The Future of Artificial Intelligence in Ecosystem Modeling

Abstract

Developing ecosystem models has traditionally been limited to a small global community of experts because of the complex skills and resources required. However, the emergence of user-friendly artificial intelligence (AI) tools with powerful generative capabilities could democratize ecosystem modeling, enabling both experts and nonspecialists to build models. We explore a speculative future where AI enables automated end-to-end model development and application. Although such tools could accelerate and enhance modeling tasks, their widespread adoption raises concerns about data integrity, bias, interpretation reliability, and the potential erosion of human expertise. We argue that regardless of AI’s technical advancement, human engagement and control remain essential. The global community must respond by identifying key factors that distinguish desirable outcomes and developing infrastructure, standards, and ethical guidelines to ensure AI use in ecosystem modeling remains scientifically robust while supporting sustainable and equitable outcomes.

Graphical Abstract

Graphical Abstract.

Across the scientific endeavor (and knowledge production more broadly), today’s artificial intelligence (AI) tools are transforming tasks such as searching, synthesizing, and coding (Spillias et al. 2024, Zambrano et al. 2023). In recent years, the broad field of AI has undergone a remarkable transformation, with rapid advances in machine learning, natural language processing, and data analytics, which have given rise to a diversity of tools and techniques (Xu et al. 2021). The AI tools being developed are already having major impacts on the way the scientific method is conducted and are unlocking new insights across a range of scientific domains (Xu et al. 2021, Royal Society 2024). The emergence of generative AI models that can perform a range of tasks including gathering data, processing text, and writing software presents an array of opportunities for modern day science. Tasks that were previously performed manually by experts can now be delegated to an AI chatbot (albeit with varying degrees of reliability) as their capabilities increase at breakneck speeds (Ott 2022) and with no clear limit in sight (Hughes et al. 2024).

In the field of ecology, deep learning has already been successfully integrated into well-established frameworks for undertaking correlative modeling of the abundance and distribution of species through space and time (Lapeyrolerie et al. 2022, Tuia et al. 2022). The use of machine learning and other AI methods has seen slower uptake in process-based systems modeling (Karniadakis et al. 2021), which has been a growing field over recent decades, although applications and discussion are emerging (Akhavan and Jalali 2024, Monks et al. 2025).

Ecosystem modeling has long been a cornerstone of ecological science, providing important insights into the complex interactions within natural systems and informing decisions in conservation, resource management, and climate change mitigation (Geary et al. 2020). This modeling approach aims to understand complex trophic and nontrophic interactions and the way in which these are shaped by environmental conditions or disrupted by human activities (Geary et al. 2020). Through time, these models have expanded to extend into more of the human dimensions (i.e., to become socioecological in scope). In the present article, we use the term ecosystem modeling inclusively to refer to a broad range of ecological models and modeling capabilities that seek to represent or simulate nature, including population, habitat, food web, and socioecological models.

Given the urgency to solve increasingly complex global environmental challenges, there is a growing need for more sophisticated, accurate, precise, and responsive modeling techniques and tools to provide guidance on how to sustainably manage socioecological systems. There is a parallel need for models to be more accessible, easy to use and rapidly deployable. Models can be diverse across scale, structure, and application but are often either process-based and rooted in mathematical formulations that aim to simulate or emulate ecological processes (DeAngelis et al. 2021) or are data driven and incorporate machine learning or statistical relationships to derive pattern and insights from data (Crisci et al. 2012). They may even be a mix (i.e., rapid statistical emulators of slow and complex process-based models; Ni et al. 2023). Successfully building a model with enough credibility to be useful for environmental management and decision-making, such as setting limits on natural resource extraction or how to spatially distribute different activities and infrastructure, requires a unique skillset that spans a combination of ecology, mathematics, and computer science and increasingly includes socioeconomics, social sciences, Indigenous knowledge, and stakeholder perspectives. Although there is a growing pool of skilled modelers, there are still significant bottlenecks. Even for skilled modelers, the large time investment required to either develop a novel modeling approach or apply an existing modeling framework to a new geographical context imposes a strong limit on the possible applications (Holden et al. 2024a, Holden et al. 2024b).

Recent and rapid advances in generative AI for a range of general information processing tasks, as well as specific domain knowledge such as mathematics (Trinh et al. 2024) and software development (Becker et al. 2023, Kazemitabaar et al. 2023), present a pivotal moment in the way ecosystem modeling could change from the way it has been historically undertaken. There are now numerous opportunities to use AI tools to gather, synthesize, interpret, interrogate, and analyze data and models. Although the prospect of applying AI to both innovating new ecosystem models and adapting existing models to new contexts is exciting; it is also fraught (Agathokleous et al. 2023, Rillig et al. 2024) with the potential to misguide if the resulting models are improperly validated, provide illusions of understanding (Messeri and Crockett 2024), or are even deliberately misused to create biased outputs and therefore flawed guidance (e.g., if models are reverse engineered from the desired outcome). Ecosystem modeling is a particularly interesting case study, given its importance for natural resource management decisions, which have profound effects on livelihoods and nature and which often incorporate a range of economic, social, and stewardship values. These values can often trade off against each other, and different stakeholder groups often hold different priorities. There is risk therefore of models being used inappropriately by stakeholder groups to support their specific agendas.

As with any transformative technology in recent memory, the deep integration of AI into ecosystem modeling is likely to pose numerous challenges, including ethical (Malakar and Lacey 2024) and legal considerations (Rodrigues 2020). The integration of AI into ecosystem modeling is a sociotechnical endeavor that necessitates a deeper engagement with the social rules, norms, and institutions that govern these technologies (Kudina and Van De Poel 2024). Currently there is a lack of clear, established protocols for developing and using AI for ecosystem modeling tools, and there is simultaneously a lack of guidelines to navigate the ethical, practical, and governance considerations of AI in ecosystem modeling. Therefore, we see a pressing need for strategic thinking and guidance on how to responsibly and effectively incorporate AI into the development and deployment of ecosystem models.

