## StatGPT: AI for Official Statistics (Executive Summary)

## Source details

**Canonical URL:** [StatGPT: AI for Official Statistics (Executive Summary)](https://www.imf.org/-/media/files/publications/dp/2026/english/saiosea.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/dp/2026/english/saiosea.pdf.md)
- [Structured JSON version](/-/media/files/publications/dp/2026/english/saiosea.pdf.json)

---

### Context and problem statement
- Official statistics are fundamental for shaping evidence-based policies, monitoring economies, and fostering public trust (United Nations Statistics Division 2015).
- Despite substantial investment, official statistics remain difficult for many users to find, interpret, and apply (IMF, Statistics Department 2018).
- Users face challenges ranging from navigating multiple websites to understanding technical metadata, often leading to frustration, data misuse, or reliance on secondary sources.
- The emergence of generative artificial intelligence seems to promise a breakthrough: conversational access to official data.

### GenAI limitation identified
- Accuracy testing reveals a major flaw: Generative artificial intelligence returns inaccurate numerical results up to two-thirds of the time, eroding trust and risking the spread of misinformation.
- With numerical information, “close” is not acceptable: numerical information returned by GenAI applications is either correct or wrong; GenAI often returns a reasonable estimate that is not the published number, creating risk because users may not fact-check plausible but incorrect figures.

### StatGPT—approach and architecture
- StatGPT interprets user prompts and translates them into structured queries against application programming interfaces (APIs) maintained by official statistical agencies.
- StatGPT does not generate statistics; it retrieves and presents the precise published estimates.
- Architecture combines LLM language capabilities with retrieval of official statistics from national statistical authorities to address the trust deficit.
- StatGPT displays the query parameters in plain English before executing the query, allowing the data consumer to confirm the parameters.

### Required actions for official statistical producers (three critical steps)
- Open their data and provide API access. Without robust APIs, AI tools cannot reliably retrieve the “published” number, leading to guesswork and inaccuracies.
- Enrich data models and metadata—making them AI ready. Standards like Statistical Data and Metadata Exchange provide a strong foundation, but more detailed, harmonized metadata are needed to ensure AI systems can map natural language requests to the correct indicators.
- Strengthen data governance and ownership. Clear attribution must be maintained, with AI systems directed to retrieve data from the primary owner whenever possible.

### Broader strategy and infrastructure
- Establishment of the Global Trusted Data Commons: a global index of AI-ready official statistics designed to serve as a bridge between GenAI applications and authoritative data sources.
- Backbone standard: Statistical Data and Metadata Exchange (SDMX) to ensure harmonization and interoperability.
- Intended outcome: all GenAI applications, including StatGPT, ChatGPT, and Gemini, will be able to interface directly with the Global Trusted Data Commons to reliably access and present official statistics.

### Empirical accuracy tests of LLMs for structured data retrieval (WEO G7 growth test)
- Experimental design:
  - Prompt: “Can you generate a table of economic growth rates for the G7 countries taking the data from the latest issue of the IMF’s World Economic Outlook. Can you provide data for 2018 to 2025. Can you provide the output in a CSV file.”
  - Performed 10 times in the same conversation, 10 times in 10 different conversations, and 5 times in the same conversation with a copy of the latest WEO loaded in memory (total of 25 prompts).
  - Comparison benchmark: April 2025 version of the IMF World Economic Outlook.
  - Total data points compared: 7 countries × 25 iterations × 8 years = 1,400 data points.

- ChatGPT (OpenAI 2025a) overall accuracy:
  - 34 percent correct when prompts were entered into the same conversation.
  - 17 percent correct when the request was made using unique conversations.
  - 14 percent correct when the latest WEO publication was loaded into ChatGPT.

- Year-level and country-level notes:
  - 2025: 79 percent of the responses were correct (using 10 continuous prompts).
  - United States: 93 percent correct (single conversation, 10 prompts).
  - Germany: 0 percent correct (continuous-conversation test reported).

- Gemini (Google DeepMind 2025) results:
  - Five iterations attempted.
  - Out of 280 data points (7 × 8 × 5), Gemini provided 35 correct responses = 13 percent correct.
  - Gemini returned identical responses across prompts in the test but with a higher error rate than ChatGPT.

### Detailed reported counts and percentages (selected findings)
- ChatGPT responses in a single conversation versus WEO (aggregated totals):
  - Total Right: 189
  - Total Wrong: 371
  - Total Possible (sample): 560
  - Percent Correct (overall single conversation): 34%
  - Country highlights (Percent Correct, single conversation):
    - Canada: 35%
    - France: 25%
    - Germany: 0%
    - Italy: 48%
    - Japan: 24%
    - United Kingdom: 13%
    - United States: 93%

- ChatGPT using 10 different conversations versus WEO (aggregated totals):
  - Total Right: 94
  - Total Wrong: 466
  - Total Possible: 560
  - Percent Correct (overall unique conversations): 17%
  - Country highlights (Percent Correct, unique conversations):
    - Canada: 16%
    - France: 10%
    - Germany: 0%
    - Italy: 15%
    - Japan: 16%
    - United Kingdom: 8%
    - United States: 53%

