Conclusions and recommendations

The planet on which we live is undergoing rapid changes. The conversion of land to agricultural and urban areas, the extraction of natural resources, the loss of biodiversity, and rising carbon emissions are altering planetary processes on land and in the oceans in ways that increase risks to people all over the world. Novel digital technologies and advancements in artificial intelligence (AI) have profoundly reshaped social interactions, resulting in a more connected global population, but simultaneously creating a diversity of novel risks.

Science plays a key role in this rapidly changing context. It provides evidence of how planetary changes could affect people’s lives, offers analyses for making informed decisions to mitigate risks and reduce social vulnerabilities, and spurs the innovation needed to drive social transformations. At its best, science can guide decisions that help societies move toward a safe future for all. Exploring such pathways is far from straightforward, however. Analyzing the multiple and interacting complexities of human behavior, ecosystems, a changing climate, and technological change remains a considerable research challenge.

While AI is already driving breakthroughs in fields from the life sciences to mathematics, its role in sustainability research is currently less established. This raises a critical question: could AI be effectively applied to the complex, interconnected challenges of sustainability?

This concluding chapter is structured in three parts. The first part summarizes an internal survey where we assess and rank the different issue areas according to their present level of AI maturity (i.e., access to data, and application of AI methods), as well as their wider accessibility. The second part summarizes high-level insights from our analysis. The third and last part ends with a list of recommendations to colleagues in academia, to governments, the public sector, entrepreneurs, and to philanthropic funders.

Comparing the Issue Areas

The use of AI methods differs considerably between the eight chosen issue areas, as shown by our literature overview and deeper analyses of each area (fig. 4a-h). In general, classical machine learning methods are more commonly used in areas exploring complex interactions in climate and nature (e.g., Understanding a Complex Earth System, Securing Freshwater for All, and Stewarding Our Blue Planet), while newer AI methods are more frequent in areas where stakeholder participation and interactions are more prevalent (e.g., Prospering on an Urban Planet, Improving Sustainability Science Communication, and Collective Decisions for a Planet Under Pressure).

How do these issue areas compare in terms of their AI capabilities, and how accessible are insights derived from such analyses to actors outside of academia? Which areas seem to have made the fastest progress? And which ones are in need of a stronger push to be able to fulfill their potential? These questions all relate to comparisons between the issue areas, thus offering insights into different levels of maturity and availability of interest to both researchers and funders.

We bring together data from internal expert assessments of these three dimensions—data availability and quality; AI model performance; and the accessibility of AI models and insights to non-expert communities (e.g., NGOs, civil society, local communities). The assessment builds on an online survey sent out to each contributing author of this report in the final part of the writing. The literature overview and review of each issue area thus informed the individual assessments. Between 15 and 23 experts contributed depending on the issue area, forming the joint assessment (from here on referred to as “the assessment”) seen in Figure 8.*

The assessment shows several notable findings. Some issue areas stand out as more mature, characterized by relatively high scores along the dimensions of AI-model performance, and data availability, and quality. This grouping includes the areas Understanding a Complex Earth System, Sustainability Science Communication, Securing Freshwater, and Urban Planet. In contrast, there are issue areas where data availability is comparatively sparse, and where AI-methods development is viewed to be less advanced. This includes the area Collective Decisions for a Planet under Pressure in particular, which scored lowest on both dimensions (see Figure 8). These areas are likely to require a step-change increase ininvestment for research and applications (including data infrastructure, compute and expertise) to enable innovation with AI-methods at scale.

Advancements in Large Language Models (LLMs), along with their increasing accessibility to the public may, we believe, have increased the availability of AI for science communication. This trend aligns with the relatively high accessibility grade given to the Sustainability Science Communication issue area. It is noteworthy, however, that two crucial issue areas Preparing for a Future of Interconnected Shocks and Understanding a Complex Earth System were graded to provide low accessibility to users outside of academia, such as NGOs, civil society, and local communities. These latter areas thus offer an opportunity for funders to invest in ways that help bridge the gap between ongoing AI research for sustainability, and its uses by society. Increased collaborations across domains and targeted research initiatives may drive scientific advances and innovative uses of AI across, and between, issue areas. In summary, there is a need to bring more collaboration into the community, including knowledge sharing, to ensure that investment is flowing to the areas of biggest impact, especially where critical issues often lie at the intersection of different domains.