In the present article, we imagine how ecosystem modeling will look in a future in which the technical capabilities of AI in ecosystem modeling are unfettered, allowing AI to perform the role of an expert ecosystem modeler in supporting efforts to better understand, manage, and engage with ecological systems. We recognize that the AI systems of the future may be drastically different from the AI systems of today, and therefore, we aim to explore both the opportunities and the risks of transferring ecosystem modeling roles and responsibilities from humans to AI, without constraining ourselves to the limitations of current AI technologies. Although the extent to which AI systems will develop to be capable of modeling efforts is unknowable, as a community of scientists and decision-makers, we can use our knowledge and understanding of ecosystem modeling to articulate a vision for how AI should and should not be used, regardless of its capabilities. In this perspective, we aim to initiate a conversation on how to best incorporate AI tools into ecosystem modeling. By outlining an initial set of guiding principles and actions, we hope to support effective and responsible use of AI that, when guided by human knowledge and values, will support well-informed and transparent decisions for the sustainable and equitable management and conservation of natural resources and ecosystems. We also hope that this prospective work will provide insights that are applicable more broadly beyond ecosystem modeling, shedding light on the risks of, benefits of, and mitigation strategies for leveraging growing AI capabilities in system modeling across disciplines.

Our perspective

A growing recognition of the emerging capabilities of generative AI prompted a range of discussions and culminated in a 1.5 day expert workshop, held in Hobart, Australia, and online in June 2024. It brought together researchers from various disciplines, including AI developers, ecologists, ecosystem modelers, and social scientists, with the aim of predicting (Cook et al. 2014, Hobday et al. 2020) the future role of AI and humans in ecosystem modeling. The central goal of the workshop was to collectively envision a future in which AI tools were unconstrained in technical capabilities related to ecosystem modeling and capable of end-to-end automated model development and application (e.g., Szymanski et al. 2023). We wanted to take an intentionally speculative approach to explore the practical, ethical, and social implications of such a scenario and to identify some of the key considerations that differentiate a desirable future scenario from an undesirable one. We then sought to identify the role of tacit human knowledge and values needed to guide AI, including the strategies that could promote a desired future where AI supports human decision-making.

During the workshop, discussions in both breakout and plenary sessions, along with structured prompts (supplement S2), served as primary sources for the ideas presented in this article. The workshop itself provided several key observations that informed the present article’s narrative. It was acknowledged that the swift pace of AI development introduces a level of uncertainty about its capabilities, making it challenging to precisely forecast its application in ecosystem modeling. This uncertainty is not isolated to the field of modeling but extends to the broader implications of AI in research and societal contexts. Furthermore, the workshop highlighted that discussions about AI’s future role are often influenced by a spectrum of perspectives, ranging from optimistic, utopian visions to more cautious, dystopian views. This range of attitudes underscores the complex emotional and intellectual responses to the rapidly evolving landscape of AI technology.

Finally, the process of the workshop itself, which was marked by interdisciplinary collaboration and the synthesis of diverse viewpoints, demonstrates the value of inclusive dialogue in shaping the trajectory of AI in ecosystem modeling. It also emphasizes the importance of continual, multifaceted discussions to navigate the ethical implementation of AI and the necessity for broad-based AI literacy to ensure its responsible and beneficial use across society.

To illustrate the potential of an AI modeler, some short demonstrations of potential use were used to frame the initial conversation (see box 1). We used Microsoft Teams to transcribe the conversations and employed the large language models GPT4 Turbo and Claude 3.5 Sonnet to generate initial outlines, identify prominent themes, and summarize broad topics for this piece. Subsequently, the narrative underwent a systematic process of review and editing, with all authors contributing to and endorsing the final article.

Box 1. Agent-based modeling with generative AI.

On 26 June 2024, as an informal benchmark test and demonstration of current AI capabilities for ecosystem modeling, used as part of the introduction to the workshop, we used Claude Sonnet to rapidly generate a simple agent-based model of sheep and wolves interacting. The AI was prompted by the lead author as part of a live demo (the sequence of prompts and resulting HTML and JavaScript can be found in supplement S1; see also figure 1) to create this toy model based on the classic NetLogo model from the late 1990s (Wilensky 1997). Our intent was to illustrate the efficiency of generative AI in producing educational and preliminary research models quickly. This toy model, although it is basic, was created within minutes, a development task that would have traditionally taken considerably longer for a human modeler. Moreover, the general structure provided was of the correct form (as verified by expert ecological modelers at the workshop). The implementation was rudimentary and in need of further refinement; nonetheless, this initial form presents a rapid entry point for newer modelers, easing access, speeding capacity development, and potentially democratizing access to these kind of sophisticated modeling approaches.

The future of AI-enabled ecosystem modeling

Ecosystem modeling is a complex process requiring numerous steps, and engagement with knowledge and values from a range of disparate sources (figure 2a). Because of its complexity, it is vulnerable to mistakes across the workflow, which human modelers must be vigilant for. We imagine that an AI modeler will have the capacity to mitigate the chance that these errors occur and reduce risks associated with implementation, but we also imagine that, if it is poorly designed or poorly implemented, an AI modeler could easily exacerbate these risks. In the present article, we explore the transformative potential of AI in ecosystem modeling, focusing not only on the enhancement of modeling processes but also on broadening accessibility to these powerful tools and the numerous risks associated with its integration. We investigate how AI might change the landscape of ecosystem research and decision-support, from the processes of building models to the agents or actors that deploy them, and how the quality of the models might be assessed. Table 1 provides a consolidated summary of the key benefits, risks, and mitigation strategies identified across the article.

