Better Data for AI – A Possible Task

Esperanza Magpantay argues that the real limiting factor for AI is not algorithms or computing power, but the quality of the data these models learn from. Official statistics, produced with transparent methodologies and internationally agreed definitions, represent one of the few truly trusted global sources of evidence. Without such characteristics, AI systems risk amplifying biases and reinforcing inequalities.

She noted that for decades, national statistical offices and international organisations have developed rigorous standards, including transparent methodologies, internationally agreed definitions, quality assurance frameworks, and strong confidentiality protections . She further illustrated this with ITU's work on measuring digital development, connectivity, and digital inclusion, arguing that AI systems will only be useful if grounded in high-quality official statistics .

Magpantay emphasises that the goal is not to replace official statistics with AI or to substitute surveys with new data sources, but rather to combine them. By integrating trusted official statistics with responsibly governed novel data sources, the result is richer, more timely, and more relevant data for decision-making. This combination approach is presented as the key lesson from existing experience.

She referenced the UN-wide initiative on mobile phone data as an example of integrating new data sources with traditional statistical systems, noting these sources provide more timely and granular data while following governance and transparency measures . She explicitly stated that the future is about combining trusted official statistics with responsibly governed new data sources .

Magpantay argues that partnership is becoming essential because no single entity can build a trusted data ecosystem alone. Collaboration across governments, national statistical offices, regulators, international organisations, academia, the private sector, and civil society is required. This multi-stakeholder approach is presented as a prerequisite for AI that serves the public good.

She stated that partnership is becoming very essential because work cannot proceed without governments, national statistical offices, regulators, international organisations, academia, the private sector, and civil society . She also described ITU's work with experts in the UN Initiative on Mobile Phone Data as an example of bringing partners together to develop methodology and help countries use new data sources .

Magpantay describes a model where national statistical offices do not pay mobile operators directly for their data, but instead offer incentives such as technical skills in data quality assurance and access to statistical data that operators do not otherwise have. This approach is being implemented across 25 countries through a World Bank project. The aim is to integrate mobile phone data sustainably into official statistics.

She described a project implemented in 25 countries with the World Bank, where the objective is to integrate mobile phone data into official statistics sustainably by having all stakeholders work together . She explained that national statistical offices can provide incentives such as technical skills and access to their own data, rather than paying operators directly for data . She also confirmed that Côte d'Ivoire is one of the 25 countries benefiting from this technical assistance .

Anu Peltola frames the central challenge of the session: AI is only as trustworthy as the data it uses and shares. Trustworthiness and well-documented, accessible data are not merely technical issues but essential foundations for reliable, accountable, and beneficial AI tools. Official statistics, produced according to internationally agreed standards with continuous public oversight, are positioned as a key source of authoritative evidence.

She noted that much of the data used to develop AI systems is fragmented, uneven in quality, insufficiently documented, and often inaccessible for verification, leading to concerns about bias, transparency, and reproducibility . She highlighted that official statistics are produced according to internationally agreed standards and methodologies with continuous public oversight .

Peltola points out that governments have already made a formal commitment to investing in their national statistical systems and data through the Financing for Development outcome document. This political commitment provides a foundation upon which further action can be built, though its practical implementation varies by country. The international community is working to support the realisation of this commitment.

She referenced the outcome document of the Financing for Development conference, noting that governments committed to investing in their national statistical systems and data, and acknowledged that the extent to which this is put into practice depends on the country .

Peltola argues that a metadata platform such as the Trusted Data Observatory would not only make data findable but would also reveal where data gaps exist. This visibility into gaps could help direct investment and action towards areas where information is most critically needed for policy. The connection between metadata, gap identification, and targeted investment is presented as a practical benefit of such platforms.

She observed that if the TDO held metadata on what data are available, it would also reveal data gaps, which could help target action to where gaps are most critical for policy needs, thereby helping to connect investment with identified gaps .

Benjamin Rothen explains that the TDO is a global discovery platform designed to make trusted data findable by both machines and humans, without centralising the data or transferring ownership. National metadata platforms are linked to a global metadata platform, allowing users to discover where trusted data resides. The data itself remains with its original custodians.

He described the TDO as a national metadata platform linked to a global metadata platform, emphasising that data stays where it is and ownership is not transferred, but a global discovery platform is created so machines and people know where to find trusted data . He noted that Switzerland has had a metadata platform for ten years where all government data comes together so machines and humans can find the right data .

