The integration of artificial intelligence (AI) multiagent systems into soil science research, as proposed by the Frontiers in Science lead article of Minasny et al. (2026), represents a compelling development in the use of AI to support scientific discovery. AI's significance clearly goes beyond the task-specific machine learning (ML) tools already established in areas such as digital soil mapping (Lamichhane et al., 2019), soil spectroscopy (Viscarra Rossel et al., 2016), and pedotransfer functions (Weber et al., 2024). By combining specialized AI agents that can synthesize literature and data, generate hypotheses, and propose experimental designs, such systems could accelerate several stages of the research process (Krenn et al., 2022).However, applying AI research agents to soil science raises important questions due to the nature of soil as the object of research. Soils are inherently complex, spatially heterogeneous (Jenny, 1941), and often only sparsely observed, while many key soil processes unfold over timescales that exceed the duration of most available datasets. These characteristics limit how fully soil systems can be captured in data, and therefore how effectively AI agents can generate meaningful insights from them. In this viewpoint, I argue that while multi-agent AI systems may substantially improve research efficiency, their successful use in soil science will depend on domain expertise to guide and critically evaluate AI-based workflows.Current applications of AI in soil science are largely based on ML models trained to perform clearly defined analytical tasks. For example, ML models can predict topsoil clay content from environmental geodata (Chen et al., 2022) or estimate the laboratory-measured organic carbon content of a soil sample from its spectral reflectance (Viscarra Rossel et al., 2016). In such applications, researcher remains central to the workflow: scientists decide which datasets to use, how to prepare them, how to handle missing data, which modeling approach to select, and how to evaluate model performance. These decisions are based on the study objective, available data, and established good practice in the field. The ML model then performs the specific task for which it was designed, at a pre-defined step within that workflow.In contrast, AI research agents go beyond this scientist-centered workflow by shifting some of these decisions from explicit human judgment to algorithmic inference. Rather than operating within a predefined workflow, an AI agent can iteratively analyze a dataset and propose analytical strategies including preprocessing steps, variable selection and transformation, and suitable modeling approaches. In effect, steps that have traditionally formed part of the researchers' analytical reasoning-including generating and evaluating hypotheses (Minasny et al., 2026)-are no longer solely human decisions, but are increasingly inferred in collaboration with AI systems. This shift could greatly increase efficiency and make knowledge and tools more accessible, particularly in interdisciplinary fields such as soil science where research often requires expertise spanning multiple domains.Although global soil datasets may appear large (Batjes et al., 2023), they remain sparse relative to the complexity of the soil systems they are meant to describe. Soils form through multi-layered interactions among climate, organisms, topography, and parent material over very long time periods, and many relevant processes operate simultaneously across wide spatial and temporal scales (Jenny, 1941). Yet, observations of these processes are limited and indirect. Only an extremely small fraction of the world's soil volume has ever been sampled, and many measurements capture static properties or simple proxies rather than the processes themselves. Greater spatial or temporal resolution does not necessarily solve this problem. For example, remote sensing imagery measures spectral reflectance only from the top millimeter of bare soil, and this signal must be locally calibrated before it can be linked to soil properties. Likewise, soil moisture sensors measure volumetric water content at a point but do not directly capture processes such as preferential flow or macropore connectivity.Even when collected at high temporal resolution, such measurements provide only a partial representation of unsaturated flow behavior. In this sense, digital soil twins derived from sensor and remote-sensing data view soil systems through a keyhole only. As in Plato's cave, interpretations are based on observable projections of the system, while important underlying processes remain only partly visible.In addition, soil datasets are often highly heterogeneous in how measurements are collected and reported. Harmonization is therefore essential, but it is often difficult to achieve because differences in measurement protocols are frequently confounded with differences in land use or pedoclimatic region. In such cases, observational variation cannot easily be disentangled from the underlying soil processes themselves (Weber et al., 2024). As a result, AI-based knowledge discovery may identify measurement artefacts as meaningful patterns rather than methodological inconsistencies.Beyond data harmonization, data access remains a recurring challenge in soil science. The effectiveness of AI research agents depends on centralized datasets, many of which are held by public institutions where data sharing often relies on trust and a clear understanding of how data will be used. The growing use of complex and potentially opaque AI systems may introduce new concerns for data owners and, in some cases, reduce rather than enhance their willingness to share data.Importantly, AI research agents do not rely solely on structured datasets; they also draw on the broader scientific literature. That literature contains a wealth of expert knowledge accumulated over decades of soil research. However, it is shaped by many of the same practical constraints that affect data collection. Research tends to focus on variables that are comparatively easy to measure and on environments where systematic data collection is feasible. Consequently, both the datasets and the literature from which AI systems learn are biased toward well-observable soil properties and processes. This observability bias is not new in soil science, but it may become more pronounced in data-driven AI research workflows. Moreover, as AI agents provide synthesized outputs, researchers may become further removed from raw data and original studies. This distancing, together with the tendency of AI research to produce outputs that appear coherent but may contain errors (Messeri and Crockett, 2024), can substantially increase the risk of biased conclusions.Examples of this bias can already be seen in digital soil mapping, where terrain is often identified as a dominant soil-forming factor because it is consistently available in spatial datasets (Chen et al., 2022). By contrast, parent material composition or local weathering processes-which may be equally or even more important in many settings-tend to rank much lower because detailed data are often unavailable. Similarly, soil texture is commonly used to explain hydraulic properties because it is widely measured, whereas soil structure and pore connectivity, both of which strongly influence water flow, are much less frequently analyzed because they are harder to sample.The limitations outlined above do not diminish the potential value of AI research agents in soil science. Rather, they underscore the continuing importance of human expertise in guiding their use. Soil science has a long tradition of conceptual frameworks-for example those describing pedogenesis, soil-landscape relationships, or biogeochemical cycles-that represent accumulated scientific understanding and can provide essential guidance for AIdriven exploration. In this sense, the most promising applications of AI agents in soil science may arise not from fully autonomous systems but from approaches that combine data-driven inference with human expertise.One possible pathway is the development of expert-in-the-loop systems, in which AI agents generate hypotheses or analytical strategies that are subsequently evaluated and refined by human experts. In such a workflow, AI agents can rapidly explore large bodies of data and literature, while human expertise provides the interpretive context needed to distinguish plausible mechanisms from spurious correlations. However, integrating soil expertise into computational systems presents its own challenges.Much of the knowledge used by experienced pedologists is tacit and qualitative, developed through field observation and comparative reasoning rather than through explicitly quantified relationships. Concepts such as soil structure development, horizon differentiation, or soil formation along landscape gradients often involve nuanced judgement that is difficult to translate directly into numerical constraints for ML models. As a result, efforts to formalize such knowledge for computational use may be difficult and even meet resistance from domain experts.These challenges suggest that the role of soil scientists in AI-assisted research may increasingly shift toward critically assessing AI-generated outputs and translating domain knowledge in ways that can meaningfully guide AI systems. In this scenario, scientists move away from manually designing every step of a project or analysis and toward supervising AIgenerated research pathways and evaluating their plausibility. Rather than reducing the need for scientific judgment, AI research agents may increase the importance of human expertise, process understanding, and interdisciplinary dialogue to ensure that computational exploration remains grounded in pedological reasoning.In soil science, AI research agents may be most valuable when used as partners in scientific reasoning rather than as autonomous discovery systems.
Madlene Nussbaum (Thu,) studied this question.