Ontologies, generative AI and the SDGs

Co-authors: Mala Kumar, Annie Brown, Anthony Ware, Lizzette Soria

This is a guest blog co-authored by Humane Intelligence and its volunteers, exploring topics related to AI evaluations and sociotechnical topics in AI.

The Initial Transformative Idea

In the late 1990s and early 2000s, experts in the United Nations (UN), academia and across international civil society came together and made a decision to, for the first time, set international quantitative standards of human development. This global exercise morphed into the Millennium Development Goals (MDGs), which outlined specific, numerical measures to define often nebulous and ambiguous ideas for things like public health, education, gender rights, and food security. The MDGs launched in 2005, covering the period of 2000 – 2015.

Fast forward ten years, and in 2015, the Sustainable Development Goals (SDGs) were launched as a second iteration of the MDGs, a main difference being the addition of climate and environment related goals. Both the MDGs and the SDGs were and are meant to be measured at the national administrative level in low- and middle-income countries, and both had the same basic structure:

Much progress was achieved with the introduction of the MDGs and the continuation of the SDGs, showing the power of international coordination and cooperation. However, despite the majority of international humanitarian assistance and international development funding to LMICs being reorganized to align to the MDGs / SDGs, most countries didn’t achieve most MDGs, and are not on track to achieve most SDGs by the 2030 end date. As we come closer to the end of the SDGs period, and the world must now contend with massive funding shortages in international assistance, we should ask the question: how can generative AI help us decide what’s next?

A New Transformational Idea

One of the major gaps with the SDG structure is that it’s presented as a taxonomy, which is a hierarchical classification scheme that groups items into categories, like organizing animals into kingdom, phylum, class, order, etc. Taxonomies can reinforce siloed analysis and rigid categorization. Since the SDGs are presented as a taxonomy, the relationships among goals are not implied, discussed or mapped.

A common and strong use case of generative AI is to synthesize complex, messy and multi-dimensional relationships in a way that any lay person without a high degree of technical or statistical knowledge can understand. With the responsible use of generative AI, what comes after the SDGs could thus be an ontology of human development goals. Ontologies provide a more expressive alternative by encoding relationships between concepts. This allows development challenges to be modeled as interconnected systems rather than discrete reporting categories, and creates a more flexible framework for understanding how interventions in one domain propagate effects across others.

At the foundation of many AI and machine learning systems are taxonomies and ontologies: the classification systems that structure how data is organized and understood. An ontology goes further than a taxonomy; it not only classifies but defines the relationships between concepts, the properties they have, and the rules that govern them.

In AI/ML systems, these ontological frameworks can determine what data gets collected, how it’s labeled and categorized, and ultimately what patterns the models can recognize. For example, a health monitoring system’s ontology shapes whether it recognizes “wellness” as merely the absence of disease or as a holistic concept encompassing mental, social, and physical health. These design choices profoundly affect what the AI can “see” and how it interprets the world.

With an human development ontology, it may be possible to answer questions such as, 

If Country X were to reduce the rate of water-borne diseases by 30%, what educational gains would they achieve?

Or

To achieve gender parity in a Country X’s education system, what investments must be made in their food security and public health?

Some efforts have developed ontologies with draft definitions and relationships, such as UNEP’s SDG Interface Ontology, SDGIO, which is “focused on specifications of terms and their interrelations in the domain of the 2030 Agenda”. In the private sector, MIcrosoft Research also published a paper on all SDGs supporting SDG 3: Good Health and Wellbeing. So the idea of the next iteration of the SDGs changing from a static taxonomy to a dynamic ontology is not completely new. What then are the challenges and opportunities to use AI to advance the idea?

The Opportunities

Community-Led Ontology Design

While the MDGs and SDGs were innovative in standardizing global metrics for human development, they lacked the flexibility to account for local contexts. Ontologies offer an alternative: they can maintain a canonical reference framework while supporting multiple context-specific iterations tailored to different communities and settings.

Community-led ontology design provides a pathway to bring impacted communities directly into the AI/ML development pipeline. Drawing on approaches like Humane Intelligence’s AI/ML evaluation work, this method uses human-in-the-loop (HITL) processes to engage communities in identifying how AI models should be adapted for their contexts. Communities can help surface model vulnerabilities, determine which categories and concepts matter in their setting, and ensure that data is collected, classified, and described in culturally appropriate and meaningful ways.

This participatory approach addresses key limitations of the MDGs and SDGs by making the design process transparent, inclusive, and responsive to local knowledge. It also builds essential community buy-in—ensuring that the frameworks actually serve the people they’re meant to represent.

Advent Of New Tech And New Understandings

Given the complex nature of human development, the relationships among SDGs and its targets are often not understood until a broad segment of a population is affected. During COVID, different countries instituted varying levels of learn- and work-from-home mandates. At the global scale and in subsequent years, it became apparent that these mandates had a profound and sometimes irreparable effect on the learning outcomes of young students. In some cases, the mandates were likely too strict given the negative impact on student learning outcomes. In other cases, the mandates were too relaxed and resulted in more COVID-related deaths with minimal student learning gains.