Figure 2.

The present and future of AI ecosystem modeling. (a) Present: the stylized role and capability (the distance from the origin) for humans to undertake modeling tasks, including innovation (the radial arrows), with limited support from AI. For a human (or AI) modeler, these tasks include the abilities to sift through research to pinpoint impactful study areas and objectives, identify human and computer resources needed, generate and refine ecosystem conceptual models, streamline data collection and synthesize domain insights, select algorithms and implement complex ecosystem frameworks, efficiently run simulations to optimize the model’s performance, conduct automated validation against real-world data, visualize and summarize findings for reporting, and consult with stakeholders throughout the modeling process to solicit feedback and field queries. (b) Future: as AI capability improves, it will be integrated across the modeling workflow. (i) It may become technically possible to automate existing ecosystem modeling workflows with AI, including having AI innovate on current modeling capabilities, however this may present unacceptable risks (ii) The division of roles that maximizes positive outcomes may involve specific roles for AI, humans, and their collaborations – with one or the other entity being excluded from some aspects of the work.

Table 1.

The benefits and risks of an AI modeler.

Theme	Benefits	Risks	Recommendations
Diversity	Facilitates diverse modeling approaches Encourages participatory modeling from a wider audience, leading to varied insights. Assists in creating multilingual models for broader accessibility.	Reinforces existing social and cultural biases (e.g., gender, race, socioeconomic status) when training data lacks representation from diverse groups. Converges models into a single form, reducing diversity of approaches.	Encourage open access and sharing of AI models and data to facilitate collaboration and transparency. Use AI to develop multiple models using different modeling approaches for the same system (allows exploration of what insights are robust to choice of modeling assumptions, and relative pros/cons of different modeling approaches)
Efficiency	Automates data assembly and preparation, software development, calibration, and report drafting. Shortens the distance between model conceptualization and implementation. Increases model development speed, allowing rapid prototyping of various approaches	Elicits an overreliance on AI, reducing critical thinking and problem-solving skills. Leads to loss of human-held detailed ecological understanding if AI is overtrusted. Potential dependence on AI for routine tasks could devalue human expertise. Difficult to establish credibility of AI-generated models	Create benchmarks to test AI abilities for ecosystem modeling to guide the deployment of AI and inform users about model trustworthiness. Mechanisms for human oversight that deliberately “slow down” AI and deliberately designed friction points for human review. Expanded checking and error routines initially, but also potentially training AI to look for these types or errors specifically. Expanding the interconnection of data and parameter collation and verification for deeper cross checking.
Knowledge	Outperforms humans at synthesizing data and reviewing complex codebases. Enhances “findability” of knowledge and data, improving model skill over time. Preserves and makes accessible domain knowledge that might be lost with retirements of senior researchers.	Loss of tacit knowledge and modeling skills among scientists and researchers. Risk of “epistemic instability” if AI-generated outputs are not distinguished from real-world information and knowledge in the development and calibration of models (e.g., risks of AI-generated data used in training). Loss of generational experience if older highly experienced modelers do not have the training to use the new tools	Contribute domain knowledge to provide AI with nuanced understanding and context. Develop methods of assessment and standards of use for ecosystem modeling with AI. Adopt procedures for auditing or recognizing AI generated data and insights. Accessible training support to help people to expand their modeling repertoire to not only use but contribute experiential knowledge to the AI mediated modeling approaches. Maintain human expertise and incorporate AI-derived methods.
Transparency	Offers explainable AI approaches with clear justifications for its choices. Readily available for interrogation, unlike human experts. Enables better understanding of causality in complex models.	AI models may become “black boxes,” difficult to interrogate or understand. Challenges in ensuring AI modeling processes are transparent and reproducible. Difficulties in replicating AI’s decision-making process for validation and scrutiny. AI innovations may be inscrutable to humans, impeding effective use of insights.	Develop clear and consistent standards for reporting AI usage in ecosystem modeling. Encourage the publication of models in standard and reproducible formats to enable critical evaluation of AI outputs.
Inclusivity/Equity	Democratizes ecosystem modeling by lowering barriers to entry for nonexperts. Translations make modeling more universally accessible, facilitating inclusivity. Promotes interactive learning and understanding across various skill levels.	AI-proposed solutions may not align with human values, causing mistrust. Misrepresentation or oversimplification of complex systems due to broad accessibility. Lack of modeling skill in users could generate convincing but inaccurate models, possibly leading to fragmented management solutions. Some users may lack the ability to critically evaluate AI models.	Develop infrastructure to support a range of stakeholders, including curated data and code repositories, validated model frameworks, and specialized AI tools that can be used safely by various expertise levels. Expand training to include school curriculum Establish robust monitoring mechanisms for AI use and model quality to balance democratization and control.
Reliability	Ability to conduct automated validation against real-world data. Enhances trust in modeling frameworks over time through improved modeling skill. Produces more consistent models “algorithmically” across modeling contexts.	Overreliance on AI could lead to underdeveloped human capacity to verify AI’s outputs. Issues around resilience to technological shocks Models will be biased by the data streams that go into them. Potential for small errors to go unnoticed AI might not handle “edge cases” AI could make up or misrepresent data and literature.	Identify and address edge cases in ecosystem models to create more resilient and reliable systems. Maintain and curate high-quality ecological data sets for training AI models. Maintain human involvement in oversight and assessment. Maintain, expand, and potentially automate conventional ways of evaluating model credibility
Innovation	Frees up human capacity for strategic and creative work. Supports the development of novel approaches and integration across domains. Innovates modeling approaches outside of its training data.	Uncertainty whether AI can truly innovate or if it is limited to existing data interpolation. May not be able to develop new problem-solving methods or modeling frameworks	Promote interdisciplinary collaboration to foster innovation. Establish processes for human–AI collaboration. Encourage exploration of edge cases as part of AI development to enhance problem-solving capabilities.
Ethics	AI can consistently apply ethical principles and societal norms, potentially reducing human bias AI can simulate a diversity of perspectives that may be difficult for any single modeler to access or represent, helping to surface underrepresented viewpoints and reduce individual bias.	Risk of malicious use or weaponization of AI. AI Could be used to create convincing but biased models that influence public opinion. Diversity of AI tools and capabilities could mean that ethics are not evenly applied. Devaluing of First Nations knowledge, data sovereignty and intellectual property.	Ensure AI integration is sensitive to ethical considerations such as stakeholder engagement, data sovereignty, and community empowerment. Involve social scientists and ethicists in the development of AI models.
Fitness for purpose	AI can tailor models to specific use cases, ensuring fit-for-purpose solutions, and rapidly explore a variety of approaches of differing complexity. AI can streamline model reuse by identifying when existing models are suitable, improving efficiency without sacrificing rigour.	AI may reinforce the reuse of models based on superficial performance metrics or biased training data, potentially perpetuating flawed assumptions and undermining model relevance in new contexts.	Develop methods to assess whether AI tools and models are appropriate for the specific requirements of ecosystem modeling tasks. Tailor AI solutions to the ecological context and ensure the models can address the targeted questions and problems with suitable precision and accuracy.

Figure 1.

Agent-based model simulation of sheep–wolf ecosystem dynamics. The left panel shows the spatial distribution of sheep (dots without bars) and wolves (dots with bars) with green bars indicating wolf energy levels. The right panel displays a time series plot of sheep and wolf population sizes over the simulation period.

Transforming how models are built

Having a highly capable AI assistant for ecosystem modelers would revolutionize the field’s potential for impact. The ability to automate many routine tasks that have clearly definable outcomes, such as finding data, software development, calibration, and report drafting, could reduce human errors, expedite model development, and allow alternate investment of human capacity. Rapid prototyping of models would mitigate the substantial costs of developing even simple models (Holden et al. 2024b), allow for the development of a diversity of models, which, either in isolation or in ensemble, could generate more robust insights about systems under uncertainty (Geary et al. 2020) and could allow for more effective identification of the appropriate model complexity to address given questions (Collie et al. 2016). Human researchers could then have more time and resources to develop skills across research domains, to engage with more diverse teams to enable diversity in problem-solving, to ask (and answer) more research questions, and to develop whole new approaches to conceptualizing and modeling ecosystems.

Faster model development enabled by AI could be especially important in time-sensitive environmental decision-making contexts. For example, the live demonstration model (box 1) illustrates how AI can rapidly generate a functioning initial agent-based model within minutes. In emergency scenarios, such as responding to a sudden ecological disturbance or planning under rapidly changing conditions, the ability to produce even a preliminary model quickly can mean the difference between informed action and proceeding without modeling support. Moreover, reducing development time from years to weeks aligns model delivery with the timeframes typically required by decision-makers, thereby increasing the likelihood of model uptake and impact.

Although there is current debate over AI’s ability to make groundbreaking leaps in approaches to modeling or to effectively handle edge cases (i.e., exceptions in data or implementation that do not conform to the expected norm and undermine performance), recent work suggests the potential for AI to transcend training data (Zhang et al. 2024) and outcompete human researchers in generating novel research ideas (Guo et al. 2025), suggesting that AI innovation within the realm of ecosystem modeling should not be dismissed out of hand. This will pose a unique challenge for the future as AI systems become more and more sophisticated, possibly eclipsing human expertise in certain areas and requiring substantial investigation to discern errors from novel discoveries (Silver et al. 2017). Similar to the peer-review process for developing ecosystem models, we foresee a need for humans to deploy specialized antagonistic AI tools, whose role is to critique and interrogate the AI tools that are used to build models (Sun et al. 2025). This red teaming (Perez et al. 2022) could potentially be applied to human-generated models to ensure best practices are adhered to (Jakeman et al. 2024).

Transforming who builds models

A particularly important potential impact is for highly skilled AI to democratize the insights of ecosystem modeling, by providing modeling capacity to people who have historically been unable to access the training or the resources needed to engage in the practice. This shift could bring about a new era of inclusivity in the scientific community, empowering a wider array of voices to contribute to environmental stewardship. AI will lower the barriers to entry by facilitating interactive learning for individuals of all skill levels, including those without programming skills (Gupta and Chen 2022, Cooper et al. 2024), by providing translations that make modeling more universally accessible or by undertaking end-to-end modeling exercises at the behest of users who are guided through the process of modeling by friendly and informed AI chatbots (in contrast to dense technical documentation and sometimes unfriendly or intimidating online fora). This change will be akin to the shift from static paper maps to having updated GIS maps in your pocket. Although things can still go wrong, especially if you use the tool unthinkingly, GIS has rapidly expanded the capacity for navigation, changed ownership hierarchies of knowledge and opened new use cases, business models, and job opportunities. This is particularly important because global environmental issues such as climate change, biodiversity loss, and overharvesting manifest differently at regional and local scales, requiring context-specific modeling expertise that is currently unevenly distributed.

