of AI-enabled tools into clinical and population health research workflows. Shared ontologies and metadata enable consistent representation of clinical and epidemiological concepts across heterogeneous data sources26. Only in ensuring consistent meaning and application of health data (e.g., quality control and data cleaning), alongside appropriate translation of that meaning to data science, can we enable semantically aware integration of AI-enabled research tools27,28. We recommend systematically setting aside time to clarify the key concepts and terminology of the other discipline. Terminology-matching is the approach in which one can respond to a collaborator by rephrasing what is heard with their interpretation of the terminology29–31. Such an approach requires translation-like resources (Table 1) as a foundation for facilitating meaningful dialog about methods and expectations across disciplines. Break down goals and processes into tasks and steps that are clearly documented AI systems and humans both tend to struggle when faced with too much data and analytic complexity. Analysis goals that are too broad or high-level often lead to ambiguity, misinterpretation, and errors. For instance, when users interact with large language models using vague or oversized prompts, the system may produce outputs that do not align with user intent (Box 1). This misalignment stems from the wide space of possible interpretations and decisions embedded in complex tasks. Such is analogous to how human decision-making (and research) can deteriorate when overwhelmed by multifaceted goals without clear structure32–35. With current AI-enabled research tools, especially generative ones, failing to manage this complexity leads to inconsistent and non-reproducible behavior36. We recommend deconstructing complex problems, including the data processing, into smaller, well-scoped steps (Fig. 1). Both human users and AI models benefit from clear well-defined and self-contained tasks. In language models, this is reflected in the effectiveness of chain-of-thought prompting, where systems internally decompose high-level prompts into intermediate steps. This not only improves results but also allows users to trace the reasoning path and verify individual steps. When designing human-AI collaboration (mixed-initiative systems), the idea is to define autonomy, competence, and responsibility at the level of individual atomic actions37,38. This granularity can improve control, transparency, and trust39. It may also be desirable to design workflows that make decomposition explicit, either through user input or by supporting AI self-decomposition, so that decision-making remains observable and revisable. In designing these, considering how they compare to data science workflows can facilitate understanding of the differing assumptions that might be embedded in an AI tool. We tested an AI-enabled tool that leverages multiple large language models to assist users with statistical analyses. The tool allows users to upload datasets and interact with the AI through a chatbot interface to summarize data, generate visualizations, and run specific analyses. It provides step-by-step output along with the corresponding Python code. We used the AI-enabled tool to answer the following causal research question: “What is the causal effect of current smoking on heart attack?” The tool was instructed to address this question using publicly available survey data. We provided two prompts. Our first prompt was simple and deliberately did not include any specific instructions regarding how we would have approached the analysis ourselves, replicating the kind of prompt that a student or an inexperienced researcher might use. “I have uploaded a dataset, and I’d like you to use it to answer the following research question: ‘What is the causal effect of current smoking on having a heart attack?’ Please tell me step by step what you would do and then run the analyses for me.” The AI tool outlined the steps to address the research question and proceeded with the analysis. It identified the exposure and outcome variables, described the data, and ran a logistic regression model, interpreting the results. Code was provided for each step of the analysis. Peer-reviewing model output. At first glance, the process appeared efficient and well-structured. We then proceeded to review the output using the traditional epidemiology workflow and identified several issues. The AI-enabled tool did not adopt a theoretical approach to evaluate the application of the data to the research question using a Directed Acyclic Graph (DAG) or formally identify appropriate variables to adjust for. Instead, it proceeded directly with statistical modeling. The AI-enabled tool did not adopt a theoretical approach to evaluate the application of the data to the research question using a Directed Acyclic Graph (DAG) or formally identify appropriate variables to adjust for. Instead, it proceeded directly with statistical modeling. The AI-enabled tool misinterpreted the odds ratio as an increase in probability (vs. in odds). The AI-enabled tool misinterpreted the odds ratio as an increase in probability (vs. in odds). The results were not reproducible. When we asked the tool to repeat the analysis using the same prompt, the results differed, raising concerns about the consistency and robustness of the tool’s output. The results were not reproducible. When we asked the tool to repeat the analysis using the same prompt, the results differed, raising concerns about the consistency and robustness of the tool’s output. Our second prompt was intended to replicate a scenario whereby experienced