How strict to make remote learning mandates for varying school-age populations was a constant debate. How to account for socioeconomic status, access to digital infrastructure, language, age, gender, and a multitude of other factors was extremely challenging, including in resource-poor areas of LMICs. Generative AI and the SDGs as an ontology may have provided key insights that were otherwise impossible to determine.

Data Interoperability

In the past, data interoperability has been a core challenge for achieving the SDGs. Over time, ontology-driven approaches combined with generative AI may reduce the resource barriers associated with data collection, synthesis, and system development. Many countries face persistent constraints in SDG monitoring: legacy statistical infrastructure, limited analytical capacity, and restricted budgets. An ontology that supports natural language queries, enables cross-domain analysis, and accommodates dynamic updates as new evidence emerges could lower these costs and extend monitoring capabilities to a broader set of institutions, including national statistical offices, civil society organizations, and community-based initiatives.

“In most OECD countries, over 40% of SDG data points are more than three years old.” [Mind the SDG data gaps, OECD 2025]. Further, “geographic coverage, timeliness and disaggregation remain areas of concern.” [IAEG-SDGs, 2025]. Ontology based generative AI solutions could add more mileage to data. Even if one data point is outdated, other data that relate to it may be more current. Unstructured but relevant data at the community level could reveal interdependencies in SDGs within a country and cross-border. 

Data interoperability in a community-led ontology system would need new standards and attributes to unlock the potential upside. From governance to data standards, the system would need to be adaptable. This opens up another opportunity for shifting how the new data needs to be collected and ways it can be collected with the available local resources rather than external funding and aid, which has diminished. 

The Challenges

Local Capacity

A hallmark challenge in international development is local capacity, especially in technical roles. This matters even more if the next iteration of SDG measurement shifts from static indicators to dynamic ontologies supported by generative AI. With proper training and well-designed ontologies, AI could help countries address complex social problems, generate roadmaps for progress, and identify gaps or bottlenecks across goals. But these outcomes are not guaranteed. They depend on whether the people closest to the problems are able to shape the underlying structure of what the system measures and how it reasons.

Community participation is essential, but not always easy. Real investments are required to build and sustain contributor pools with lived experience, including training, compensation, quality support, and long-term coordination. Without these, ontology design risks becoming an academic exercise that looks complete on paper but fails in practice. In many cases, researchers and professionals will miss key concepts, relationships, and edge cases that determine what outcomes look like on the ground. Lived experience surfaces informal systems, cultural norms, local definitions of success and harm, and the constraints people navigate every day, all of which are often invisible in national statistics or global frameworks.

Degrees and technical credentials matter, but they are not sufficient for representing human development realities. The most valuable SDG knowledge is often held by frontline practitioners, community leaders, and affected populations who understand the tradeoffs and context that drive real-world outcomes. Ontologies that treat this knowledge as secondary will be structurally incomplete, and the AI systems built on top of them will be less accurate, less trusted, and less useful.

Finally, local capacity is not only about people. It is also about tools. Many AI and data systems are designed for data scientists and technical teams, which creates a barrier to participation from non-technical stakeholders. If ontologies are meant to support better development decision making, the infrastructure must be accessible to anyone who has relevant expertise, not just those with the technical training and resources that are disproportionately concentrated in high-income countries. In practice, the strongest approach will likely involve mixed local teams of technologists and non-technical domain experts, supported by translators who can bridge technical concepts and local cultural realities.

Responsible Design and Use

With varying amounts of data of varying quality and complexity, using generative AI to define a human development ontology will have varying factual accuracy, usefulness, and interpretability context-to-context. Countries that have long historic records of highly accurate, machine readable, structured data for a given SDG, target or indicator, will likely have better results than countries that don’t. There are no quick fixes to these structural data issues, which makes the process and transparency of a generative AI-powered ontology both difficult and important. Among those design decisions are who and how to decide what data is included, what is an acceptable output, and how to mitigate harmful behavior. 

The promise that AI can help solve some of the world’s hardest problems is within reach, but it will not be achieved through today’s default approach: scaling models without equal investment in the data foundations, governance, and human systems that make them trustworthy in real-world settings. In development contexts, the goal is not to produce answers faster, it is to produce answers that communities can rely on, contest when needed, and adapt over time. That means treating the ontology itself as critical infrastructure: designed transparently, maintained responsibly, and shaped by the people most impacted by its outcomes.

Moving from possibility to impact will require deliberate choices about how these systems are built and governed. That means pairing generative AI with participatory design, investing in local capacity and accessible tooling, and establishing clear standards for evaluation, auditability, and harm mitigation across contexts. 

Done well, a generative AI-powered ontology could become more than a technical upgrade to the SDGs. It could become a new way to learn, coordinate, and make decisions together, grounded in both evidence and lived reality, thereby ushering in a new way to achieve better human development outcomes.