AI’s role in encouraging learning and questioning is another vital aspect. On one hand, automating tedious tasks will allow more capacity to pursue more innovative and rewarding work, but on the other, we must explore whether we will be losing valuable learning and experience that comes from engaging with tedious work. For example, the process of calibrating and validating large and complex ecosystem models can be time consuming and mundane work, but through such a process, a human modeler can gain valuable insight into the dynamics of a system, how it functions, and its sensitivities. Conversely, it might be possible that this loss will be offset by the benefits of having bespoke AI learning companions or collaborators (Chen et al. 2024) who can rapidly bring the researchers up to speed on novel modeling techniques and tools. By enabling more people in more places to build and adapt models, AI could help decentralize modeling capacity and support more locally relevant applications of ecosystem science. To a certain extent, some of these processes have already been automated (Fournier et al. 2012), but having more interactive AIs may help calibrate the models while also sharing insights into the system dynamics that a modeler would have gained had they done the calibration themselves and, potentially, insights that they might not have gained through the manual process.

The democratization of model-building capabilities suggests a future where diverse stakeholders, from small communities to large corporations, are better equipped to direct their own modeling efforts and manage their data. This shift could enable various groups to find better environmental management strategies and engage more equitably in resource use negotiations. The custodianship of models and data by stakeholders might increase accountability, requiring them to consider the broader implications of their actions more thoroughly. However, the proliferation of modeling capabilities could lead to situations where differing approaches result in varying outcomes and policy recommendations, potentially increasing uncertainty and conflict. Moreover, it could result in misinterpretation of model capabilities and limitations by less-experienced users. In our complex, global environment where access to capital and technical expertise varies widely, the process of democratizing modeling tools raises important questions about data ownership, access, and the balance between local autonomy and collaborative efforts in model development and application.

Transforming the quality of models

Improving the quality of modeling insights hinges on three key factors: the skill of the AI-generated models, the quality of the data used to build them, and adherence to established modeling norms, such as documentation, calibration, and sensitivity and uncertainty analyses. Although AI has the potential to enhance model quality by automating parts of these processes, thereby increasing reliability and credibility over time, it also introduces new challenges.

One major challenge is how and whether to establish consistent workflows across diverse contexts that adhere to modeling best practices (Bennett et al. 2013, Hipsey et al. 2020, Jakeman et al. 2024). In theory, AI systems could be trained and designed to apply best practices in modeling. However, there is no guarantee that generalist AIs will reliably incorporate or enforce these standards. Current generative AI can produce convincing but superficial ecosystem models (see box 1), which may mislead users if the underlying data or literature has not been critically evaluated. This raises concerns about whether appropriate variables were measured, whether experimental methods were sound, and whether biases are present. Another key challenge will be to determine when to reuse existing models and when to develop new ones. Although reusing models can reduce costs and build credibility through continuity, it also risks perpetuating substandard or contextually inappropriate models. AI could play a critical role in this decision-making process by assessing model performance across different contexts, identifying mismatches between model assumptions and new applications, and flagging when a model’s reuse may compromise reliability. This capability could help balance efficiency with rigor, ensuring that model reuse supports rather than undermines modeling quality.

If they are successful, AI modeling techniques will no longer produce just a single model but potentially a plethora of viable and reasonable alternatives. This abundance introduces a new challenge: deciding which models are most credible, appropriate, and worth incorporating into decision-making. Although this diversity could enhance robustness by allowing comparison across models, it also increases the risk of selecting lower-quality or misleading models, especially when they appear superficially convincing.

Distinguishing between models will require more than just evaluating outputs. Ecosystem models (particularly mechanistic ones) depend not only on data but also on sound theoretical understanding and well-founded conceptual frameworks, which AI may not readily or reliably integrate. Without expert oversight, there is a risk that AI-generated models become black boxes that are difficult to interrogate, interpret, or reproduce (Rudin 2019). This opacity poses a particular challenge for understanding causality, which is essential not only for establishing model credibility but also for enabling scientific learning and decision-making. Even experienced modelers can overlook critical assumptions or make errors during model development, and the stakes are even higher when AI can rapidly generate numerous plausible models. Worse still, malicious actors could exploit AI to create biased but persuasive models (think deep fakes but for nature) that serve vested interests while appearing scientifically credible. This underscores the urgent need for clear, transparent criteria and protocols to guide model selection, taking into account the appropriate level of complexity, transparency, causality in the model and contextual fit. Without such safeguards, we risk mistaking plausibility for validity and efficiency for rigor (Essington and Plagányi 2014, Hamilton et al. 2022).

Seizing opportunities and mitigating risks

The powerful capabilities of generalist AI systems to perform the task of building (even simple) ecosystem models (box 1) suggests that its adoption into modeling frameworks is inevitable. However, the opportunities and risks that arise through this adoption, and the extent of adoption are yet to be determined. The workshop outcomes outlined above highlight that critical awareness and adaptability will be central to ensuring the ecosystem modeling community can capitalize on opportunities and mitigate risks associated with using AI (table 1).

Ideally, AIs incorporated in ecosystem modeling processes will, like expert human modelers, make well-reasoned decisions pertaining to the appropriate modeling framework and data for addressing the particular questions being posed. Assuring the efficacy of ecosystem models is a specialized task requiring a depth of understanding of the theory underpinning the models, the relationships between model components, and the transformations likely to occur to input data giving rise to outputs. Most researchers (let alone stakeholders) are unfamiliar with this kind of detail, which can lead to model misspecification, misplaced trust in model outputs, and inappropriate acceptance of misleading models in the literature. These problems already exist in peer-review of papers based on ecological and ecosystem models (Jakeman et al. 2024). Therefore, consideration of the problems of enhancing ecosystem modeling with AI can help highlight requirements for improving ecosystem models and the application of models in science generally.

