OpenMined Foundation

The focus should be on the global governance of different forms of decentralized AI, including federated learning, agentic AI, and more generally the emerging "compute-to-data" paradigm, where data remains local and computation is distributed to where the data resides. Public data on the internet has already been extensively scraped, but vast amounts of untapped private data could still be leveraged in a trustworthy way through decentralized AI technologies. By contrast, the current centralized "data-to-compute" paradigm tends to concentrate power in the hands of a few actors and does not enable the use of sensitive, highly valuable private data for AI training in the way that decentralized compute-to-data approaches can. The summit should therefore acknowledge both the opportunities and the challenges raised by these decentralized technologies. It should explore what needs to be done to ensure that they strengthen AI and data sovereignty for nations globally, rather than being impeded, captured, or taken over by the few actors that currently dominate the centralized AI and data paradigm. A key part of this discussion should be the role and importance of open-source technologies. Open-source tools can help make decentralized AI more transparent, auditable, interoperable, and accessible to a broader range of public and private actors. They can also reduce dependency on a small number of dominant providers and support more distributed, sovereign, and trustworthy AI ecosystems. It is also important to examine different forms of global governance mechanisms for decentralized and agentic AI, and how these mechanisms could effectively and efficiently address, at global level, questions such as interoperability, liability, cybersecurity, privacy, identity management, user adoption, robustness, output verification-attribution-transparency and private data control & monetization by their owners.

• Safe, secure and trustworthy AI
• Open-source software, open data and open AI models
• Interoperability of governance approaches
• Transparency, accountability, and human oversight

Nations worldwide do not want to depend on a handful of providers of extremely powerful, centralized AGI or superintelligent systems that only a few actors can develop by concentrating the world's data and building vast, oversized data centres. A better future would be one in which a multitude of specialized less powerfull AI systems and AI agents interoperate in a highly efficient and effective way based on the instructions and accountability of their owners. Such an approach would help avoid overdependence, inefficiencies, and market dominance resulting from the centralized AGI/superintelligence narrative.This is a more sovereign world for all nations worldwide without any AI and Data colonialism.

Yes, the list completely misses the importance of access to training data, especially private data, for building the next generation of AI systems. Data remains the most important factor. Ultimately, everything comes down to increasing the capability of AI systems - whether large or small, centralized or decentralized, open or closed - to process more data, faster and more effectively. This can be achieved either by applying more computing power or by developing faster and more efficient algorithms that can extract more value from the available data. The processing of private data is the next major frontier and one of the most promising paths to faster progress in AI. However, private data cannot be centralized for sensitivity, confidentiality, and privacy reasons. Nor should it be centralized, because doing so would further concentrate power in the hands of a small number of actors, as is already the case in the current centralized AI paradigm..

In my region of the world, changing geopolitical alliances are reshaping the strategic landscape. Close and trusted allies of yesterday are increasingly becoming strictly transactional commercial actors, ready to weaponize trade, technology, energy, and defence dependencies more aggressively and strategically to their own economic advantage. They are also seeking ways to constrain the enforcement of sovereign democratic policy choices by exploiting strategic choke points in energy, technology, and defence. The most significant opportunity arising from this shift is the transformation of my region into a more technologically, energetically, and militarily sovereign area. This requires a new industrial policy, major strategic investments, and a renewed effort to boost our economy and internal market. It also clearly means diversifying our trade partnerships more effectively, favouring partners that are less likely to weaponize technology, energy, or defence dependencies.

We need a high-level plan to ensure that the next layer of the internet now emerging before us — driven by agentic and decentralized AI systems — does not end up reinforcing the concentration of power already held by today's dominant actors. This requires developing a clear and acute awareness of the risks at stake, as well as a strategy to level the playing field. The objective should be to prevent excessive concentration of power and ensure that the development of agentic and decentralized AI supports more open, competitive, interoperable, and sovereign digital ecosystems.

We need a new benchmark demonstrating that the combination of decentralized, open-source AI agents and systems with controlled access to private data can be as powerful as, or even more powerful than, classic centralized AI approaches. Such a benchmark would help prove that decentralized AI is not only more trustworthy and sovereignty-preserving, but also technically competitive. This could accelerate user adoption by showing that decentralized architectures can deliver comparable or superior performance without requiring private data to be centralized. We still need to build the decentralized technologies and protocols that would allow the owners of AI agents to authorize other AI agents to train on, query, or otherwise process their private data for specific and verifiable purposes. This requires a framework, protocol, or set of de facto rules to be followed by all AI agents. Such rules should cover the legal identity of AI agents, the accountability of their owners, the traceability of their operations, broad interoperability, and the mechanisms through which AI agents can securely access controlled private data in order to leverage and monetize it without requiring that data to be centralized.

We need to focus all key stakeholders on building the next ARPANET of AI: a major global legal and technical sandbox involving the deployment of interoperable AI agents across different jurisdictions and continents. This sandbox should be organised around a concrete problem of common interest, possibly linked to the Sustainable Development Goals, public-interest AI, or the health sector. The objective would be to test how decentralized and agentic AI systems can operate across borders, while enabling controlled access to private or sensitive data and preserving data sovereignty. From this initiative, a shared technical and legal architecture should emerge, together with the governance bodies needed to oversee agentic AI. In the same way that the early internet gave rise to protocols, coordination mechanisms, and institutions for global interoperability, this AI Dialogue should help define the protocols, rules, safeguards, and governance structures for the next generation of decentralized AI systems.

In general, private data owners are not involved as closely as they should be. Yet they are essential to advancing the next generation of decentralized AI systems and driving user adoption. Many data owners are concerned that their assets could be scraped, misused, or exploited in the absence of adequate protocols. Existing mechanisms such as robots.txt are no longer sufficient for the AI era. Private data owners therefore need to be convinced that it is in their interest to help build a new vision of powerful AI agents in which data remains local, and data owners retain control over the specific AI predictions, purposes, and use cases for which their data may be used

The best format would be to begin with a clear presentation of the stakes of decentralized and agentic AI, including the role played by open-source technologies and private data. This should include a precise explanation of what is currently missing: the technical protocols, legal frameworks, governance mechanisms, and trust infrastructure needed to enable AI agents to interact, interoperate, and access private data under the control of its owners globally. In a second set of parallell sessions, participants could then be divided into breakout groups and asked to design a global sandbox project of global public interest. The purpose of this global sandbox would be to help develop and test the missing technical and regulatory elements needed for decentralized and agentic AI systems to operate across jurisdictions (with no legal risks for the parties involved in the project) In a third consolidating session, the proposed sandbox projects could then be presented, discussed, validated, and voted on by all participants. The winning proposal could be endorsed as a global initiative launched as an outcome of the IGF, and driven by the appropriate stakeholders to ensure its success.Such an initiative should be supported financially by a coalition of public and private bodies, and should aim to produce a concrete technical, legal, and governance architecture for the next generation of decentralized AI systems and Agentic AI.

European Union AI regulatory sandboxes and real-world testing provisions, established by the AI Act, provide useful models that could be generalized at the global level. They show how innovation can be supported in a controlled environment, while allowing regulators, developers, data owners, users, and civil society to identify risks, test safeguards, and refine governance mechanisms before large-scale deployment. A global AI Dialogue could build on this approach by promoting cross-border legal sandboxes for decentralized and agentic AI systems, enabling practical experimentation across jurisdictions while developing the missing technical protocols, accountability rules, interoperability standards, and governance structures needed for trustworthy global deployment.