In an essay published in the European Journal of Epidemiology, Ulrich Mansmann examines whether artificial intelligence can broaden the conceptual space within which epidemiological reasoning occurs. The analysis traces automated deduction back more than three centuries to Gottfried Wilhelm Leibniz, who conceptualized a formal language, the characteristica universalis, designed to represent scientific concepts with sufficient mathematical precision to allow relations between them to be resolved through a calculus ratiocinator. Faced with intellectual, political, and religious disagreement, Leibniz envisioned disputants sitting down to declare, Calculemus—let us calculate. Leibniz sought a deductive sequence moving from rational structure to formal representation, calculation, and conclusions. Conversely, modern generative artificial intelligence approaches reasoning from the opposite direction. Large language models do not derive from an external system of rational truth; rather, they process the accumulated historical record of human discourse, converting written text into learned representations from which they reconstruct patterns of thought and argumentation. This inversion raises the central question of whether computational systems trained on inconsistent human discourse can help researchers conduct more systematic scientific inference.
The essay highlights that generative artificial intelligence can facilitate a transition from ordinary cognitive tasks to systematic meta-reasoning. Because the textual corpora utilized to train large language models incorporate scientific claims alongside related criticism and debate, an advanced system can compare arguments across distinct domains. Specifically, such systems can juxtapose competing hypotheses, detect incompatible assumptions, propose alternative explanations, and examine whether conclusions logically follow from specified premises. This mechanism does not merely summarize previous statements but interrogates the underlying structure within which competing scientific claims are made. Human epidemiologists inevitably work within limited cognitive and disciplinary horizons, selecting literature from familiar traditions and concentrating on hypotheses deemed plausible within those frameworks. By maintaining simultaneous access to multiple competing explanations and disparate bodies of evidence, artificial intelligence can help investigators counter narrow disciplinary boundaries, allowing researchers to explore scientific questions across broader conceptual parameters.
Crucially, the expansion of a reasoning space must not be mistaken for an automatic progression toward reliable scientific deduction. The analysis underscores empirical research by Zhou and colleagues demonstrating that scaling the size and instructability of language models does not simply produce increasing reliability. Instead of avoiding difficult or uncertain queries, more capable models frequently replace avoidance with answers that appear superficially sensible but are factually incorrect, while human supervisors fail to reliably detect these errors. This finding indicates that the capacity to broaden a reasoning space is fundamentally distinct from the capacity to evaluate that space accurately. An artificial intelligence system deployed in epidemiological research may generate numerous additional hypotheses, yet some of those propositions may be irrelevant, unsupported, or completely false. Conceptual breadth is not equivalent to completeness, and textual consistency does not establish truth. Consequently, artificial meta-reasoning requires a mandatory second component: all generated propositions must connect directly back to empirical evidence, formal logical reasoning, computation, and reproducible analysis.
Epidemiological inference serves as an instructive test case because it rarely involves a direct mathematical deduction from empirical observations to definitive public health conclusions. Instead, epidemiological practice relies upon a multi-stage evidentiary chain progressing from the scientific question to study design, data collection, statistical analysis, causal interpretation, population inference, and final decision-making. Artificial intelligence could contribute by interrogating the coherence of this entire evidentiary chain. Such interrogation involves determining whether a statistical estimand corresponds to the initial scientific question, identifying which causal assumptions are required, surfacing alternative biological or social mechanisms that might explain the observations, and locating contradictory evidence. Furthermore, computational interrogation can assess whether an apparently negative study excludes an intervention entirely or merely disproves the specific implementation tested, while verifying whether relevant target populations or clinical outcomes have been overlooked. These capabilities address the logically prior epistemological question of whether researchers have adequately represented the complete reasoning space surrounding a public health problem.
The concept of a reasoning space also helps distinguish two legitimate sources of variation in scientific inference: epistemic variability and conceptual neighbourhoods. Epistemic variability arises because identical empirical data may support multiple defensible inferential pathways depending on the foundational assumptions, causal structures, statistical estimands, or target populations selected by investigators. An effective artificial intelligence system could make this plurality visible by showing that under assumption A one conclusion follows, under assumption B an alternative interpretation becomes plausible, and evidence C discriminates between them while evidence D does not. Simultaneously, an epidemiological problem can be approached through different conceptual or semantic neighbourhoods. One formulation may foreground chronological age, another individual biological risk, another healthcare implementation, and another societal consequences. While language models are adept at moving across neighbouring conceptual structures to expose overlooked perspectives, their learned representations organize concepts in ways that activate specific associative networks. Consequently, AI-supported meta-reasoning must deliberately search for alternative framings of the question itself rather than merely reproducing the most dominant textual associations found in its training material.