researchers provide additional guidance to the AI-enabled tool. “I’m uploading a dataset to answer the following research question: “What is the causal effect of current smoking on having a heart attack?” Please first create a DAG and then conduct the analyses for the corresponding DAG.” Peer-reviewing model output. The AI-enabled tool generated a plausible DAG that was conceptually meaningless (e.g., not defensible based on known predictors in the literature). The DAG was also not used by the model in the subsequent analytical steps. After having produced the DAG, the tool attempted to run the analysis but was not able to convert a string to a numerical value, failing to complete the task. This was unexpected, as the same issue had not occurred with the prompt 1. Exercise Reflection. This simple experience highlighted some strengths and limitations of using AI for data analysis. The tool generated reasonable analytical steps, explained its methods clearly, and provided usable Python code. However, there were substantive errors in these outputs (particularly those unique to epidemiology, such as DAGs). When outputs are plausible but incorrect, they risk reinforcing misconceptions and giving users, especially those with limited expertise, a false sense of security that can ultimately hinder rather than support learning. Define the purpose of an AI-enabled tool with respect to the workflow step it is applied to and the degree of human involvement An important distinction in the use of AI-enabled research tools concerns (1) the workflow building block that is AI-enabled and (2) the degree of AI versus human involvement40. When humans and AI are paired without clarity about how decisions are being made and at what level of autonomy, collaboration tends to underperform regardless of how well their capabilities could theoretically complement each other10. Lack of clarity on an AI-enabled tool’s role in the research workflow can lead to a pedagogical research trap, where researchers underuse critical thinking skills and lose efficiency gains40. Opportunities exist to integrate AI in each step of the research workflow41. However, not all steps in a scientific workflow carry equal risk, nor do they require the same level of judgment or responsibility. Each step in the health research workflow is aligned with specific overarching scientific principles such as reliability, reproducibility, or validity. Likewise, as is done with automated vehicles42, AI-enabled tools (agents) can be classified into levels of autonomy (Fig. 2), depending on the amount of control the user wants to give the tools. Rather than blindly applying AI to a step in the research workflow, we must deliberately decide which tasks can be safely and ethically supported by or delegated to AI-enabled research tools and which require more nuanced human judgment and oversight. As such, these level assignments are also an expression of trust into AI-autonomy, given current AI capabilities, analysis objectives, error tolerance, and liability. Classification of artificial intelligence (AI)-enabled tools for research based on their level of automation and user involvement. Classification adapted from that used for autonomous vehicles by Soder et al.42. In interpreting this figure, note that the degree and form of human involvement, and thus control, can vary widely within each category. For example, some of the research tools explain rationales or provide the code used to support the analysis (as such allowing identification of assumptions and errors made in the analysis), while others do not (e.g., black box-like tools). Making the application of AI-enabled tools and human oversight decisions transparent and defensible is essential for building trust in research conducted with the support of AI and for producing high-quality science, especially given the volume of data currently available43. The respective roles of humans and AI in carrying out the research task should also be clearly spelled out in journal publications and conference papers, rather than being hidden due to stigma around the use of AI-enabled tools. In doing so, we recommend specifying, and ideally prespecifying within protocols, the AI-enabled tools used, as well as the tool’s purpose, limitations, and user knowledge or skill applied in the human-AI collaboration3. Do not assume how AI-enabled tools work Effective human-AI collaboration requires users to understand how these tools work so they can apply the appropriate level of skepticism. AI-enabled tools carry implicit assumptions about epistemology (i.e., the theory of how knowledge is generated), acceptable approaches to output production, and acceptable levels of automation44,45. These assumptions are not always apparent, may not align with epidemiology, and can impact the overall quality and validity of health research3. For example, seemingly transparent models (i.e., models that provide rationales or programming code showing how an analysis was internally conducted) may still mislead users by creating an illusion of human understanding when it may not be how the answer was generated46,47. The use of AI-enabled tools that are epistemically opaque (e.g., in how confounders were selected for an analysis) can also impede the actualization of key scientific principles of transparency and reproducibility. Knowledge of the data-generating process is a key strength in current epidemiology training programs and is foundational to the use of epidemiology for understanding and remediating health disparities. Thus, researchers must develop new skills to evaluate the ability of AI-enabled research tools to provide context-sensitive epistemic