The 2025 AI Action Summit revealed that there are even more challenges in this area than initially anticipated in the absence of dedicated governance actions. Hence, we propose a set of actions that encompasses lifecycle management, impact assessment, regulatory policy, and shared mechanisms to ensure that open-source AI solutions tangibly advance progress toward the SDGs.
Integrate sustainability into open-source AI lifecycle management
Integrating sustainability across the AI lifecycle requires a multi-pronged strategy that addresses computing hardware, model development, and application deployment25, which could shift from fragmented reinvention to governed commons. The computing hardware phase requires a fundamental shift toward high-efficiency computing, built on the principles of energy-efficient architectures, liquid cooling, and renewable energy, which have demonstrated significant environmental benefits such as reducing data center energy demand by up to 20% and water consumption by up to 52%26. In the model development phase, creating shared, modular AI architectures can eliminate the immense waste from redundant development efforts. For example, using model distillation and pruning techniques can result in an overall Large Language Model compression of around 70%27. In the deployment and application phase, it is essential to mitigate waste from the redundant deployment of numerous similar models by transitioning towards a deployment-centric framework that prioritizes the use of multimodal foundation models as reusable infrastructure28. In addition, many technical efficiency metrics are vital for sustainability in the AI lifecycle, such as Power Usage Effectiveness and carbon emissions per training or inference. Beyond these metrics, effective governance requires a broader “Return on Environment” indicator29. This evaluates the net contribution of open-source AI to the SDGs by balancing its direct environmental costs against the benefits it generates.
Establish a framework for quantitative AI impact indicators
Currently, there is a lack of standardized quantitative indicators to analyze the complex impacts of open-source AI on the SDGs across different spatial and temporal scales. The India-AI Impact Summit 2026 shifts from dialogue to demonstrable impact, highlighting that governance requires quantitative impact frameworks as a prerequisite, especially for countries in the Global South30. The critical action is to incorporate the impacts on the SDGs by establishing composite frameworks that collect, integrate, and associate multiple baseline indicators, accurately measuring the facilitating and inhibiting impacts of multiple dimensions generated by different AI models. A typical example is the 13 Sustainability Criteria for AI Systems, which cover over 40 indicators and measure the social, environmental, and economic sustainability of AI clearly31. In addition, the establishment of local AI impact indicators adjusts local social, economic, and environmental characteristics under a quantitative impact framework, which is essential for countries to systematically collect and develop corresponding datasets and model repositories aligned with these indicators. This enables normative indicators to be transformed into analytically robust evaluation outcomes, ultimately operationalized as actionable policy practices.
Assure trustworthy open-source AI through regulation and scrutiny
The greater the misuse of AI, the more difficult it becomes for humans to distinguish authentic content from harmful synthetic media. Establishing robust governance for these high-stakes SDG applications requires a clear accountability chain involving developers, users, and governments32. For developers, especially at major tech companies, it is crucial to use security-audited training datasets to mitigate bias and to adopt explainable AI technologies appropriate to different model types. Improving transparency in the use, modification, and distribution of AI models relies on licensing agreements like the Open Source AI Definition23. Governance of open-source AI security calls for mandatory third-party audits and the implementation of simulated adversarial attacks to mitigate specific vulnerabilities. Furthermore, governments worldwide hold developers accountable and prevent users from spreading malicious content by establishing regulatory policies, such as the EU AI Act and the U.S. Transparency in Frontier AI Act. Finally, users are accountable for applying AI ethically and helping to identify and report misuse, such as strictly adhering to the “Prohibited AI Practices” defined under the EU AI Act.
Build global knowledge-sharing mechanisms and regional cooperation
Although the transition to open-source AI has improved democratizing access globally, inequality persists in other areas, such as computing infrastructure, talent distribution, and data sovereignty. From a global perspective, building open-access platforms that adhere to the Findability, Accessibility, Interoperability, and Reusability principles is crucial to ensure equitable access to computation, data, and foundational AI models. The United Nations would play a critical role in governing these AI infrastructures, helping reduce governance inequalities across Global South countries. For example, an institution named Global Dialog on AI Governance could strengthen global cooperation mechanisms by publishing policy-oriented annual reports on AI infrastructure performance and by conducting intergovernmental negotiations alongside broad consultations with diverse stakeholders33. In addition, a Global Fund for AI would help ensure that all countries secure a minimum irreducible AI capacity across skills, compute, data, and models34. From a regional or national perspective, valuable datasets remain confined within local government institutions and private corporations due to data sovereignty regulations that prevent global sharing35. Hence, cooperation between global platforms and regional research centers is essential to leverage shared foundational models while integrating proprietary, localized datasets. To navigate complex data-privacy regulations such as the General Data Protection Regulation, partnerships can adopt data-sharing models that provide controlled interfaces to sensitive data or rely on metadata, representative samples, and synthetic datasets through Retrieval-augmented generation and Parameter-Efficient Fine-Tuning. These regional research institutions, such as The Australian Artificial Intelligence Institute and The National Institute for AI Evaluation and Security in France, serve as critical intermediaries that enable the localization and deployment of foundational models by synergizing global resources with regional data to develop context-specific solutions.