Rothen argues that a critical problem is not a lack of data but a lack of visibility: trusted data is often not findable, not interoperable, and not harmonised. Investing in metadata — descriptions of what data exists and what it means — is essential because large language models and machines search for text, not raw numbers. Without good metadata, machines cannot locate or use the underlying data.

He stated that the problem is often not a lack of data but that trusted data is not visible, not findable, and not interoperable . He emphasised that if organisations do not invest in metadata, data sets will never be found, because LLMs search for text data and not raw figures, making good textual descriptions essential .

Rothen highlights that the work of national statistical offices has changed fundamentally in the past two to three years, requiring organisations to restructure how they operate and present their data. Switzerland is redesigning its entire website to be AI-ready so that machines can find data more effectively. This is described as a significant investment and a major organisational challenge.

He noted that Switzerland is changing its webpage completely in October so it is AI-ready, enabling machines to find data much better, describing this as a huge task and a significant investment . He also acknowledged that Switzerland has been working for ten years on a metadata platform where all government data comes together for both machines and humans, and that this remains a difficult task .

Rothen acknowledges the importance of open government data and FAIR principles as foundational steps, but argues that metadata platforms are needed at an even earlier stage to describe what data exists. Even data that is not openly accessible should be described in metadata so that governments and researchers know it exists and can begin conversations about access. This approach helps governments understand their own data landscape.

He described five steps to reaching a level of open government data publication and noted that the TDO tries to start much earlier by identifying what data exists and how different datasets can be linked . He explained that even micro-level data that is not openly accessible should have a description in the metadata platform, so governments can discover what data other ministries hold and begin conversations about access .

Rothen describes the TDO as a collaborative initiative led by the Swiss government, involving around 20 countries from across the world and 30 to 40 international organisations. The platform is intended to be Geneva-based, drawing on the existing concentration of international knowledge in Geneva and expanding globally. He invites further participation and investment to help build the platform.

He stated that approximately 20 countries and 30 to 40 international organisations are working with Switzerland on the TDO, with the platform intended to be Geneva-based to bring together existing knowledge and expand globally . He expressed the ambition to show a prototype at the AI summit in Geneva in 11 to 12 months and invited all present to be part of the initiative .

Vibeke Østreich-Nielsen identifies a structural problem in how data and statistics are funded: investment tends to be sectoral and project-based rather than systemic, and donor funding for the Global South is often not linked to official statistics or long-term capacity building. This fragmentation means that data systems lack the continuity needed to develop sustained statistical capacity. The focus is often on meeting specific data needs rather than making data broadly available.

She observed that because statistics and data are a cross-cutting field, investment is often sectoral and not always linked to official statistics, with the focus being on specific data needs rather than making data available for countries generally or for AI purposes . She noted that this is part of the broader challenge facing data and statistics ecosystems .

Østreich-Nielsen describes the SEVIA Platform for Action, developed in the context of the Financing for Development agenda, which produced four recommendations to operationalise commitments to data and statistics. These recommendations cover treating data as a public good, improving national coordination, strengthening data governance and innovation, and investing in national statistical capacity. The first recommendation explicitly includes making data available in AI-ready formats.

She outlined the four recommendations: treating data and statistics as a public good, improving national data and statistical system coordination, strengthening data governance and innovation, and investing in national data and statistics capacity . She noted that the first recommendation includes an end goal of publishing data and making it available in AI-ready formats, which remains a major challenge .

Østreich-Nielsen, speaking from the perspective of a donor agency, argues that donor organisations need to develop more practical approaches for how they evaluate and decide on data-related projects. This includes ensuring that data work funded through donor projects is ultimately made available in AI-ready formats. She presents this as a concrete step the donor community can take to improve the data ecosystem.

She described work from a donor agency context to develop practical approaches for representatives in donor organisations on how they review and decide on projects, particularly regarding data work . She noted that this is intended to help ensure funded data work contributes to making data available in AI-ready formats .

Østreich-Nielsen highlights that a significant untapped resource exists in the form of administrative data held by government ministries that is not being made publicly available. This data could provide valuable information for both decision-makers and AI systems, particularly in African contexts where very little data is publicly accessible. Making this data available is identified as a major challenge and priority.

She noted that many ministries sit on relevant administrative data that are not being made available, but that could bring a lot of information to decision-makers and in an AI context, particularly in an African context where very little data is publicly available .

Daniel Power uses the example of the May 2023 Ebola outbreak in eastern DRC to demonstrate the practical value of mobile phone data in humanitarian crises. By tracking the movements of subscribers who had been in the outbreak area, Flowminder was able to rank health zones by their connectivity to the outbreak and predict where Ebola was likely to spread. This analysis proved highly accurate within a week of publication.

He described how Flowminder, using data from Vodacom Congo through an eight-year partnership, conducted a rapid cohort study identifying all subscribers who had been in the outbreak area and tracking which health zones they visited in subsequent days and weeks . He noted that a week after releasing their first report, Ebola was detected in ten more health zones, eight of which were among the top regions Flowminder had identified using mobility data .

Power acknowledges that mobile phone data, while valuable, is not representative of the whole population and systematically underrepresents certain groups. This bias must be addressed by combining mobile data with survey data to adjust for these imbalances before the data is used in AI systems. He argues this is critical both for general use and even more so when data feeds into AI systems that may lack the discretion to account for such biases.

He stated that mobile phone data underrepresents women, children, the very old, the more rural, and the less wealthy, and that many data types privilege men and the wealthy . He noted that when releasing monthly data on population distribution, it is necessary and critical to run surveys to understand how representative the data is and to adjust for bias, and that this applies equally or more strongly when data is used for AI .

Power argues that protecting the privacy of individuals who contribute data to mobile systems is essential, particularly given how much information subscriber data contains about individual movements. Flowminder's approach is to deploy software on the mobile operator's premises that aggregates data before it is exported, meaning individual-level data never leaves the operator's systems. This is presented as a model for responsible data use in AI contexts.

He explained that Flowminder always processes data at the mobile operator's systems, so individual-level data did not leave the Congo and did not leave Vodacom's premises; instead, software deployed on their premises aggregates the data before export . He emphasised the importance of protecting the privacy of people contributing data to AI systems .

Power highlights that having the support of regulators is critical for legitimising and enabling the responsible use of mobile phone data, particularly in complex country contexts. Long-term relationships with both mobile network operators and regulators provide the foundation for accessing and using such data responsibly. Clear governance principles and frameworks are presented as prerequisites for this work.

He noted that having a signal from the regulator ARP in DRC that they were comfortable with Flowminder's work was really important for navigating the complex environment of that country . He also described the importance of understanding operator needs and building long-term relationships, as well as the role of the World Bank's global data facility in Côte d'Ivoire .

Power describes the genuine tension that arises when negotiating access to mobile operator data, as operators are often told their data is a source of revenue to be monetised, while governments may have hard lines against paying for data. He notes that paying for data does not necessarily correlate with better data quality, and that his organisation takes a pragmatic approach depending on the country and operator. The considerations are complex and vary significantly by context.

He described the tension between governments' reluctance to pay for data and operators' expectations of monetisation, noting that in the Congo, Flowminder pays only a small fee for server management rather than for the data itself . He observed that paying for data does not necessarily correlate with data quality, citing cases where organisations paid for data that turned out to be of poor quality .

Power emphasises that Flowminder's ability to respond rapidly to crises such as the DRC Ebola outbreak was made possible by an eight-year partnership with Vodacom Congo. Similarly, having regulator support is critical for navigating complex country environments. These long-term relationships are presented as essential infrastructure for responsible mobile data use.

He noted that Flowminder's rapid response to the DRC Ebola outbreak was enabled by a longstanding eight-year partnership with Vodacom Congo . He also described how support from the regulator ARP in DRC was really important for legitimising their work in a complicated country context involving conflict and other considerations .

Alexandre Barbosa argues that nationally representative ICT statistics are essential for designing, implementing, and evaluating public policies across multiple domains including digital inclusion, education, and health. Brazil's CETIC model produces data aligned with international standards and SDG frameworks. The availability of this data in multiple languages further enhances its accessibility and utility.

He stated that nationally representative ICT statistics from CETIC support the design, implementation, and evaluation of public policies on digital inclusion, education, health, and several other dimensions, and are aligned with SDG dimensions . He noted that data are available in Portuguese, English, and some in Spanish, accessible via their website and reports .

Barbosa presents Brazil's CETIC as an example of a self-sustainable financing model for ICT statistics, funded entirely by revenues from the .br country code domain registry rather than government budgets or donor funding. This model has enabled 20 years of continuous ICT surveys. The Brazilian Internet Steering Committee, a multi-stakeholder body, oversees this arrangement.