elements, such as by applying discipline-specific knowledge to evaluate the selection of a variable adjustment set48. Reject, revise, and accept; approach AI-enabled workflow steps like a peer review Validation of the usefulness of an AI-enabled tool for its intended purpose is important; validation of its output for a specific task is equally important. No tool is perfect; yet the limitations of AI-enabled tools are new and may not be noticed. For example, generative AI models (e.g., ChatGPT) often produce plausible-sounding outputs that can be incorrect (i.e., hallucinations), misleading, or subtly biased49. Generative model outputs are also very heavily tied to the quality of the prompt provided49. These limitations challenge the notion that these tools can provide authoritative guidance or independently act on tasks50. However, without training on use of AI for research, students may assume generative tools are authoritative or intelligent in a human sense, when in fact they rely on probabilistic patterns rather than understanding51. Thus, AI-enabled research tools require a paradigm shift in which researchers and trainees must simultaneously correct the AI-enabled tool while leveraging it for their specific analysis goal. Navigating this new division of labor requires new skills and a reminder of the need to maintain researcher accountability. Practically speaking, researchers must learn to “peer-review” their AI-enabled tools, just as they would scrutinize the work of a colleague (Box 1). This means not only verifying results but also understanding how they were produced, what assumptions were embedded into the system, and what limitations must be communicated to audiences by the authors who take accountability for the research. Researchers can also use chain-of-thought prompting (as previously mentioned) or repeat prompting with varied terminology. In the latter, a researcher would evaluate the consistency of the model’s output in each iteration of the prompt. In this way, using AI limits the risk of shortcut thinking while offering opportunities for deeper engagement with the data and methods, which may be complemented by more efficient output generation. As a secondary benefit, training researchers to review, interrogate, and contextualize AI-generated outputs might also help remediate the human cognitive biases that exist (e.g., when humans overstate their own knowledge or blindly trust heuristics). Using AI-enabled research tools in this way could thus amplify human insight, rather than replace or obscure it. Practice self-reflection in the use of AI-enabled research tools Reflecting on the long-term implications of using AI-enabled tool is part of a broader commitment to preserving the integrity and depth of scientific inquiry. Researchers and students alike should critically reflect upon their motivations, needs, and assumptions when choosing to use AI-enabled research tools. Specifically, they should ask themselves: Should I use an AI-enable tool here and what role is the tool playing in the overall research workflow? The proportionality principle in this context can be a good guide: the benefits of using a new tool should outweigh its risks, and they should also outweigh the benefits of the method it replaces52. Researchers must be able to pause and interrogate their own decision-making: Am I relying on this output? What assumptions underly the AI’s model? Could alternative interpretations or methods be more appropriate here? Such reflective practices do not merely safeguard against error; they form the basis of epistemic responsibility and scientific integrity. For trainees, this is particularly important, as they may lack the experience and methodological proficiency to assess the quality of these tools, the appropriateness of the analytic techniques applied, or to detect errors53. Relying too early on AI-enabled tools that outsource key cognitive tasks will hinder the development of competencies essential to critically evaluating their outputs and achieving long-term research success. We advise that students must first develop foundational health research skills such as critical appraisal, study design, and conceptual reasoning. Implications for teaching and training health researchers Guidance put forth in this paper has implications for the skills needed by health research trainees (Table 2). In the context of AI-enabled research tools, educational programs must instruct trainees on how to breakdown the research processes of their own disciplines, as well as data science, into modular workflows (Fig. 1). Health research trainees should learn how to coordinate, not just execute, the steps of a research workflow with AI-enabled tools. This includes clarifying how responsibility is actioned at different levels of human–AI collaborations and how to detect and address errors in AI outputs. Table 2Actioning the proposed recommendationsPractical adviceSkills, knowledge, and abilities needed by health researchersEducation initiatives example1. Never assume shared meaning; review terminology of disciplines that develop AI-enabled research tools prior to their use1. Conceptual understanding of data science and AI2. Ability to integrate terminologies from multiple disciplines.3. Ability to effectively elicit definitions from collaborators4. Comfort with rapidly acquiring new disciplinary languagesDevelop opportunities for health researchers to be introduced to common AI, machine learning, and data science terminology, then engage in interdisciplinary group activities to practice defining these terms in their own words and explore differences in disciplinary interpretations. Practice in