Protecting data quality in the AI era of official statistics

As generative artificial intelligence (AI) gains traction in our economies and societies, a growing share of citizens, journalists and policymakers are accessing official statistics through chatbots and AI assistants rather than through the official websites of National Statistical Institutes (NSIs) or governments. With AI-mediation, the path to data from producer to user has lengthened considerably. Information now passes through a series of digital intermediaries that retrieve, interpret and redistribute it before it finally reaches its intended audience. Along the way something can be quietly lost: a reference period, a revision footnote, the methodological notes that allows the user to judge whether the figure responds to their particular question. As a result, OECD discussions on statistical quality are increasingly recognising that quality is no longer just a property of the data itself, but of the entire information chain.

AI is reshaping demand and usage of official statistics.

AI can substantially improve access to official statistics by making data easier to find and, possibly, easier to comprehend. Instead of engaging directly with tables or dashboards, users increasingly receive AI-generated summaries in plain language. While this can make statistics more accessible, particularly for non-expert users, this convenience comes with a trade-off: the underlying information has already been selected, interpreted and summarised by an AI system in ways the user almost never sees. A figure can arrive without the specificities that allows users to analyse the source, reference period, method and caveats. A revision footnote, for example, may not survive the AI journey, and, unless prompted by a request of the AI user, the institution that produced the figure may not appear at all. Users receive a large amount of information through experiences that are often light-hearted and positive, yet lack access to the context and underlying data structures necessary to evaluate the quality of that information and use it effectively.

AI has other impacts on the use of official statistics, take for example citizen-generated data (CGD), data that people or communities produce themselves through surveys they run, reports they file and platforms they use to document realities that official statistics may miss. For this type of data, AI intermediation raises significant issues of ownership, validation and representativeness. While AI processing can help amplify the reach of CDG, it can also absorb them at a scale where the citizen's intentional act of producing data dissolves into raw input (Lotfian et al., 2026).

The largest transformation of all may be the rise of simulated users, AI-generated representations of the public built from synthetic personas, behavioural models and automated testing by statistical offices. These tools are useful for anticipating demand and stress-testing communication, but a model of the public is not the public. Recent work shows that simulated respondents tend to flatten the diversity of real groups and to misrepresent minority voices (Wang, A., J. Morgenstern and J. P. Dickerson, 2025; Bisbee et al., 2024), confirming that simulation is a reliable tool for exploration, not representation.

Once AI mediates the relationship between data and users, quality is no longer just about the figure itself. It becomes a property of the entire data chain, raising a number of technical challenges.

• Traceability: AI systems have no built-in awareness of whether a source has been revised, corrected or withdrawn. A 2025 study asked ChatGPT to assess 217 retracted research articles, thirty times each (6 510 assessments in total) and not a single one mentioned the retraction (Thelwall et al., 2025). This may also apply to official statistics as a discontinued or superseded statistical series can continue to circulate downstream with its original authority excluded.
• Machine-readability : The reliability of an AI answer depends on how the underlying source is structured, with retrieval accuracy reaching 90.7% on well-structured policy text but falling to 85.6% on a less-structured corpus (Yun et al., 2024). The structural quality of what statistical offices publish now conditions the accuracy of what AI systems communicate on their behalf.
• Representativeness: General-purpose AI systems are trained on whatever data is most abundant on the open web, and abundance is not representativeness. A 2024 study comparing AI-generated synthetic survey responses with real human responses, in a deliberately favourable case, found that 48% of regression coefficients estimated on synthetic data differed significantly from those estimated on real data, and a third of those reversed direction (Bisbee et al., 2024). This highlights perhaps the most consequential issue, as representativeness which is the assurance that a figure speaks for the whole population, is precisely what AI systems struggle to provide and what official statistics are designed to deliver.

