StatGPT：官方统计的人工智能

STATISTICS DEPARTMENT StatGPT: AI forOfficialStatistics Prepared by James Tebrake, Bachir Boukherouaa,Jeff Danforth, and Niva Harikrishnan INTERNATIONAL MONETARY FUND STATISTICS DEPARTMENT DEPARTMENTAL PAPER StatGPT:AI for OfficialStatistics Prepared by James Tebrake, Bachir Boukherouaa,Jeff Danforth, and Niva Harikrishnan1 Cataloging-in-PublicationDataIMFLibrary Names: Tebrake, James, author. | Boukherouaa, El Bachir, author. |Danforth, Jeff, author. | Harikrishnan, Niva, author. | International Monetary Fund, publisher. Title: StatGPT: AI for official statistics / James Tebrake, Bachir Boukherouaa, Jeff Danforth, Niva Harikrishnan.Other titles: AI for official statistics. | Departmental paper. | International Monetary Fund. Statistics Department.Description: Washington, DC : International Monetary Fund, 2026. | Includes bibliographical references.Identifiers: ISBN: The authors thank colleagues from across the official statistics community for testing StatGPT and for providingvaluable feedback that helped improve and enrich the platform. They also thank colleagues from the StatisticalData and Metadata Exchange sponsor agencies who actively participated in testing and offered constructiveinput at various stages of the development. The authors further acknowledge the contributions of staff fromthe IMF’s Information Technology Department for their technical support during the development and imple-mentation, as well as EPAM staff members who provided technical expertise and implementation support. Anyremaining errors or shortcomings are the responsibility of the authors. The Departmental Paper Series presents research by IMF staff on issues of broad regional or cross-countryinterest. The views expressed in this paper are those of the author(s) and do not necessarily represent theviews of the IMF, its Executive Board, or IMF management. ABSTRACT:National statistical systems generate the statistics that underpin policy, economic analysis, and publictrust. Yet, despite decades of investment in statistical capacity, two persistent challenges, data accessibility andinterpretability, limit the impact of these official statistics. The rise of large language models and generative artifi-cial intelligence (GenAI) applications such as ChatGPT and Gemini appeared to offer a solution by enabling usersto retrieve statistics using natural language. However, testing demonstrates that although the GenAI applicationsexcel at synthesizing text, they perform poorly at delivering official statistics: they frequently provide dangerously“reasonable” but incorrect figures. This paper introduces StatGPT, an initiative by the IMF Statistics Department thatleverages large language models not to generate statistics, but to generate structured queries against applicationprogramming interfaces of official statistical agencies. StatGPT ensures that users receive the exact publishedfigures, every time, while benefiting from natural language interaction. This paper examines the limitations ofoff-the-shelf GenAI applications, outlines how StatGPT overcomes these limitations, and proposes a road map formaking official statistics AI ready through open data access, enriched metadata standards, and strengthened datagovernance. By aligning technological innovation with statistical rigor, StatGPT represents a critical step toward afuture where official statistics remain authoritative, trusted, and universally accessible in an AI-driven world. Publication orders may be placed online or through the mail:International Monetary Fund, Publication ServicesP.O. Box 92780, Washington, DC 20090, USAT. +(1) 202.623.7430publications@IMF.orgIMFbookstore.orgelibrary.IMF.org Contents Executive Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .v1. Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .12. Current Dissemination Challenges Facing the Official Statistics Community. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .23. The Potential of GenAI to Address These Challenges. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .84. StatGPT—AI for Official Statistics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .185. Making Official Statistics AI Ready. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .25References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .