1 00:00:09,717 --> 00:00:27,398 So welcome to the Global Network webinars. My name is Alexander Loshki and I will be moderating this webinar today together with my colleague Sean Lovell. Today, we have with us Kala Yem Amesaga from the 2 00:00:27,398 --> 00:00:48,760 World Economic Forum, who will speak about AI governance and data sovereignty. This is the third of five webinars in our AI for Official Statistics series. As AI reshapes economies, institutions, and societies, 3 00:00:48,760 --> 00:01:15,174 data sovereignty is emerging as a practical governance challenge rather than a purely legal or geopolitical concern. This webinar will examine how sovereignty should be understood, not as isolation or self-sufficiency, but as strategic control, resilience, 4 00:01:15,174 --> 00:01:37,957 and trusted in the and interdependence across the data and AI value chain. Moving beyond narrow debates on localization, the session will explore why sovereignty must reflect the differentiated capacities of states, cities, communities, platforms and multilateral actors. 5 00:01:39,237 --> 00:02:00,757 With AI agents and cross-border data use rising. Raising new questions of confidentiality, equity, and control, the discussion will focus on governance models, interoperable standards, and regional collaboration that enable innovation while safeguard agency. 6 00:02:00,797 --> 00:02:02,597 Trust and shared value. 7 00:02:04,597 --> 00:02:27,916 So let me introduce now our speaker. Carla is Lead of AI and Data Governance at the World Economic Forum. She's working on topics such as AIAI agents governance, data sovereignty, and data equity. She's passionate about exploring the ways in which data, AI, 8 00:02:27,916 --> 00:02:36,757 and technology can promote digital transformation of governments, public-private collaboration, and inclusive development around the world. 9 00:02:38,957 --> 00:02:54,677 Prior to joining the World Economic Forum, she worked at the Inter-American Development Bank, supporting digital government and statistical capacity building projects. Before that, she worked also at the World Bank, OECD. 10 00:02:55,757 --> 00:03:11,879 Ashoka Changemakers and the Mexican Ministry of Foreign Affairs. Carla holds a master's degree in international affairs from the University of California in San Diego and a bachelor's degree in political science 11 00:03:11,879 --> 00:03:18,677 and in international relations from the Autonomous Technological Institute. of Mexico. 12 00:03:21,477 --> 00:03:37,999 So just before I hand over to Carla, just some practical remarks, as always. Please write your comments and questions into the meeting chat. You can do that at any time. And then during the Q&A session, 13 00:03:37,999 --> 00:03:45,477 we would encourage you to ask the question yourself if you can. Otherwise, we will read the question out for you. 14 00:03:47,317 --> 00:04:02,357 As always, this Global Network webinar is being recorded and will be made available on the Global Network of Data Offices and Statisticians on LinkedIn. And you will find the link to join this LinkedIn group in the meeting chat. 15 00:04:04,837 --> 00:04:18,597 We also invite you to follow the link to subscribe to our monthly newsletter. And that would be everything from my side. So now over to you, Carla. The floor is all yours. Thank you. 16 00:04:21,477 --> 00:04:33,025 Thank you so much, Alexander, Sean, and thank you to the Global Network of Data Officers and statisticians for the invitation. And of course, thank you all for joining today. Good morning, good afternoon, 17 00:04:33,025 --> 00:04:48,882 and good evening, wherever you are. Know that this is a complex time for the data and official statistics community. You've been asked to produce faster, more trusted, and more granular statistics. You're also being asked to work with new 18 00:04:48,882 --> 00:05:02,341 data sources, new technologies, and new institutional expectations. sometimes with budget constraints. And that was already true over eight years ago when I was at the Inter-American Development Bank, 19 00:05:02,341 --> 00:05:20,924 working closely with NSOs on capacity building and digital transformation. And it is even more true today as AI is moving from experimentation into core institutional workflows. And at the same time, more data is being collected, processed, and reused by private actors, platforms, 20 00:05:20,924 --> 00:05:34,437 and automated systems, requiring even more cross-sector collaboration. So AI governance and data sovereignty have become key practical questions for public institutions. And this is a topic that we're going to discuss today. 21 00:05:35,397 --> 00:05:39,557 Now I'm going to start sharing my slides. So bear with me for a second. 22 00:05:50,437 --> 00:06:05,666 The ideas that I will share with you today are based on the work carried out at the World Economic Forum. including through our Center for AI Excellence. And this community brings over 600 high-level experts and innovators to 23 00:06:05,666 --> 00:06:12,597 drive AI and data usage in ways that foster innovation, economic transformation, and human-centric outcomes. 24 00:06:14,437 --> 00:06:27,466 This work is also derived from the insights from our experts that are part of the Global Future Council Network, in particular the Global Future Council on Data Frontiers. This is part of the network of our 700 25 00:06:27,466 --> 00:06:47,237 experts from all sectors in all parts of the world. And this council seek to advance ideas for innovative and positive impact in the world. During my presentation, I will also pose some questions for self-reflection that will help us explore some of the concepts being discussed. 26 00:06:48,517 --> 00:07:00,377 And throughout the call, I invite all of you to share examples, tools, case studies, research, or any reactions that you have on the topic presented on data sovereignty and AI governance more broadly. 27 00:07:00,377 --> 00:07:07,957 We would love to hear from you. And we will also have some time at the end of the the presentation for questions and answers. 28 00:07:12,437 --> 00:07:32,041 So today I will focus on 7 key themes, why data sovereignty matters now, how data sovereignty is becoming operational, why sovereignty is not the same as localization, why the sovereignty landscape has been changing and now it's heterogeneous, 29 00:07:32,041 --> 00:07:45,317 why we need to move from ownership to operational rights, how AI agents make sovereignty more urgent, and what trusted interdependence AI infrastructure And digital embassies mean for statistical offices. 