interdisciplinary settings is key.2. Break down goals and processes into tasks and steps5. Ability to decompose various research approaches into modular workflows6. Structuring and sequencing research workflow tasks efficiently (systems thinking skills)7. Understanding interdependencies within research workflows8. Capturing and communicating decomposed tasks clearly (e.g., documentation)9. Analytical thinkingFacilitate exercises where researchers deconstruct AI research papers or case studies into discrete goals, tasks, and steps, mapping these visually to build an understanding of implementation processes.3. Define the purpose of an AI-enabled tool with respect to workflow step and degree of human involvement.10. Understand traditional research workflows, including the role of pre-specified analyses and protocols.11. Knowledge of human factors affecting research outputs (i.e., appreciating how human cognitive load interacts with automation)12. Ability to evaluate the value and impact of applying AI-enabled research tools13. Understanding barriers to (and reasons for not) integrating AI into workflows where appropriateConduct scenario-based workshops where participants analyze research workflows, identify components suitable for AI automation, and discuss implications for human roles in each step. Practice automating specific research workflow components and evaluating differences.4. Understand how AI-enabled tools work to ensure responsible use.14. Understand common AI models, their limitations, and strategies used to mitigate their limitations.15. Recognize common AI mistakes (e.g., calculation errors, coding mistakes, hallucinations)16. Evaluating input data integrity, particularly in relation to an AI-enabled tool’s design17. Fairness and bias assessmentUtilize journal club or case study formats to critically appraise examples of AI research tools uses and outputs, focusing on model assumptions, limitations, and potential biases, and practice identifying errors or misinterpretations.5. Reject, revise, and accept; approach AI-enabled workflow steps like a peer review.18. Foundational coding skills19. Debugging skills (e.g., diagnosing code or output errors systematically)20. Critical appraisal skills for AI-enabled outputs (e.g., model output proofreading or peer-reviewing skills)21. Familiarity with established research and reporting guidelines to assess outputs from AI-generated analyses22. Feedback delivery skills for providing constructive, actionable critiques in interdisciplinary contexts.Incorporate practical sessions where researchers review AI-generated analytic code or outputs, propose revisions, and discuss their reasoning, simulating peer-review processes but for AI outputs.6. Practice self-awareness in use of AI-enabled technology in research23. Reflective practice skills (e.g., structured reflection on choices and biases)24. Ethically reflect on the impact of automation25. Ethical and legal awareness in AI for health26. Policy literacy, including familiarity with institutional, national, and international AI research policies.27. Version control and reproducibility (i.e., good documentation practices)28. Commitment to updating knowledge with evolving AI ethics and standards.Foster teaching and research environments that promote values in addition to efficiencies, such as methodological rigor, critical thinking, and epistemic humility.Identify and discuss the potential pitfalls of AI research tools, such as epistemic opacity, over-reliance, or unjustified delegation, with trainees as part of routine practice. Develop modules on reproducible research practices, including version control, and facilitate discussions on the ethical and legal implications of AI use in health research, as well as personal responsibility for its outcomes.Skills needed by health researchers to leverage AI-enabled research tools and examples of education initiatives that can enable instruction of these skills.AI artificial intelligence. Actioning the proposed recommendations 1. Conceptual understanding of data science and AI 2. Ability to integrate terminologies from multiple disciplines. 3. Ability to effectively elicit definitions from collaborators 4. Comfort with rapidly acquiring new disciplinary languages 5. Ability to decompose various research approaches into modular workflows 6. Structuring and sequencing research workflow tasks efficiently (systems thinking skills) 7. Understanding interdependencies within research workflows 8. Capturing and communicating decomposed tasks clearly (e.g., documentation) 9. Analytical thinking 10. Understand traditional research workflows, including the role of pre-specified analyses and protocols. 11. Knowledge of human factors affecting research outputs (i.e., appreciating how human cognitive load interacts with automation) 12. Ability to evaluate the value and impact of applying AI-enabled research tools 13. Understanding barriers to (and reasons for not) integrating AI into workflows where appropriate 14. Understand common AI models, their limitations, and strategies used to mitigate their limitations. 15. Recognize common AI mistakes (e.g., calculation errors, coding mistakes, hallucinations) 16. Evaluating input data integrity, particularly in relation to an AI-enabled tool’s design 17. Fairness and bias assessment 18. Foundational coding skills 19. Debugging skills (e.g., diagnosing code or output errors systematically) 20. Critical appraisal skills for AI-enabled outputs (e.g., model output proofreading or peer-reviewing skills) 21. Familiarity with established research and reporting guidelines to assess outputs from AI-generated analyses 22. Feedback delivery skills for providing constructive, actionable critiques in interdisciplinary contexts. 