Ethical Use of AI in Earth and Environmental Sciences: Principles and Challenges

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Nov 3 2023

In a recent publication in the journal Nature, researchers explored a set of principles and guidelines for the use of artificial intelligence (AI) and machine learning (ML) tools in the fields of earth, Space, and environmental sciences.  

Background

Within Earth, Space, and environmental sciences, a spectrum of technologies, ranging from sensors to satellites, furnishes intricate insights into the planet, its ecosystems, and its chronological evolution, spanning all spatial dimensions. The pervasive adoption of AI tools is increasingly conspicuous, extending to domains such as weather forecasting and climate modeling, energy and water resource management, and expedited disaster assessment to facilitate swifter relief efforts and reconstruction.

The American Geophysical Union (AGU), in collaboration with NASA and a community of ethicists and researchers, including the present authors, has undertaken the objective of formulating a compendium of principles and guidelines governing the application of AI and ML tools in the realm of Earth, space, and environmental sciences.

Six principles to foster trust

Diligent adherence to these best practices plays a pivotal role in averting harm when employing AI in pursuing scientific inquiry. The first four principles for researchers and those remaining for scholarly organizations are:

• Transparency: meticulously document and disclose details pertaining to participants, datasets, models, inherent biases, and uncertainties.
• Intentionality: ensuring that the AI model and its implementations are elucidated, replicable, and amenable to reutilization.
• Risk Mitigation: Consider and manage potential risks and biases to which datasets and algorithms are susceptible, as well as their potential influence on results or unintended consequences.
• Inclusive Approaches: Strive for inclusive research design, engage with vulnerable communities, and enlist domain-specific expertise.
• Outreach, Training, and Exemplary Practices: Extend educational resources and support to all stakeholders and career stages.
• Ongoing Commitment: Implement, scrutinize, and advance these directives.

The landscape of responses will inevitably evolve as AI continues to advance, yet the bedrock principles will remain grounded in the cardinal tenets of sound scientific methodology, encompassing the collection, treatment, and utilization of data.

Addressing biases and risks in AI-driven research and policy integration

In the context of AI-driven models, researchers must possess a profound understanding of training and input datasets. This includes meticulous scrutiny of inherent biases, especially when the model's outputs impact consequential decisions, such as disaster responses, investments, or healthcare choices. Inadequately described datasets heighten the risk of producing studies that perpetuate biases, falling into the "garbage in, garbage out" category, rendering outcomes either meaningless or perilous.

Environmental data exhibit disparities in coverage and fidelity across regions and communities. Areas with frequent cloud cover, such as tropical rainforests, or those lacking in situ sensors and satellite coverage, such as polar regions, are inadequately represented. Data abundance and quality often unintentionally favor affluent areas and populations while marginalizing vulnerable or historically discriminated communities.

The amalgamation of data sources, vital for offering guidance to the public, policymakers, and businesses, can exacerbate these issues. Unintended repercussions may occur when sensitive information is inadvertently disclosed. The proliferation of diverse datasets elevates the risk of adversarial attacks, which can corrupt or degrade data without researchers' awareness.

Malicious or erroneous use of AI and ML tools can distort model outputs and undermine conclusions. The interplay between AI and ML models amplifies their impact and magnifies risks through error propagation. To mitigate these risks, adherence to data deposition recommendations is advisable.

Institutions should equip researchers with the skills to scrutinize data and models for spurious results, considering environmental justice, social equity, and implications for sovereign nations. Institutional review boards should oversee AI models' integration into policy decisions. By focusing on these critical aspects, researchers can navigate the challenges of AI-driven research and contribute to more equitable and reliable outcomes.

Transparency initiatives in AI-driven research

In the realm of scholarly research, a divide exists between classical and AI-driven models. While classical models require code access, AI models lack clear protocols for elucidating limitations. This opacity stems from AI's inherent inscrutability and lack of transparency. Understanding results, assessing uncertainty, and reconciling model disparities are challenges. Machine learning's adaptability leads to variability, even with identical data and algorithms.

In scholarly publications, researchers must elucidate AI models for external evaluation. Cross-model comparisons and data segregation are invaluable for robustness. Clear standards for explaining AI models, akin to statistical confidence levels, are needed. Efforts to enhance transparency through explainable AI (XAI) aim to make AI more comprehensible. XAI aids interpretation in contexts like short-term weather forecasting and Earth sciences.

To boost transparency, data and code sharing, replicability testing, risk and bias mitigation, and uncertainty reporting are crucial. Collaboration with data specialists and affected communities is key. Adoption of FAIR guidelines emphasizes data accessibility and reusability. Discipline-specific repositories, supported by generalist ones, maintain data quality. Scholarly organizations promote ethical standards. Financial investment ensures repository viability without compromising research funding.

Conclusion

In summary, researchers explored AI tools in the Earth and environmental sciences. They discussed the six principles for AI use in these sciences, focusing on transparency, intentionality, risk mitigation, inclusivity, outreach, and ongoing commitment. The current study highlights that addressing biases and risks in AI-driven research is crucial, as inadequate dataset descriptions may perpetuate biases. Environmental data disparities pose challenges, and malicious AI use can distort results. Transparency initiatives, like XAI, aim to enhance AI comprehensibility by emphasizing data sharing and ethical standards in research.