For example, to ensure AI’s benefits are obtained while its risks are mitigated, we could implement safeguards and best practices that uphold principles such as diversity, efficiency, knowledge, transparency, inclusivity, reliability, innovation, and ethics while ensuring that these principles collectively support the positive aspects of AI integration. For example, existing standards such as the ODD (for overview, design concepts, and details) protocol for model description and the TRACE (for transparent and comprehensive model evaluation) framework for documenting modeling processes could be applied to AI-generated models (Grimm et al. 2020, Aylln et al. 2021).

Achieving a balance between the democratization of AI tools and the prevention of misuse requires careful governance, including community oversight, AI-specific peer review, and auditing standards. In addition, equitable access to AI tools must be ensured to avoid exclusivity to well-resourced entities. Lowering entry barriers presents an opportunity to build infrastructure such as data repositories, validated model frameworks, and training programs accessible to diverse users. The ecosystem modeling community should participate actively in shaping this future by contributing expertise, developing assessment methods, setting usage standards, and integrating AI education into relevant school and scientific curricula.

Contributing domain knowledge

Although it is possible for AI systems to incidentally pick up the best practices of ecosystem modeling through access to articles and content online, in our experience, there is a huge amount of tacit knowledge used by experts in the field, developed over years of experience, that will not be available to train AIs. We suggest, therefore, that the best way to ensure that the models that are built by AI comply with best practices of the ecosystem modeling community (e.g., Ayllón et al. 2021, Jakeman et al. 2024) and take into consideration the themes listed in table 1 will likely be for ecosystem modelers to actively provide AI with the nuanced understanding and context that it may miss from solely learning from written content. This transfer of knowledge may well form a new branch of research, because experienced modelers can struggle to articulate the tacit (and in some cases effectively subconscious) understanding they are drawing on in edge cases and other less common circumstances.

Today’s large language models, such as ChatGPT, are not well equipped to address the milieu of management decisions and, if they are used to inform model development, could naively exacerbate ongoing issues of data justice (Sworna et al. 2024, Urzedo et al. 2024). Human modelers, working in concert with social scientists and ethicists, are well placed to contribute their expertise in helping to resolve how an AI modeler would approach these considerations. There are suggestions that well-designed AIs themselves could enable the exposure to diverse perspectives (Li et al. 2024); however, we caution that techniques such as these should not be used as substitutes for ongoing, genuine engagement with stakeholders in the socioecological systems being modeled. That is, we acknowledge that the model development process as outlined in figure 2a, is not one-dimensional in a sense that stakeholder engagement is part of, for example, scoping, developing conceptual frameworks, and data gathering. On the basis of this argument, we envisage a future that is not necessarily as per figure 2bi but one where human oversight is still part of the mix of roles and capabilities in the human–AI interaction (figure 2bii). Another important need is for experts in ecosystem modeling to help lift the capability of the end users (e.g., stakeholders and decision-makers) to interpret AI generated model output for themselves, equipping them with the skills needed for at least a basic assessment of model integrity and reliability and the capacity to judge the usefulness of a model for their needs and whether further expert input is required. This could then go hand in glove with the kinds of benchmarks and standards discussed below. Furthermore, end users will need the ability to recognize that even well-constructed models may be unable to distinguish among management alternatives because of limitations in available data or inherent system variability. Addressing this requires that AI-supported models incorporate and communicate uncertainty through confidence bounds, sensitivity analyses, or scenario-based outputs so that decision-makers understand not just what a model predicts but how robust those predictions are.

Developing methods of assessment

Given that the range of ecosystem models and tasks is as diverse as the AI tools that can be used to support or undertake them, a rigorous program of evaluating and validating tools against modeling tasks will help in guiding future modelers on how to safely deploy AI and to understand the capabilities of each tool. One way we envision this being achieved is to build a robust suite of benchmarks explicitly designed to test AI abilities to perform ecosystem modeling. These benchmarks would provide an ongoing and dynamic understanding of the fitness for purpose of various AI tools and for understanding the limitations of an AI applied to each task. These benchmarks should not only assess technical performance but also contribute to establishing model credibility by evaluating transparency, robustness, and alignment with best practices in model development and documentation. This will be important both for providing clear guidance to modelers (AI and human) as to the best choices of AI but also to inform consumers of model outputs and insights as to the pedigree and potential trustworthiness of the model in question. Benchmarks should be designed with transparency in mind, allowing users to understand the decision-making process of AI and to evaluate its outputs critically. Furthermore, AI systems with built-in reflexivity and the capacity to report on model precision or accuracy on the fly would aid in avoiding the adoption of spurious insights from poorly specified modeling efforts.

Developing standards of use

The use of AI-developed models should be reported clearly and consistently, following standards agreed on by scientists and decision-makers regarding responsible AI development, data handling, ethical usage, and reporting. Other domains have started formulating such standards (Crossnohere et al. 2022), and there is a growing body of literature that can inform the development of guidelines for the domain of ecosystem modeling (Amershi et al. 2019, Jobin et al. 2019, Tuia et al. 2022). These guidelines should serve both the community of professional ecosystem , and the growing cohort of nonexperts who will be empowered to use AI tools to build their own models. The development of these guidelines would have the dual benefit of increasing the transparency and reliability of deployed ecosystem models regardless of whether they were developed by humans, AI, or a hybrid approach. In addition, the question of who will oversee the assessment of AI tools is critical. With large corporations often driving these technologies, accountability and governance become particularly challenging, mainly because of the potential prioritization of profit over transparency or fairness (Kloppenburg et al. 2022). The community must work toward establishing independent, unbiased mechanisms for AI tool assessment to ensure the integrity and reliability of ecosystem models.