30 00:07:49,237 --> 00:08:05,677 First, let's start with setting the baseline for our conversation. Today we'll be talking about data sovereignty, not from a legal term per se, or not from the geopolitical angle, but more from a practical governance question. 31 00:08:06,877 --> 00:08:20,517 When data is collected, shared, processed, used to train or query AI systems, or acted on by AI agents who has meaningful control. And please allow me to emphasize on the word meaningful because this really 32 00:08:20,517 --> 00:08:33,576 matters. A government may say they die sovereign because it is stored luckily. A company may market a product as sovereign because the cloud region is domestic. Or a regional block may use sovereignty 33 00:08:33,576 --> 00:08:48,037 as a strategy for competitiveness or autonomy. And these three claims may all be valid in part, but none is sufficient on its own. So today I want us to explore together what sovereignty looks like in practice. 34 00:08:51,357 --> 00:09:07,606 First, let me begin with a complex yet increasing common scenario that you might be experiencing. Imagine a national statistical office preparing a labor market estimate. The workflow uses survey data, admin tax data, 35 00:09:07,606 --> 00:09:14,437 private sector job posting data, and machine learning model hosted through a cloud provider. 36 00:09:16,357 --> 00:09:33,297 Then, on the other hand, you have a GenAI assistant helping analysts drop documentation, or maybe an AI tool checking anomalies in the data and doing quality control. Some data states in the country, some metadata or data crosses borders, 37 00:09:33,297 --> 00:09:49,157 and maybe there's some even some processing happening to a trusted provider. And then the outputs of this analysis are maybe reused by another agency in another country. So now ask yourself, where is the data? 38 00:09:49,157 --> 00:10:05,397 After all this back and forth and all these changes, where's the data? And this is a key question. But this question by itself is no longer enough. We also need to ask ourselves, who approved the data use? 39 00:10:06,697 --> 00:10:25,962 What model uses data? What agent. To net. What metadata follow it? Can the statistical office audit the process later? Can it stop, correct, or revoke downstream use? and the people, businesses and communities represented 40 00:10:25,962 --> 00:10:27,837 benefit from the value created. 41 00:10:31,277 --> 00:10:41,917 These are not straightforward or easy questions to answer sometimes, but these are important and kind of like give us an insight of where the conversation has been shifting. 42 00:10:43,917 --> 00:10:58,245 Because we're talking about a shift in perspective. If sovereignty is not only about where data sits, but about who can govern, what happens to it, then the first practical question would be, how is data sovereignty changing as AI 43 00:10:58,245 --> 00:11:07,277 becomes part of everyday institutional work? And that brings us to the main first point, data sovereignty is becoming operational. 44 00:11:09,197 --> 00:11:26,037 That is no longer simply collected, stored, analyzed, and published. It is increasingly used to train models, fine-tune systems, generate synthetic data sets, power AI assistance, support automated workflows, and informed decisions in real time. 45 00:11:28,717 --> 00:11:38,877 So the core question is shifting. It is no longer is the data inside our borders. It is also, can we govern what happens to the data across its full life cycle? 46 00:11:41,677 --> 00:12:03,871 And for statistical offices, this life cycle is familiar. It includes collection, cleaning, linkage, classification, analysis, quality insurance, assurance, dissemination, archiving, and reuse. But AI adds new layers around model training, prompt logs, 47 00:12:03,871 --> 00:12:09,997 synthetic data generation, automated coding, and agentic workflows. 48 00:12:12,157 --> 00:12:32,434 And as we explore these new implications that are brought by AI, we find ourselves in a position where there's no necessarily straightforward definition of data sovereignty. At the forum, we've been exploring with our experts, 49 00:12:32,434 --> 00:12:47,837 like what are the different angles of data sovereignty in the context of AI. And what we found is that there's a joint definition that kind of like touches on concepts like governance, values, control, flexibility, 50 00:12:47,917 --> 00:13:08,752 resilience, localized investment, but also trusted collaboration, autonomy, and international cooperation. In some of these concepts embedded in data sovereignty, first talk about sovereignty in the form of capacity and not necessarily just 51 00:13:08,752 --> 00:13:26,157 formal authority. They also convey the idea that sovereignty is not isolation. A country can keep data inside its borders and still lack control if compute, model updates, encryption keys, or access controls sit elsewhere. 52 00:13:27,357 --> 00:13:41,935 And 3rd and very important is that sovereignty must support interoperability and trusted collaboration. For many countries, institutions, and communities, sovereignty will be achieved through well-governed interdependence and not 53 00:13:41,935 --> 00:13:43,597 through full self-sufficiency. 54 00:13:46,637 --> 00:13:51,037 So that's why sovereignty is increasingly a governance issue, not only a legal one. 55 00:13:53,037 --> 00:14:00,557 And it's about whether an institution can act, verify, correct, exit, and benefit from these data interactions. 56 00:14:04,077 --> 00:14:25,632 Now let's pause for a moment and try to internalize some of the concepts that we have discussed. And I invite you to think about one AI or data workflow in your institution. And if you had to explain who controls the data, the model, the compute, and the audit trail. 57 00:14:25,632 --> 00:14:29,757 Could you answer all of these questions confidently? 58 00:14:41,677 --> 00:14:54,797 Once we define sovereignty as operational capacity, one important implication follows, that we cannot really measure sovereignty only by looking at geography. Location matters, yes, but it's only one layer of control. 59 00:14:56,317 --> 00:15:13,408 A dataset might be stored locally while real authority sits elsewhere, either in the cloud architecture, the model provider, the encryption systems, the logs, or the vendor contract. So the next step is to separate sovereignty from one of its most common 60 00:15:13,408 --> 00:15:15,197 associations, localization. 61 00:15:18,797 --> 00:15:33,837 Sometimes localization is necessary, yes, for sensitive government data, national security data, health data, or any other confidential statistical microdata. But localization alone does not create sovereignty. 62 00:15:35,237 --> 00:15:50,397 A data center inside the country does not automatically answer the key questions. Who controls the hardware? Who operates the cloud? Who access the logs? Who audits the model? and what exit options exist if the provider relationship changes. 63 00:15:51,437 --> 00:15:54,317 So localization can move data without moving power. 64 00:15:56,157 --> 00:15:59,837 An assistant can look sovereign on a map and still be dependent in practice. 65 00:16:01,757 --> 00:16:18,957 And at the same time, data stored abroad is not automatically outside sovereign control. Strong legal safeguards, technical controls, audit rights, encryption, and trusted institutional arrangements can all support control across borders. 66 00:16:20,477 --> 00:16:34,237 So instead of treating sovereignty as a yes or no question, might be more useful to think in terms of a full stack. Instead of asking, is the data local? Perhaps we can ask, where does the authority actually sits? 67 00:16:35,357 --> 00:16:41,437 and where do we need control, where do we need partnerships, and what safeguards make that partnership trustworthy? 68 00:16:45,677 --> 00:16:57,677 And this is a moment also to reflect. For example, let's pause for a few seconds. Think of one technology vendor or technology platform your institution relies on. 69 00:16:59,917 --> 00:17:09,757 Now, ask yourself, would you answer who controls the encryption keys, the audit logs, the subcontractors, and exit process? 70 00:17:11,677 --> 00:17:27,074 Perhaps you can answer these questions, or perhaps you have not fully stopped to understand if you have like all the information needed from your either vendors or external providers. And these are the important questions to 71 00:17:27,074 --> 00:17:35,917 ask yourself as you are. either engaging in contracts as you are processing or handling data, especially if it has a sensitive nature. 72 00:17:43,037 --> 00:17:57,256 And this full stack view that I introduced also shows that data sovereignty is not held by 1 actor alone. Once we ask where authority actually sits, we see many layers and many actors involved. We see governments, 73 00:17:57,256 --> 00:18:11,601 statistical offices, cities, communities, tech providers, researchers, and regional institutions. which means that sovereignty is not uniform. It's different across actors and have different force and control, 74 00:18:11,601 --> 00:18:13,757 responsibilities and vulnerabilities. 75 00:18:17,357 --> 00:18:21,917 And that brings us to the broader sovereignty landscape and to the question of data equity. 76 00:18:24,477 --> 00:18:40,517 The concept of sovereignty naturally makes us think of states. And indeed, states are central to this conversation. They make the laws, negotiate the agreements, regulate the markets, invest in infrastructure, and set the national strategies. 77 00:18:41,357 --> 00:18:54,824 But in the data and AI era, sovereignty is shaped by many actors. by statistical offices because they're, and you are the trusted data stewards, setting standards, protecting confidentiality and deciding 78 00:18:54,824 --> 00:19:12,317 what evidence is produced and how it is communicated. But also other actors are very important cities as they collect different kinds of urban data, communities, indigenous communities, language communities, migrant communities, diasporas. 79 00:19:15,197 --> 00:19:29,963 Also, private platforms matter, hyperscalers, model providers, cloud vendors, data brokers. really shaping what is technically possible and commercially viable. And regional and multilateral institutions matter. 80 00:19:29,963 --> 00:19:35,677 They can support shared standards, regional collaboration, capacity building, and interoperability. 81 00:19:38,157 --> 00:19:56,345 So in this landscape with sovereignty is shared across various actors, we see that not every actor exercises control the same way. The national government may exercise sovereignty to regulation, procurement, 82 00:19:56,345 --> 00:20:06,237 and infrastructure investment. But the Statistical Office may exercise it through confidentiality protocols, access rules, metadata standards, and quality frameworks. 83 00:20:10,957 --> 00:20:21,517 And this is a great moment to introduce also why data equity has become essential, has always been important, but has become essential in this conversation. 84 00:20:24,157 --> 00:20:32,717 And what I mean by data equity is the shared responsibility for fair data practices that respect and protect human rights, opportunity, and dignity. 85 00:20:34,637 --> 00:20:48,535 And this concept is important in the conversation of data sovereignty because it makes us look at the full data life cycle, the collection, curation, processing, analysis, sharing, application, retention, 86 00:20:48,535 --> 00:20:50,637 reuse and disposal of the data. 87 00:20:54,157 --> 00:21:07,517 it makes us ask ourselves, like, who is visible in the data? Who's missing? Are the attributes used to describe people or communities accurate and appropriate? who can access the data, tools and benefits. 88 00:21:09,437 --> 00:21:20,877 and also who benefits from the data, who might be harmed or who may want to seek a remedy from any negative externality or any negative outcome. 89 00:21:25,597 --> 00:21:38,513 So groups that are underrepresented in data are often underserved in policy. So that's why this is a key consideration, especially for statistical offices and government offices, because poor quality data about our 90 00:21:38,513 --> 00:21:53,357 community can lead to poor quality decisions. about that community. And AI in the context that we're living in can also amplify these patterns, these biases, these gaps, if they're not addressed deliberatively. 