The risks can also  go beyond the technical aspects. Official statistics are a public good in the strict economic sense: they are non-rival, meaning one person's use does not diminish another's, and they are non-excludable, meaning access cannot be restricted. This is precisely why official statistics have traditionally been provided by the state rather than the market. Yet AI is beginning to place this second characteristic under subtle pressure. The content itself stays public, the figures remain free to access, but meaningful access is migrating behind premium assistants, paid interfaces and infrastructure that is highly concentrated. Today, approximately half of the world’s secure internet servers are hosted in a single country (World Bank, 2025). As a result, a public good can become excludable in practice, without any deliberate decision to exclude.

Adding to this challenge, AI systems are not neutral conduits of information. Every model embeds value judgements in what it highlights, what it simplifies and what it omits, choices that follow from how it was trained, not from any explicit decision. For producers of official statistics, the act of mediation has therefore become a quality issue in its own right. A figure can be perfectly accurate at source and still be distorted by the values of the system that conveys it (Schoeffer, J., M. De-Arteaga and N. Kühl, 2024; Hilliard et al., 2025).

These shifts converge on a question that has so far received no clear empirical answer: what happens to trust in official statistics when AI sits between the institution and its public? The effect could go either way. AI mediation may raise trust by delivering statistics in accessible and conversational language, making them easier to integrate into everyday decisions; or it may erode trust by detaching the figure from the institution that produced it. When a user no longer sees who is behind the number, the basis for trusting it changes. Which of these dominates in practice is one of the most consequential questions of the next few years for statistical data.

If quality now resides in the entire information chain, someone must take responsibility for safeguarding that chain. Statistical offices are uniquely positioned to take this role, extending their existing mandate. This means ensuring official statistics are sufficiently structured, documented and accessible to serve as trusted reference points for AI systems; and that they are able to monitor the “last mile” of dissemination to understand  how statistics are reformulated downstream once they leave their office. It also means setting standards for non-official data, including citizen-generated data, that are increasingly used in public decisions; and preserving the institutional independence that AI intermediaries can borrow from but cannot manufacture for themselves. Trust is the one asset private intermediaries can build on but cannot generate on their own, and protecting it is now a core part of what producing official statistics involves.

The word that best captures this expanded role for NSIs is stewardship. This is not about expanding the mandate for its own sake; it is about ensuring quality and trust in a changing information ecosystem. It seems to now be a necessary condition to continue performing the task statistical offices have always had, producing statistics that people can trust and  use when taking critical decisions.

The question raised by AI is not whether official statistics still matter. In an AI-mediated world, they matter more than ever.

Bisbee, J. et al. (2024), "Synthetic replacements for human survey data? The perils of large language models", Political Analysis, Vol. 32/4, https://doi.org/10.1017/pan.2024.5.

Hilliard, E. et al. (2025), Measuring AI Alignment with Human Flourishing, https://arxiv.org/abs/2507.07787.

Lotfian, M. et al. (2026), as cited in WISE analytical work on citizen-generated data and AI mediation, OECD.

Schoeffer, J., M. De-Arteaga and N. Kühl (2024), "Explanations, fairness, and appropriate reliance in human-AI decision-making", Association for Computing Machinery, New York, https://doi.org/10.1145/3613904.3642621.

Thelwall, M. et al. (2025), "Does ChatGPT ignore article retractions and other reliability concerns?", Learned Publishing, Vol. 38/4, https://doi.org/10.1002/leap.2018.

Wang, A., J. Morgenstern and J. P. Dickerson (2025), "Large language models that replace human participants can harmfully misportray and flatten identity groups", https://arxiv.org/abs/2402.01908.

World Bank (2025), Digital Progress and Trends Report 2025: Strengthening AI Foundations, World Bank, Washingtion, https://doi.org/10.1596/978-1-4648-2264-3.

Yun, X., Yun, J. & Xue, J. (2024), "Improving citizen-government interactions with generative artificial intelligence: Novel human-computer interaction strategies for policy understanding through large language models", PLOS ONE, https://doi.org/10.1371/journal.pone.0311410.