Promoting open access and sharing AI models and data will facilitate collaboration and transparency within the scientific community. This openness will not only support collective advancement but will also build trust in AI technologies. Encouraging the publication of models in formats that allow for easy reproducibility is another important step. It ensures that other scientists can run, test, and validate the models themselves, contributing to the overall robustness of the research. Finally, there should be an explicit focus on edge cases in ecosystem modeling. For example, modeling the ecological impact of an invasive species suddenly introduced into a previously isolated ecosystem (e.g. cane toads in Australia) can test how well the model handles rare but disruptive biological events. By using AI to test and challenge these models, researchers can identify and address potential weaknesses or bias, leading to more resilient and reliable AI-built ecosystem models. Researchers will also need to maintain and curate high-quality repositories of ecological data, theory and knowledge on the processes driving ecological change that can be used to train AI models—similar to how AlphaFold used curated protein structure data for its groundbreaking work in protein folding prediction (Dessimoz and Thomas 2024).

Incorporating AI in how we learn to model

Fundamentally, the value of any modeling exercise is to synthesize and enhance understanding of our own (human) mental models of how nature works. To the extent that AI can facilitate the development of these mental models, maintaining a critical stance toward AI outputs and methodologies will be important to avoid cognitive traps and overreliance on the technology (Messeri and Crockett 2024). This may mean that people who interact with AI modeling systems need to maintain the same (or higher) levels of skepticism and trust as when working with human colleagues, no matter how certain an AI output might appear. This includes having a detailed understanding of the capacity of AI tools to perform the various tasks to which they are applied.

Training the next generation of professionals to more effectively use AI will ensure the continuity of skills and will foster the replicability and verification of AI-produced models and will require many of the same skills required for conventional modeling. The potential for specialized AI systems trained for specific tasks also underscores the ongoing need for human experts who can decompose larger, complex, or ill-defined tasks into smaller, manageable components that the implemented AI systems can address more effectively.

As AI becomes increasingly integrated into ecosystem modeling, updating curricula to train the next generation of modelers in AI methods will also help to encourage that these tools are implemented responsibly. These educational initiatives must emphasize critical thinking and model interrogation skills, as the ability to evaluate AI outputs and understand their limitations becomes paramount. Developing training programs that focus on working collaboratively with AI tools, teaching researchers how to effectively decompose modeling tasks for AI assistance, and how to integrate AI outputs with human expertise are essential.

Inevitable limitations of AI modelers

Let us now assume that we can seize all the opportunities and mitigate all the risks that we have just described and any others we haven’t yet been able to imagine. What role is left for the human? Although it is possible to envision a future in which an AI system becomes so advanced that it becomes technically possible to entrust it with some aspects of environmental decision-making, we would caution against such an approach. No matter how technically capable AI systems become in the future, they will still have several insurmountable limitations that will prevent complete automation and will require human guidance and decision-making at various levels of the modeling workflow. Ecosystem modeling is embedded in a complex social context (DeAngelis et al. 2021); therefore, there are some key issues to consider.

Whose AI to use?

Even if difficult questions of AI alignment are eventually resolved at some levels (Gabriel 2020) and meaningful mechanisms for control are implemented (De Sio et al. 2022), the questions of who builds, designs, controls, and maintains the AI systems will remain. The values and biases of developers can inadvertently be reflected in the AI (Roselli et al. 2019), leading to a potential cultural hegemony that may not be representative of all stakeholders involved in ecosystem modeling. In today’s context, where only a handful of private companies are leading the development of cutting-edge AI systems, there is uncertainty as to who will have control over powerful AI systems and how widespread their availability will be, raising questions about equity, the democratization of the benefits of the technology, and the resilience of the system.

AIs are not stakeholders

AI systems, regardless of their sophistication, cannot access the full spectrum of human intuition, experience, and knowledge. Traditional cultural knowledge, individual human values, and privacy concerns are examples of important considerations that may elude AI understanding. These gaps necessitate active human involvement to provide consent and cocreate knowledge, ensuring that AI models are not only technically sound but also culturally sensitive and ethically grounded according to those who would employ and be affected by them. The limitations of AI in capturing the nuance and depth of human knowledge underscore the need for human-centered design and usage of AI in ecosystem modeling, prioritizing human input throughout the process (Bondi et al. 2021). There is a need for strong systems, procedures, and norms, governed by humans, around how AI is used to facilitate ecosystem modeling.

AIs cannot be responsible

Accountability in decision-making is a fundamental principle that AI, as a nonsentient entity, cannot fulfill. No matter how advanced or persuasive an AI system may appear, it remains a tool to assist human work. Humans can use AI to generate insights and support decision-making but must ultimately accept responsibility for the outcomes of those decisions (Schleiger et al. 2024). The responsibility cannot be transferred to the AI, because it lacks the capacity for moral judgment and accountability. This underscores the necessity for robust human governance systems to oversee the use of AI in ecosystem modeling, ensuring that decisions are made responsibly and transparently.

The need for human-centered ecosystem modeling

As we navigate the rapid evolution of AI in ecosystem modeling, the imperative for a human-centered approach becomes increasingly clear. This approach is essential to ensure that our models remain not only trustworthy and effective but also firmly aligned with human values, scientific integrity, and the complex realities of ecosystems. The swift progress of AI technologies presents a unique challenge: We must balance technological innovation with rigorous scientific standards and ethical considerations, all while maintaining the irreplaceable role of human expertise and judgment. To address this challenge, guidelines are needed for both expert and nonexpert users of AI modeling tools that adhere to scientific norms (Blau et al. 2024) and that evolve alongside AI advancements.