To illustrate how expanding a reasoning space operates in actual epidemiological practice, the essay reviews colorectal cancer screening within the German FARKOR programme. FARKOR evaluated screening among individuals aged 25 to 50 with a familial risk of colorectal cancer. Conventional screening in Germany was predominantly age-based, excluding individuals under age 50 from the national screening programme. However, persons with a positive family history develop colorectal cancer more frequently and at younger ages, with reported relative risks approximately two- to four-fold higher depending on family constellation. Although clinical guidance already recommended earlier colonoscopy for first-degree relatives, population screening remained constrained by age thresholds. The decisive conceptual transition involved shifting the operational question from asking at what age screening should start to asking at what level of risk screening should start. Incorporating family history, degree of relationship, and age of diagnosis transformed an essentially one-dimensional age representation into a multidimensional risk space, illustrating how alternative conceptual neighbourhoods fundamentally alter epidemiological problem formulation.
The empirical findings of the FARKOR programme simultaneously demonstrate why expanding a reasoning space does not produce an automated, definitive policy decision. Among FARKOR participants undergoing colonoscopy, detection rates of adenoma and advanced adenoma were of approximately the same magnitude as those documented in older populations already eligible for conventional screening. An artificial intelligence system examining this clinical scenario could simultaneously survey chronological age, familial risk, the reliability of family-history documentation, screening participation, clinical harms, economic costs, and alternative screening modalities. However, the value of an artificial intelligence system does not lie in declaring a screening age autonomously. Its actual scientific contribution lies in revealing that an apparently coherent, age-based policy argument may rest upon an incomplete representation of the underlying health problem. By making the multidimensional structure visible, computational meta-reasoning prevents premature closure while leaving policy resolution to rigorous empirical validation and human deliberation.
When epidemiological evidence enters public health governance, the distinction between analytical broadening and collective decision-making becomes even more critical. The World Health Organization recently described prospective roles for artificial intelligence across evidence-informed health policy, identifying contributions in problem identification, evidence synthesis, policy design, scenario simulation, implementation, and monitoring. Crucially, the World Health Organization explicitly frames artificial intelligence as augmenting rather than replacing human judgement, emphasizing multidisciplinary deliberation and sustained human oversight. This boundary is vital for population health policies whose consequences unfold over extended periods. Policymaking must simultaneously evaluate current resource constraints, distributional equity, individual autonomy, and competing societal objectives. Artificial intelligence possesses a distinct advantage in maintaining multiple perspectives and extended time horizons concurrently. Once policy objectives are defined, computational tools can examine whether proposed interventions are mutually compatible, identify trade-offs, and explore outcomes under alternative assumptions, making contradictions in collective public health reasoning transparent.
This intersection exposes the inherent boundary of Leibniz’s calculative ambition, requiring a clear demarcation between three domains: evidence, consistency, and values. Artificial intelligence can survey empirical evidence and test consistency between assumptions, arguments, and consequences; however, neither capability determines what society ought to value. Whether public health frameworks prioritize longevity, equity, autonomy, economic prosperity, environmental sustainability, or future generations cannot be derived from epidemiological data alone. Flawless logical reasoning confers neither moral nor democratic authority. The philosophical movement of effective altruism illustrates this boundary by using reason and evidence to optimize limited resources for maximum benefit, assessing opportunity costs and expected impacts in global health priority setting, animal welfare, and existential risks. Yet as calculations extend toward distant or global outcomes, quantitative comparisons become dominated by uncertain probabilities, disputed valuations, and subjective assumptions regarding whose welfare counts. Computational models can make these trade-offs and assumptions explicit, but reason alone cannot specify the objective function that society ought to optimize.
Ultimately, the modern realization of Leibniz’s ambition requires a more modest posture than an uncritical call to calculate. Where Leibniz sought a language from which rational calculation could proceed deductively, generative artificial intelligence starts from the heterogeneous, contradictory record of human thought to reconstruct patterns of reasoning. Its most promising contribution to epidemiology lies in widening the space within which questions are formulated: uncovering neglected hypotheses, connecting evidence across domains, exposing conflicting assumptions, and evaluating coherence across the evidentiary chain. However, because an expanded reasoning space can harbor more error alongside greater insight, artificial meta-reasoning demands rigorous empirical anchoring, methodological scrutiny, and human oversight. The contemporary scientific workflow must therefore heed a three-part invitation: before calculating, researchers must ask whether the problem space is adequately represented; after calculating, investigators must determine whether conclusions follow from evidence; and before taking action, decision-makers must distinguish empirical findings from societal values, ensuring that artificial intelligence illuminates human reasoning rather than replacing it.
Subject of Research: Artificial intelligence and the expansion of epidemiological reasoning spaces
Article Title: From Calculemus to AI: can artificial intelligence broaden the epidemiological reasoning space?
Article References: Mansmann, U. (2026). From Calculemus to AI: can artificial intelligence broaden the epidemiological reasoning space?. European Journal of Epidemiology. https://doi.org/10.1007/s10654-026-01483-2
Image Credits: AI Generated
DOI: 10.1007/s10654-026-01483-2
Keywords: epidemiological reasoning, generative artificial intelligence, meta-reasoning, evidentiary chain, epistemic variability, semantic neighbourhoods, Gottfried Wilhelm Leibniz, calculus ratiocinator, FARKOR programme, colorectal cancer screening, familial risk, large language models, evidence-informed health policy, World Health Organization, effective altruism, causal inference, empirical verification
News Source: Scienmag.