He explained that CETIC has been running continuous ICT surveys for 20 years through a self-sustainable financing mechanism funded by the .br country code domain registry, managed by the Brazilian Internet Steering Committee and NIC.br, with surveys 100% funded by registry activities . He noted that Brazil ranks sixth largest domain name database among G20 and OECD countries, providing sufficient financial resources to fund the surveys .

Barbosa identifies three major advantages of the CETIC financing model: stability enabling continuous survey programmes and long-term time series; independence from government budgets and individual donors, ensuring methodological consistency; and innovation, allowing continuous updating of surveys to address emerging technologies. Stable financing also enables the development of innovative data collection methods such as big data and mobile phone data.

He outlined three advantages of the model: stability enabling continuous survey programmes and long-term statistical series for monitoring digital transformation ; independence from public budgets and individual donors, maintaining professional independence and methodological consistency ; and innovation, allowing continuous updating of surveys and development of alternative data sources including big data and mobile phone data .

The GIZ Egypt Representative raises the practical challenge that many countries, including Egypt, face foundational preconditions that must be met before AI-ready data can be produced. These include digitising existing data, enabling government-to-government data sharing, and creating incentives for both public and private sector actors. Research suggests these preconditions must be in place before more advanced data governance and AI data initiatives can succeed.

She described Egypt's recent open data policy and GIZ's efforts to support its implementation, noting that research she encountered identified many preconditions that need to be in place first, including digitising existing data, enabling government-to-government communication, creating connected platforms, and incentivising government and private sector actors .

Jacques Péguet raises the question of downstream payment for data, asking whether there is any consideration of charging for better data or whether this is ruled out for political reasons. This question touches on the tension between treating data as a public good and the practical need to fund its production and maintenance. The question invites reflection on sustainable financing models for high-quality data.

The Côte d'Ivoire Representative expresses interest in learning more about the CETIC model and how it was established, reflecting a broader demand from countries in the Global South for practical guidance on replicating successful data integration and financing models. This highlights the need for knowledge sharing and technical assistance across countries seeking to build similar capacities.

The representative from Côte d'Ivoire expressed interest in learning more about the CETIC model and how it was set up , prompting Esperanza Magpantay to confirm that Côte d'Ivoire is one of the 25 countries already benefiting from the World Bank mobile phone data integration project and to invite further discussion .

The Moderator argues that it is not enough to produce statistics and numbers using the same definitions and procedures as before, because the nature of information in society has changed over time. Innovation must be built into the entire process of data production, not treated as an optional add-on. This point is made in response to Alexandre Barbosa's discussion of financing and innovation.

The Moderator noted that what we understand as information in society has changed over the years, and that statistics cannot simply be produced using the same definitions and the same procedures, emphasising that innovation must be incorporated into all aspects of data production .

The Moderator raises the question of what concrete progress would look like at the next summit or conference, pushing the discussion beyond general principles towards specific, demonstrable outcomes. This framing encourages panellists to identify tangible deliverables rather than aspirational goals. The TDO prototype at the Geneva AI summit is cited as one such concrete milestone.

The Moderator asked what a single concrete piece of progress in this area would look like at the next summit or conference, prompting Benjamin Rothen to identify the TDO prototype at the AI summit in Geneva in 11 to 12 months as a key deliverable .

The Moderator synthesises the session's key messages by emphasising that no single actor can deliver trusted data for AI and that partnerships are indispensable. She also stresses that while AI tools are powerful and should not be underestimated, the results they produce must be critically evaluated, because data shapes debates and decisions and it matters what numbers underlie any analysis.

The Moderator concluded the session by stating that no single actor can deliver this, that partnerships and different types of data are needed, and that AI tools are powerful but require good data to ensure robust results, adding that it is easy to produce outcomes with AI tools but that one must be very critical about them .

The Moderator commits to carrying the discussion forward within the UN system and the broader network of chief statisticians, framing this as the appropriate institutional channel for scaling up and connecting with partners. This reflects a recognition that the issues raised require coordinated international action beyond the session itself.

The Moderator stated that she would take the discussion back to the UN system and the broader network of chief statisticians so that the international community can consider how best to advance this agenda to scale and connect with partners .

The Moderator draws attention to the fact that data serving the public good is not exclusively held by public institutions; privately held data, such as mobile operator data, can be equally important for societies and decision-makers. She also highlights the value of such data for anticipating future needs, as demonstrated by the Ebola outbreak example.

Following Daniel Power's presentation on the DRC Ebola outbreak, the Moderator observed that public good data can also be privately held and can be equally important for societies and decision-makers, and can help foresee what will be needed .