91 00:21:55,197 --> 00:22:09,416 So sovereignty should not be reduced to state control over data. It should also include different agencies and different actors and help advance the ability of institutions, communities, and people to shape how data about them 92 00:22:09,416 --> 00:22:14,717 is used. how data is shared and how data is benefiting the communities that is coming from. 93 00:22:20,797 --> 00:22:33,837 So if sovereignty is shared across many actors as we saw, then the language of simple ownership quickly becomes insufficient. the practical question is no longer just who owns the data. 94 00:22:35,837 --> 00:22:46,797 But now, and especially for statistical offices, the more useful question could be, who has and which rights, under which or what conditions, and with what accountability. 95 00:22:49,837 --> 00:23:05,386 So we need to move from ownership as a broad claim to operational rights as a governance tool. Many data governance institution discussion still rely on the language of ownership. The language can be used in some context, but it's often too blunt for public 96 00:23:05,386 --> 00:23:07,357 sector data and official statistics. 97 00:23:09,197 --> 00:23:28,751 Can a business own its survey response once it is used for official statistics? Can a statistical office own admin data collected by another ministry? These questions become quickly complicated. So for a practical standpoint and practical definition of governance, 98 00:23:28,751 --> 00:23:34,957 it is often more useful to shift from ownership to operational rights and controls. 99 00:23:36,957 --> 00:23:54,077 So for any important data set or AI enabled workflow, we should ask questions like this instead. Who has the right to access this data? Who can approve a new use? Who can audit the system? Who can correct or challenge outputs? 100 00:23:54,797 --> 00:23:57,757 and who can exit or revoke any access. 101 00:24:01,037 --> 00:24:18,409 And for statistical offices, these questions are concrete and important. take administrative data. The issue is not whether a tax authority or health ministry shares data with the statistical office, but the greater issue is whether the statistical purpose is clear, 102 00:24:18,409 --> 00:24:32,717 quality limitations are documented, linkage is governed. Access is logged and the resulting statistics are trusted. And trust, as we know, is at the center of all data interactions and exchanges. 103 00:24:38,397 --> 00:24:56,717 So moving on. Sober team practice means being able to answer. questions like. What is this data? Where did it come from? What has been done to it? What rights has attached to it and what happens next? 104 00:24:59,847 --> 00:25:03,287 And this makes Providence. Traceability and. Central. 105 00:25:05,997 --> 00:25:11,047 A statistical and data offices already understand metadata better than. Any other institution? 106 00:29:11,676 --> 00:29:18,876 We should define roles, set permissions, explain escalation rules, monitor performance, and review access over time. 107 00:29:21,436 --> 00:29:28,556 Every significant AI agent. working in a statistical or government environment should have a clear authorization profile. 108 00:29:30,796 --> 00:29:47,996 related to the purpose, the data you can read, where actions are prohibited, where actions require a human in the loop, how lots are kept and such. And this is how institutions can preserve trust while using more capable tools. 109 00:29:49,276 --> 00:30:01,836 And this also connects directly to data sovereignty because if an agent can act on data, move data, transform it, and trigger decisions, then sovereignty depends on the controls that we set around that agent. 110 00:30:07,516 --> 00:30:15,916 let me go back to a bigger point that was made earlier, that control is essential, but control does not mean isolation. 111 00:30:17,756 --> 00:30:34,507 So even the strongest governance frameworks depend on infrastructure, such as compute, connectivity, storage, talent, cybersecurity, skills, vendors. So the next question becomes, especially in a global context, 112 00:30:34,507 --> 00:30:45,796 how can countries and institutions preserve agency while participating in an interconnected AI ecosystem? And this is where trusted independence becomes very relevant. 113 00:30:51,116 --> 00:31:05,196 So for many countries and institutions, isolation is neither realistic nor desirable. AI infrastructure is expensive, compute is concentrated, talent is unevenly distributed, and data standards vary. 114 00:31:11,356 --> 00:31:31,005 Also, many statistical challenges are cross-border by nature. Migration, climate, trade, health, digital platforms, supply chains, financial flows. So through research at the World Economic Forum, we have found that AI infrastructure can be built around 3 main building blocks, 115 00:31:31,005 --> 00:31:37,356 compute, which is data in use, connectivity, data in motion, and storage, data at rest. 116 00:31:39,276 --> 00:31:54,716 So to scale these building blocks, economies need technical prerequisites, energy, water, land, hardware, and cybersecurity. but they also need institutional prerequisites, policy, talent, and financing. 117 00:31:56,636 --> 00:32:01,996 And this is why AI sovereignty should be understood as a strategy and not just a purity test. 118 00:32:04,316 --> 00:32:18,324 And also because economies sit on a spectrum, at one end is extensive domestic ownership, where there's local control of compute, connectivity and storage. And at the other end, you find trusted international 119 00:32:18,324 --> 00:32:24,236 partnerships. such as shared infrastructure and governance with reliable partners. 120 00:32:27,276 --> 00:32:45,669 So we find that in practice there is. very few instances where, countries operate at the extremes, and actually most economies will use hybrid strategies. And this is where digital embassies concept becomes relevant. 121 00:32:45,669 --> 00:33:02,636 And this is not a new concept, but it's not well known. It's basically a secure digital space, just like a regular embassy, but think about it in the digital context. It's a secure digital space in a host country where a guest. 122 00:33:02,716 --> 00:33:21,166 country can, the guest countries' data, workloads, storage, and sometimes compute remains governed under the guest countries' own jurisdiction. So you can host data workloads. somewhere else, but they're still obeying your own 123 00:33:21,166 --> 00:33:38,833 jurisdiction. And here we're not just talking about placing data abroad. We're talking about extending sovereign control through legal, technical, and operational safeguards. And Stonia pioneered the concept by hosting government data and essential 124 00:33:38,833 --> 00:33:45,596 registries in Luxembourg to preserve continuity of core state functions if domestic systems were disrupted. 125 00:33:47,876 --> 00:34:00,613 And now the concept is re-emerging and gaining more traction in the AI infrastructure debate because many economies need access to sovereign government capacity before they can build everything domestically. 126 00:34:00,613 --> 00:34:15,641 And this is important because countries should not necessarily aim to develop the whole AI stack in-house, but can rely on collaboration and exchanges to do this. So digital embassies can help governments 127 00:34:15,641 --> 00:34:35,676 preserve continuity when domestic infrastructure is disrupted. expand access to computer storage when local buildout is low or expensive, keep data and operations subject to home country law, and scale capacity with demand rather than building large assets upfront. 128 00:34:38,636 --> 00:34:52,938 But digital embassies only work if trust is designed into the arrangement. And when we talk about trust, we mean there has to be political commitment and willingness, a clear legal basis and scope of the 129 00:34:52,938 --> 00:34:58,316 agreement, strong data management, technical safeguards, and operational rules. 130 00:35:02,156 --> 00:35:15,916 And these arrangements need to be safe and need to be audited. And they have to also envision scenarios where political conditions change and the data needs to be withdrawn. 131 00:35:18,556 --> 00:35:29,836 So the digital embassy can help countries strengthen sovereignty and collaboration, but only if it's legally robust, interoperable, auditable, encrypted, resilient, and portable. 132 00:35:35,316 --> 00:35:49,817 And this is an important lesson for official statistics, because trusted interdependence means investing domestically where control is essential, using shared or regional capacity where it is sensible, and designing every partnership so that 133 00:35:49,817 --> 00:35:54,156 accountability. Observability, exit, and benefit sharing are real. 134 00:35:57,596 --> 00:36:14,213 So taking all these together, these examples point to one conclusion that data sovereignty in the AI era is not a single policy choice or a single technical control or procurement clause. It's a set of capabilities that must be 135 00:36:14,213 --> 00:36:21,516 designed Tested and maintain across data models, infrastructure, vendors, communities, and institutions. 136 00:36:23,716 --> 00:36:41,426 So now please allow me to translate all of this into five practical actions to consider. First. Map the data and AI stack for critical statistical functions. For your most important workflows, know what data enters, where it is stored, 137 00:36:41,426 --> 00:36:59,701 where it moves, what compute processes it, what models or tools touch it, what vendors are involved, and what safeguards apply. Then apply data equity as part of the sovereignty governance. So ask who is represented, who is missing, who may be harmed, who benefits, 138 00:36:59,701 --> 00:37:00,876 and who has a voice. 139 00:37:03,276 --> 00:37:20,956 Third, make provenance and metadata non-negotiable. You should be able to trace not only a data set, but also major transformations, AI assisted steps, synthetic data generation, model versions, human approvals, agent actions, and downstream dissemination. 140 00:37:24,796 --> 00:37:39,098 Then create authorization profiles for AI agents in high impact tools. So be clear about what the tool can read, write, recommend, publish, or trigger. And always keep humans in the loop, especially where confidentiality, rights, 141 00:37:39,098 --> 00:37:55,836 public trust, or irreversible consequences are involved. And finally, test for sovereignty washing in procurement and partnerships. Make sure that true sovereignty is verifiable. And it's not just being said on paper. 142 00:38:01,036 --> 00:38:16,876 And always ask yourself some sanity questions. Can you export your data? Can you port your model? Can you rotate your own encryption keys? Can you access process and transfer events be independently verified? 143 00:38:16,876 --> 00:38:18,956 Is there a tested exit plan? 144 00:38:22,476 --> 00:38:29,916 So across these actions, there's a clear message that sovereignty becomes meaningful only when it can be exercised in practice. 145 00:38:32,556 --> 00:38:52,574 And to close, sorry, and to close, let me share 3 final takeaways. First, data sovereignty is agency, not isolation. It is the ability to shape and govern data use across the data and AI value chain. Second, sovereignty is layered and differentiated. 146 00:38:52,574 --> 00:39:10,621 So states matter, but so do statistical offices, cities, communities, platforms, researchers, and regional institutions. Each has a different layer and leverage. In this equation. And third, sovereignty must be operationalized. 147 00:39:10,621 --> 00:39:22,036 It becomes real to rides, metadata, provenance, auditability, authorization, interoperability, benefits sharing, technical safeguards. The next adoptions. 148 00:39:25,396 --> 00:39:29,876 So for the statistical community, this is not a departure from your mission. It's an extension of it. 149 00:39:32,916 --> 00:39:43,236 You've always depended on a social contract, working with people, businesses, institutions to provide data because they trust it and trust it will be used responsibly, confidentially, and for the public value. 150 00:39:44,436 --> 00:39:53,876 And AI now is just changing what institutions can do with data and also changing what and who participates in data value chains, who has visibility and who has leverage. 151 00:39:56,196 --> 00:40:06,836 So in this context, how do we all preserve agency, trust and share value in a world where data models and infrastructure and agents are increasingly interconnected? 152 00:40:09,076 --> 00:40:12,756 That's the practical challenge of data sovereignty in today's world. 153 00:40:16,516 --> 00:40:29,156 Thank you so much for listening. My contact information is on the screen. I'll also share some links in the chat. Please reach out to me if you want to learn more about this work. And I look forward to the discussion. 154 00:40:33,396 --> 00:40:51,676 Thanks so much, Carla, for these very, very useful insights on data sovereignty and what it entails and what should be included or taken into consideration when talking about data sovereignty and data governance. I would like 155 00:40:51,796 --> 00:41:14,789 Now, to invite our colleagues in this webinar, if you have questions, please feel free to now raise your hand and we will give you the floor to ask your question. You can also... Write your questions into the meeting chat and we can read it out to you, 156 00:41:14,789 --> 00:41:34,640 but I think it would be nice if you could maybe ask the question yourself. So I saw there was one. Question in the meeting chat, so I just go in order of the questions as they appear in the chat. So Jose Miguel Guzman, 157 00:41:34,640 --> 00:41:39,236 do you want to come in? Otherwise I can read it out. 158 00:41:42,676 --> 00:41:45,076 So, Jose, please go ahead. 159 00:41:48,676 --> 00:41:53,316 And you are muted, so if you want to talk, please unmute yourself. 160 00:41:59,156 --> 00:42:16,156 I see your hand, Jose, but if it does, now. Okay, I think, yes. Thank you very much. It was so incredible, amazing presentation, Carla. Thanks. You have show us so many concept ideas. It's like we need to 161 00:42:16,596 --> 00:42:37,724 process everything because it's a condensation of so many things that are happening. I would like to ask you some practical question, and this is what I post in the chat about, you know that The UN is now changing a lot, the context, international context and financing 162 00:42:37,724 --> 00:42:53,199 context is changing quickly and then international organizations such as UNFPA or UNICEF, others must deal with this issue of... how to invest, more productively. And particularly in this area of data, 163 00:42:53,199 --> 00:43:07,476 UNFPA is an organization that, help countries to conduct sensors, to improve vital statistics, et cetera. And most of these institutions are now trying to develop a strategy in relation to data to update. 164 00:43:07,716 --> 00:43:22,311 their framework because they are from the past. So what complete recommendation you will give to those organization in terms of how what they should include in the data strategy that they ensure that issue of 165 00:43:22,311 --> 00:43:27,556 governability and issue sovereignty is considered in the framework that you have presented. 166 00:43:30,756 --> 00:43:45,596 Thank you so much for your question. And I mean, we could talk about this for a long time because linking about or thinking about linking, funding to like what are the priority actions that statistical offices can take and 167 00:43:46,516 --> 00:43:58,996 especially in the context of sovereignty. I think the first thing that I would say is that there should be a clear mapping and awareness of the capabilities and what are the gaps and understand like where are we sitting when it comes to the 168 00:43:58,996 --> 00:43:59,556 Mhm. 169 00:44:02,516 --> 00:44:17,565 the data ecosystem and in terms of like the skills that we have in-house and what are the processes that we are developing in-house and what are the processes that we're maybe doing through third parties. So I think that initial inventory can 170 00:44:17,565 --> 00:44:29,716 definitely help. make an assessment when it comes to like what are the areas that should be invested in. One, of course, critical area for investment is talent. You want to have in-house talent, but at the same time, 171 00:44:29,716 --> 00:44:42,946 Mhm. you want to understand like what are the ways through which you can maybe collaborate with other countries, use other countries' tools, or collaborate with external vendors. And that is important to us, 172 00:44:42,946 --> 00:45:00,276 as long as you know that you are retaining control, especially about sensitive data and you have agency, And the final say when it comes to like what happens to the data when the moment it leaves your office and when it's processed, when it's exchanged. 173 00:45:01,396 --> 00:45:13,945 So I would start by doing this inventory and understand like what are the gaps in terms of critical infrastructure or skills, the access to compute, access to energy. Then we'll identify what are those, 174 00:45:13,945 --> 00:45:32,449 if there's investment that can be done, go to some of these critical areas and where there's collaboration possible. And this is not new, just exacerbated by AI. There should be a strong focus about ensuring that data is produced with 175 00:45:32,449 --> 00:45:53,876 quality, that there is clear metadata, there's clear standards to produce this and that there's also data sharing. So you're not duplicating the work, you're not maybe asking too much of users, but you're doing your data processes in an efficient way. And I think that 176 00:45:54,916 --> 00:45:56,596 Then I've seen that in practice that. 177 00:45:58,596 --> 00:46:14,544 development banks and international organizations, they're also starting to look more into like the stage of development when it comes to data and AI processes to, determine funding allocation and funding 178 00:46:14,544 --> 00:46:19,076 decisions. So Just building these capacities is essential also to. 179 00:46:21,316 --> 00:46:29,716 have and to have transparency and to have accountability that can help also attract funding for strategic investment. 180 00:46:32,996 --> 00:46:45,076 Thank you. Thanks, Carla. Then we have a question by or comment by Christian Casanova. Do you want to come in, Christian? Christian? 181 00:46:48,276 --> 00:47:11,198 Sure. My question is around, because we talked a lot about myself, I'm actually I'm part of the Global Shaper community as well for the forum. But I wanted to ask in relation to the human data supply chain. And so you have a lot of companies like 182 00:47:11,198 --> 00:47:32,029 Markor, Scale AI, and I can name like another 10 of them that really are in the business of producing human data to train their models and essentially sell data to a lot of the frontier model companies here in the US or also in Europe. So my question to you, 183 00:47:32,029 --> 00:47:52,154 as we talk about AI governance, compute and privacy, and et cetera. And especially in relation to statistics, where does the layer, where is this human data center layer stand now and where would indeed possibly be regulated as critical AI 184 00:47:52,154 --> 00:48:07,578 infrastructure? Because these human data is being produced at a scale that may be outside the country at the moment, at least Markor operates that way. But if Europe wants to access that data for any form of statistical significance, 185 00:48:07,578 --> 00:48:20,276 is that going to be? an issue or is that going to be a regulation as part of some form of statistical production, supply chain or any other form of data that they're using from these human data centers? 186 00:48:23,076 --> 00:48:37,796 there's a, that's a complex question because there's definitely been a lot of talk about shortage of quality data to feed the models. And there's even been an increased movement towards smaller models and more efficient models just to 187 00:48:38,116 --> 00:48:53,556 kind of like cut that trend that was going into, let's create more data. And the companies that you're talking about, like Scale AI and such, are definitely playing a key role in this, you know, value chain because of. 188 00:48:53,716 --> 00:49:09,707 They're sourcing data and sourcing data that is machine readable and AI optimized. So that's another key point. So there is definitely. risks when it comes to how, really fully understanding and having 189 00:49:09,707 --> 00:49:25,322 full traceability of how this data is collected, because ultimately you don't want to end up with synthetic data that is generated, but that is like so far off from, the ground truth. the populations that is seeking to represent that is, 190 00:49:25,322 --> 00:49:45,734 no longer usable and can lead, to hallucination or problems in the model outputs. So there should definitely be clear standards and quality checks and assessments when it comes to like, what type of data is being produced in this context for massive ingestion by the 191 00:49:45,734 --> 00:50:03,737 models and if it's really accurate. And in that sense, I think the role of the National Statistical Office is there any official data providers become even more essential to really, because I mean, these agencies have the closest link to the ground through data, 192 00:50:03,737 --> 00:50:23,544 to the populations, to the communities, to the businesses, to help, you know, ensure that There's data quality and there's, precision and accuracy when it comes to the data is collected, but... answering to the to the second part of your question. 193 00:50:23,544 --> 00:50:42,633 There's definitely a need for I wouldn't say stronger regulation, but definitely the standards that ensure that the data that is produced and commercialized follow like all the quality, the key quality criteria, and that there is enough assessment when 194 00:50:42,633 --> 00:51:06,774 it comes to ensuring that. the data remains through or close to the populations that you seek to reflect and that we consider like moving towards more efficient models versus just moving to having just more data and having more of these models consuming it, 195 00:51:06,774 --> 00:51:19,676 because ultimately the more optimized models and optimized data sets can lead to high quality results versus just like large volumes of data. So I hope I answer your question. 196 00:51:23,796 --> 00:51:41,076 Thanks, Carla. Then I have a question, actually. I also put it in the chat. So certain governments control and limit the use of AI infrastructure and AI models for people in other countries due to national security considerations. 197 00:51:42,036 --> 00:51:56,356 So what can the international community do to avoid a new AI divide which comes then on top of the digital divide that already exists around the world? Maybe you can say a few things about that. 198 00:51:58,396 --> 00:52:21,571 I mean, there's definitely some amount of data divide. I mean, just between, open source models and non-oper source. And there's different models developed in some parts of the world. And we're seeing that the national security plus the trade restrictions are 199 00:52:21,571 --> 00:52:39,175 impacting who has access to what model, right? And a lot of times, we know that companies service customers all over the world, but there's we're seeing like in practice a lot more restrictions when it comes to 200 00:52:39,175 --> 00:52:56,039 access to data or access to these platforms from some countries. So that is happening and there is definitely competition in geopolitical lens embedded into this. But what can international organizations 201 00:52:56,039 --> 00:53:12,234 do? I think the first thing is definitely help advance alignment when it comes to like the principles that should shape the development of these models, because no matter what model you use, it ultimately, or you would want that it serves the same 202 00:53:12,234 --> 00:53:27,464 principles, you know, that it's preserves equity, that preserves, a protection, especially for high-risk uses, that preserves, the privacy of the users, that there's like education and support, like building to it, 203 00:53:27,464 --> 00:53:45,137 especially for vulnerable populations, for kids, as they use these tools. So you will want alignment on those principles. So that's one thing that organizations, international organizations and the global community can do. And there are examples around this. 204 00:53:45,137 --> 00:53:59,116 I mean, the OECD, the UN are doing great work in these areas, advancing standards. governments like the government of Japan with the Hiroshima AI process, like really advancing standards on how governments and industry can develop. 205 00:54:00,716 --> 00:54:17,082 these models and adopt them based on shared principles. And the other the other part is really trying to understand the application of the models and try to look into use cases and see where are the specific 206 00:54:17,082 --> 00:54:35,029 considerations and either positive or negative outcomes depending on the. on the specific use cases and try to develop, if not regulation, at least some level of self-regulation standards that can help ensure that 207 00:54:35,029 --> 00:54:50,236 ultimately the application of these models also protects humans and that innovation can be done in a way that is secure. They're always going to be that competition and that. 208 00:54:52,316 --> 00:55:06,956 complex set of providers, but we should ultimately aim that these solutions will have alignment when it comes to like core principles and that its applications will be closely monitored and assessed, especially if they have like a. 209 00:55:07,036 --> 00:55:15,196 high risk or high negative implications for any of the target population or populations that might be impacted from it. 210 