Key insights

1. AI offers vast potential across the sustainability sciences

Recent advancements in AI, combined with increasingly powerful computing capabilities and methodologies, are enabling the discovery of patterns and relationships within vast datasets from various fields of science, leading to intriguing scientific discoveries. This capacity for data integration, generation, and analysis across global and local sustainability domains allows for insights that were previously too complex, or too interdisciplinary, to tackle. This capability extends to diverse data sources enabling automated feature extraction in realms ranging from marine acoustics and species identification, to water resources monitoring and urban climate risk mapping (see Issue Areas). Hybrid simulation-AI models, which combine AI algorithms with process-based models, are proving particularly powerful, with reduced computational cost compared to traditional methods (p. 16, 19, 34, Theme box 3). There are, however, large differences in capabilities between the different analyzed issue areas (p. 65-67).

2. AI can inform decisions and help communicate complex sustainability challenges

AI can bolster decision-making and science communication for sustainability. Examples include AI-augmented analysis supporting marine protected area (MPA) management (p. 39), water resource planning and management (p. 41-43), and the development of early warning systems for hazards and conflicts (p. 27-29). Large language models (LLMs) and multimodal language models (MLLMs) can help personalize, localize, and make complex scientific information more accessible, engaging, interactive, and actionable for diverse audiences (p. 57).

3. Environmental footprint, biases, hallucinations, and unequal access to AI pose sustainability risks

While AI as a research method for sustainability holds important potential, the increasing use of AI in society in general, particularly for generative AI, demands energy and freshwater across the whole supply chain, leading to increased CO₂ emissions and e-waste, and often impacting already vulnerable communities. This growing footprint, including toxic substances from e-waste, and water withdrawals for energy production and/or cooling in water-scarce regions, represents a sustainability risk and climate justice issue on its own, thus requiring serious consideration and mitigation (p. 10-11). Realizing AI’s potential for sustainability will require overcoming issues of data scarcity, quality, and geographical imbalance, which are particularly acute in low-income countries (p. 12). Underrepresentation of low- and middle-income countries in AI research, design, and use may perpetuate existing inequalities. Algorithmic biases and “hallucinations” introduce substantial ethical and accuracy challenges, and can lead to allocative harms that disproportionately affect vulnerable groups (p. 11-12). Interpretability and vast computational demands are major hurdles for AI adoption and trust. Complex deep learning architectures until now often function as “black boxes,” making it difficult to understand the causal mechanisms behind AI predictions and acting as a barrier to trust for decision-makers and users (p. 11). Training and optimizing these models, especially foundation models, require vast computational resources, creating access and usability challenges, but also opportunities once these models are made accessible (see Theme Box 1).

4. Responsible uses of AI for sustainability research are possible

As all individual chapters illustrate, addressing risks associated with uses of AI for sustainability research is possible. Responsible AI development for sustainability requires fostering interdisciplinary collaboration; augmenting rather than replacing human expertise; supporting academic freedom, peer review, and open science; and establishing robust governance mechanisms addressing AI’s limitations proactively. Several pioneering private, science, and civil society-led initiatives have emerged in this domain. This includes, for example, tools and methods for the development of responsible foundation models; consensus checklists to help advance valid, reproducible, and generalizable AI-based science; and Indigenous Data Sovereignty principles. The co-production of knowledge and benefit-sharing with local communities, decision-makers, and others offer ways to empower and augment human expertise, requiring careful selection of problems and rigorous evaluation.

5. Breakthroughs in sustainability research driven by AI are within our reach and essential for our collective future

Current applications of AI for sustainability-related research are extensive and diverse, and show that an “AI for Sustainability Science” community has begun to emerge.

Its diversity in focus and expertise (see Issue Area chapters), as well as in different levels of AI maturity (p. 65-66), can be viewed as a source of innovation. The potential is substantial, and could transform existing areas of sustainability science, and open up new research areas and practical applications. The growth in public and private investments in AI hardware, AI research, and “AI for Science” in general, however, will not drive scientific breakthroughs of relevance for sustainability at the pace and scale needed. Lack of technical expertise, bottlenecks in access to AI compute, and limited collaborations between AI scientists, developers, and sustainability researchers create obstacles that limit the current potential of AI. Supporting cross-domains collaborations; scaling up AI funding for integrated biodiversity, climate, and social-ecological research; developing clear guidelines for responsible AI development and use for sustainability; and building ambitious international research programs across domains could lead to important advances that benefit people and the planet.