23. Reflective practice skills (e.g., structured reflection on choices and biases) 24. Ethically reflect on the impact of automation 25. Ethical and legal awareness in AI for health 26. Policy literacy, including familiarity with institutional, national, and international AI research policies. 27. Version control and reproducibility (i.e., good documentation practices) 28. Commitment to updating knowledge with evolving AI ethics and standards. Foster teaching and research environments that promote values in addition to efficiencies, such as methodological rigor, critical thinking, and epistemic humility. Identify and discuss the potential pitfalls of AI research tools, such as epistemic opacity, over-reliance, or unjustified delegation, with trainees as part of routine practice. Develop modules on reproducible research practices, including version control, and facilitate discussions on the ethical and legal implications of AI use in health research, as well as personal responsibility for its outcomes. Skills needed by health researchers to leverage AI-enabled research tools and examples of education initiatives that can enable instruction of these skills. AI artificial intelligence. Equally important is the need to teach with AI-enabled tools (vs. only about AI) so that trainees gain firsthand experience with their capabilities and limitations54,55. Integrating tool use into the learning process can demystify the technology, sharpen students’ evaluative skills, and foster a mindset of shared accountability between human researchers and machine assistants56,57. Educating trainees is particularly important, as their exposure to and collaboration with AI-enabled tools occurs whilst they are developing research skills and experience. Case studies and opportunities to practice as well as receive feedback on their ability to critically appraise the outputs of AI-enabled tools can equip students with the necessary skills to become successful independent researchers (Table 2). Revamping education in light of technical advancements is not a new phenomenon. Previously, with the growth of statistical software systems (i.e., SAS, R, and STATA), epidemiological education shifted from emphasizing hand calculation to computer-enabled statistical tests. Doing so involved many iterations of identifying (and instructing) how one should trust a software’s output. We can learn from this experience and apply analogous principles to the integration of AI-enabled tools in health research (Table 2). Embracing and preparing for change With the increasing use of AI-enabled tools in clinical and population health research areas, we face an urgent need to reflect on how we structure and teach their integration into existing research workflows. Effectively leveraging them in existing research workflows requires understanding the research workflows of the disciplines that develop AI-enabled research tools (data science) and of the health research disciplines that use those tools. In applying these tools to health research disciplines, reflection is needed on terminology differences, the modular components that constitute research workflows, the components of research workflows that AI should support, and the technical structure of the AI-enabled research tools. Application of these tools must be complemented with self-reflection on the ultimate role they serve, peer-review of model outputs, and transparent reporting of their use. Leveraging AI-enabled tools for health research in this way, with humans at the center and health research standards upheld, can support the continued production of high-quality, robust, ethical, and impactful health research. By highlighting key considerations for the integration of AI-enabled tools in health research our findings provide a useful starting point for researchers, educators, and trainees, enabling the development of concrete frameworks for validating AI outputs in high-stakes health research contexts like the one addressed in the paper. Acknowledgements The authors would like to acknowledge the University of Zurich Global Seed Fund grant for their financial support of this work. Author contributions G.P. and J.S. led the drafting process of the initial manuscript and coordinated revisions. S.T. and F.S. led the analysis exemplifying the application of peer-review to identify potential pitfalls of AI-enabled tool. All authors S.T., F.S., S.E., A.B., A.C., J.B., L.C.R, and V.v.W. provided conceptual input and reviewed the manuscript critically for important intellectual content. All authors read and approved the final manuscript. Data availability For the analyses described in Box 1, we used a publicly available dataset (https://www.kaggle.com/datasets/kamilpytlak/personal-key-indicators-of-heart-disease). No additional datasets were generated or analyzed during the current study. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Gemma Postill, Jana Sedlakova. These authors jointly supervised this work: Laura C. Rosella, Viktor von Wyl. Change history 9/21/2026 A Correction to this paper has been published: https://doi.org/10.1038/s41746-026-03286-x Contributor Information Gemma Postill, Email: gemma.postill@utoronto.ca. Laura C. Rosella, Email: laura.rosella@utoronto.ca Viktor von Wyl, Email: viktor.vonwyl@uzh.ch. References Associated Data Data Availability Statement For the analyses described in Box 1, we used a publicly available dataset (https://www.kaggle.com/datasets/kamilpytlak/personal-key-indicators-of-heart-disease). No additional datasets were generated or analyzed during the current study.