00:55:17,476 --> 00:55:37,622 Thank you, Carla. Anyone else wants to ask a question? So please raise your hand. That's the easiest. I see Sean, my colleague Sean has a question. Please come in. thanks for the presentation. I just wanted to ask how you see the 211 00:55:37,622 --> 00:55:57,619 intersection between data sovereignty and the data privacy conversation that typically happens. And by that I mean Most NSOs are familiar with the fundamental notion of privacy. You protect individual responses. And the fundamental tension between 212 00:55:57,619 --> 00:56:15,590 privacy and statistical offices is that statistical offices are tasked with producing derivative inferences about broader populations, to infer things about broad groups of people that, on the basis of the specific individuals. And we'd like to protect not only 213 00:56:15,590 --> 00:56:29,204 confidentiality, but through these privacy frameworks. We like to protect the ability of anyone to draw information about individuals against the intention of the data collection. So, if it's a census, 214 00:56:29,204 --> 00:56:45,587 we use differential privacy to, you know, remove the effect of any one particular person from that data set. Now, in this model, the sovereign is the individual. Maybe I'm misunderstanding this, but essentially because they have, 215 00:56:45,587 --> 00:57:02,437 we're treating their claim over their personal identity as a fundamental right. Now, I understand that data sovereignty is also looking at a higher level at this issue of, well, upstream to this, you have the fact that if I collect data about a lot of people, 216 00:57:02,437 --> 00:57:16,681 I can determine population characteristics about people at large. I can put together models about individuals that can and make claims about them that they didn't intend to make. I might infer that you're pregnant because you go into a supermarket at 3 A. 217 00:57:16,681 --> 00:57:31,076 m. Sort of innocuous behavior becomes very identifiable behavior. And I wonder if this is part of the sovereignty conversation that might be missed. But I just was curious about how you draw the line between privacy and sovereignty. 218 00:57:31,076 --> 00:57:45,076 Is it that there are legitimate parties that like to make, you know, that have a claim over our data? Is it or do individuals fundamentally reserve all rights to their identity. So just an interesting. 219 00:57:45,916 --> 00:58:02,021 blow thoughts through my mind. I don't know if you have anything to say about that. Yes, that's a great question. And they're definitely intertwined in any aspects about privacy when it comes to, you know, the signing privacy, 220 00:58:02,021 --> 00:58:17,812 enhancing or preserving processes and the agency of the subject and, issues related to consent, effective consent as well, and effective right to withdraw, data. I mean, those questions still hold because, 221 00:58:17,812 --> 00:58:30,788 I mean, ultimately sovereignty means. making sure that data, no matter where it is, it doesn't have to be just in your country, no matter where it is, it will follow the same principles that it will follow in-house. 222 00:58:30,788 --> 00:58:49,494 So if your country is providing like all these privacy preserving guarantees, just like you said, like to ensure that data can be collected, but individuals cannot be identified and all the statistics can be produced with all the privacy preserving mechanisms. 223 00:58:49,494 --> 00:59:09,356 All of those principles hold. It's just that You do it, I mean, you're able to maybe host data somewhere else, maybe you're able to exchange data, and ultimately the concept of sovereignty conflict expands those protections beyond. 224 00:59:09,756 --> 00:59:25,116 maybe makes the privacy conversation even more complex. And because of the context in AI, not necessarily because the context of the sovereignty part of the conversation, just by introducing AI, privacy just becomes a 225 00:59:26,956 --> 00:59:41,716 a lot more complex because of the complexity sometimes of tracing how like training data has been used and the implications of the outcomes are coming from it. So I would say that they're definitely intertwined and 226 00:59:42,086 --> 00:59:59,426 And the privacy should remain at the core of, you know, data production processes and you also have different approaches to privacy, right? You have some systems that are more rights based on systems that are maybe more market based in allow. 227 00:59:59,746 --> 01:00:15,506 for more room towards like what private actors are collecting and doing with the data. So those conversations pre-AI still hold, the GDPR era and before conversations, but now with the sovereignty is about 228 01:00:15,506 --> 01:00:33,106 like ensuring that all these guarantees hold. no matter where you're hosting the data, no matter where the compute happens, and then ultimately you can ensure the privacy of your citizens or your communities and in the responsible use of data ultimately. 229 01:00:34,066 --> 01:00:51,862 thank you so much. I really appreciate that response. You're welcome. thank you. Thanks to everyone for all your questions. Unfortunately, we are at time now. But before we all go, maybe you can switch on quickly your camera and your microphones. 230 01:00:51,862 --> 01:01:09,180 We have a nice tradition here in the global network. Please follow and switch on your camera. Today maybe the crowd is a little bit shy. Well, a few faces pop up. So please join me and give Carla a big 231 01:01:09,180 --> 01:01:11,826 round of applause. Thank you. 232 01:01:13,746 --> 01:01:28,593 It's going to mute the noise, I think, with this silent clapping. Okay. Thank you. Yes, I think the noise cancelling kind of Okay. Okay. Anche o collo. filters out the applause. But in any case, you see a lot of happy faces here. Thanks, 233 01:01:28,593 --> 01:01:48,372 everyone, for joining. Thanks, Carla, for this great presentation, for your thoughts on this and your your inputs. I would like to invite everyone also to join our next two webinars in our AI series. So we have still two to go. 234 01:01:48,372 --> 01:02:06,026 You can see the registration links in the meeting chat and you have also received them. in the invitation to this webinar today. So thanks, everyone, and have a great rest of your day. Thank you. Bye, everyone. 235 01:02:06,026 --> 01:02:09,026 Thank you so much. Bye bye. Thank you, bye-bye.