The integration of AI into ecosystem modeling is not just likely but inevitable and imminent, but how this is done remains to be seen. This integration brings with it the critical responsibility to ensure that AI serves as an amplifier of human expertise, not a replacement for it. However, we must also recognize that there may be scenarios where AI’s capabilities could lead to better or faster outcomes than human intervention alone. Drawing parallels from other industries further along in AI integration, such as aviation and power grid management, we can anticipate instances where AI might need to take a more active role in decision-making processes, particularly when rapid responses are crucial. In the realm of ecosystem modeling, AI integration may provide an essential capability for building increasingly complex models to address large-scale issues while maintaining known low-level ecological mechanisms. Nevertheless, despite its convincing capabilities, AI lacks personhood, moral responsibility, and the ability to comprehend the full consequences of its outputs. Therefore, the onus falls squarely on us, the designers and users of AI and AI-assisted modeling tools, to claim ownership of the models we deploy and the decisions they inform. In cases where AI takes a more active role, human oversight becomes even more important and should focus on setting boundaries, defining ethical guidelines, and maintaining a bigger-picture perspective that ensures the AI actions align with the broader ecological and societal goals. This responsibility extends beyond the scientific community to include private entities and local communities seeking to make informed decisions about ecosystem management. In this context, the field of interaction design, which studies how to optimize interactions between humans and products, plays a crucial role in shaping how humans and AI collaborate effectively. As Borthwick and colleagues (2022) argued, interaction design is uniquely positioned to contribute to the conservation and restoration of the Earth’s biosphere by creating interfaces and processes that facilitate meaningful human oversight and decision-making.

As AI models become more complex and numerous, the challenge for human modelers to maintain necessary oversight will intensify: People will still need to keep a line of sight on the detailed mechanisms of the model and also verify that any AI-generated software faithfully adheres to the model. We envision AI playing a critical role in supporting but not supplanting this oversight. AI can swiftly analyze model components, identify potential inconsistencies, and uncover flaws that might elude unaided human scrutiny. For instance, AI could rapidly compare multiple ecosystem models, highlighting discrepancies in predictions or assumptions that warrant further investigation by human experts.

Human-centered design remains essential in developing AI software, models, interfaces, and processes for ecosystem modeling. This approach ensures that AI systems remain comprehensible, accountable, and ultimately controllable by human operators. It acts as a safeguard against the potential pitfalls of fully automated, inscrutable decision-making processes that could have far-reaching consequences for both ecosystems and human societies. Implementing human-centered design in ecosystem modeling might include creating intuitive visualizations of AI model outputs, developing explainable AI techniques that allow users to understand the reasoning behind model predictions, and establishing clear protocols for human intervention and override in AI-assisted decision-making processes.

In conclusion, the roles we have discussed throughout this article highlight the symbiotic relationship between AI’s computational prowess and human ingenuity. As we progress further into the AI era, we must remain open to the possibility that AI may sometimes outperform humans in specific tasks or decision-making processes. However, this does not diminish the importance of human involvement. Instead, it shifts our role toward strategic oversight, ethical guidance, and big-picture thinking. In the face of AI’s relentless advancement, we must commit to a human-centered methodology, especially in the delicate realm of ecosystem management, where the stakes of environmental integrity and societal well-being are exceptionally high. It is our collective responsibility to guide AI’s application with foresight, ethics, and a deep respect for the natural world we strive to understand and conserve. To operationalize this commitment, we propose the following guiding principles for integrating AI into ecosystem modeling responsibly and effectively: (i) maintain human oversight by implementing mechanisms for human review and responsibility in AI-assisted ecosystem modeling decisions, (ii) ensure transparency by developing clear standards for reporting AI use and encourage publication of reproducible models to enable critical evaluation, (iii) promote modeling diversity by using AI to develop multiple modeling approaches for the same system to explore robust insights and mitigate bias, (iv) democratize with safeguards by creating infrastructure supporting diverse stakeholders while establishing monitoring mechanisms to balance accessibility and quality control, (v) integrate ethics by involving social scientists and ethicists in AI model development to address ethical considerations and ensure alignment with human values, and (vi) enhance AI literacy by updating educational curricula and foster interdisciplinary collaboration to promote responsible AI use in ecosystem modeling.

Author Biography

Scott Spillias (scott.spillias@csiro.au), Rowan Trebilco, Fabio Boschetti, Piers Dunstan, Javier Porobic, Pascal Hirsch, Alistair J. Hobday, Jess Melbourne-Thomas, Karen Wild-Allen, Skipton N. C. Woolley, and Elizabeth A. Fulton are affiliated with CSIRO Environment, in Hobart, Tasmania, in Australia. Scott Spillias, Rowan Trebilco, Fabio Boschetti, Javier Porobic, Alistair J. Hobday, Jess Melbourne-Thomas, Eva Plaganyi, Cara Stitzlein, and Elizabeth A. Fulton are affiliated with the Centre for Marine Socioecology, University of Tasmania, Hobart, Tasmania, Australia. Matthew P. Adams is affiliated with the School of Mathematical Sciences and the Centre for Data Science, at the Queensland University of Technology, Securing Antarctica’s Environmental Future, in Brisbane, Queensland, Australia. Andrew Constable is affiliated with the University of Tasmania, Hobart, Tasmania, Australia. Einat Grimberg, Sarah Kaur, Cécile Paris, Cara Stitzlein, Viveka Weiley are affiliated with CSIRO Data61, in Eveleigh, New South Wales, Australia. Nicola Grigg and Simon Ferrier are affiliated with CSIRO Environment, in Canberra, Australian Capital Territory, Australia. Mike Harfoot is affiliated with Vizzuality, in Madrid, Spain. Matthew Holden is affiliated with the University of Queensland, in Brisbane, Queensland, Australia. Trevor Hutton, Denham Parker, Éva Plagányi, and Jacob G. D. Rogers are affiliated with CSIRO Environment, in Brisbane, Queensland, Australia. Viveka Weiley is affiliated with the CSIRO Concept Lab, in Sydney, New South Wales, Australia