- ChatGPT in same conversation with latest WEO uploaded (aggregated totals):
  - Total Right: 40
  - Total Wrong: 240
  - Total Possible: 280
  - Percent Correct (with WEO uploaded): 14%
  - Country highlights (Percent Correct, WEO uploaded):
    - Canada: 13%
    - France: 0%
    - Germany: 13%
    - Italy: 13%
    - Japan: 13%
    - United Kingdom: 25%
    - United States: 25%

- Gemini responses using a single conversation versus WEO (aggregated totals):
  - Total Right: 35
  - Total Wrong: 245
  - Total Possible: 280
  - Percent Correct: 13%
  - Country highlights (Percent Correct, Gemini single conversation):
    - Canada: 25%
    - France: 13%
    - Germany: 13%
    - Italy: 13%
    - Japan: 0%
    - United Kingdom: 13%
    - United States: 13%

### Average absolute percentage point differences reported
- Table 6 (ChatGPT, 10 prompts in a single conversation versus WEO):
  - Average Difference—All: 0.20.30.30.90.50.30.30.0
- Table 7 (ChatGPT, 10 prompts in unique conversations versus WEO):
  - Average Difference—All: 0.50.62.72.71.30.40.50.5
- Interpretation:
  - On average, the difference in 2025 is less than 0.1 percentage points.
  - Despite small average differences, LLM outputs are only 79 percent accurate in 2025 (single-conversation result), far short of the official statistics community’s demand for 100 percent precision, and the outputs appear reasonable—raising concern.

### External corroboration and model evolution
- Market Research Society (MRS 2025) tests (June 4 and 5, 2025) across 10 different LLMs: estimates generated by LLMs often appear credible but are frequently inaccurate and inconsistent across runs and platforms; offline models relying solely on internal data show particularly unreliable performance.
- Tests cited were conducted in August 2025, before OpenAI’s 5.0 LLMs.
- ChatGPT 5.0 behavior on same prompt: did not return requested data and instead stated inability to locate a full, clean IMF WEO time series for 2015–2025 for all the G7 countries, describing the IMF’s WEO Data Mapper requirement for individual retrieval (OpenAI 2025b) — characterized as a new guardrail preventing false citation and erroneous structured-data outputs.

### StatGPT test results and empirical findings
- StatGPT test (DIAL X (EPAM Systems) 2025b) using five iterations:
  - The query in one conversational example returned 35 series.
  - StatGPT returned the same data as published by the IMF (by design).
- Aggregate test outcomes (StatGPT Responses Using 5 Prompts in Different Conversations versus Actual World Economic Outlook):
  - Total Right: 280
  - Total Wrong: 0
  - Total Responses: 280
  - Percent Correct: 100%
  - Country-level entries (Canada, France, Germany, Italy, Japan, United Kingdom, United States):
    - For each country: counts per year 2018–2025 were 5, 5, 5, 5, 5, 5, 5, 5
    - Total Right per country: 40
    - Total Wrong per country: 0
    - Total Responses per country: 40
    - Percent Correct per country: 100%

### Operational example and capabilities
- Typical flow for “economic growth and inflation indicators for G7 countries”:
  - Data set agent lists datasets in StatGPT’s global index with relevant indicators (examples: StatCan, National Balance Sheet Accounts; ESTAT, GDP, and main components; WB, World Development Indicators (WDI); IMF, World Economic Outlook (WEO)).
  - Beta global index datasets included: FRED (FOR), StatCan (National Balance Sheet Accounts), IMF (WEO), ESTAT (GDP and main components), OECD (TiVA 2023), WB (WDI), ECB (Balance Sheet Items), BIS (Debt securities statistics).
  - LLM recommends WEO when the user requests forecasts for 2025 because WEO contains forecasts.
  - User refines time period to 2010–2025; null values returned where data do not exist.
  - User narrows series to CPI Period Average and GDP Constant Price Percent Change; StatGPT executes the query, retrieving two indicators for 2010–2025 for G7 countries.
  - The conversation constructs a URL query passed to the IMF API to retrieve the data as published by the IMF.
- StatGPT can query multiple data catalogs, combine outputs, cite each indicator retrieved, classify data producers as data owners or data providers, transform data per user specification, and present original and transformed data plus the transformation script.

### Making official statistics AI-ready: standards, metadata, ownership, and APIs
- Two perspectives on responsibility:
  - Software providers should ensure their AI apps and LLMs deliver fully accurate results when querying official data.
  - Official statistics producers should structure and publish data in formats accessible and understandable by AI systems.
- IMF authors assert primary responsibility lies with official statistics producers to ensure correct access to official statistics.
- Three pillars for AI-readiness:
  - A robust, forward-looking AI-ready data standard tuned to GenAI so GenAI applications can accurately map user intent to authoritative statistics.
  - Full implementation of the standard, enriching metadata and referential data sets so GenAI applications and LLMs can reliably interpret queries through the Global Trusted Data Commons.
  - Open APIs to make data resources accessible, empowering GenAI platforms to query the Global Trusted Data Commons directly and deliver precise results as published.

- SDMX as foundation:
  - SDMX provides a globally recognized framework for defining data structures, dimensions, and metadata attributes.
  - Metadata must be complete, precise, harmonized, and semantically meaningful—not merely technically compliant.
  - Enrichment should include full indicator descriptions, clear definitions, time coverage notes, methodological explanations, and links to source documentation.
  - With enriched metadata, StatGPT can map “inflation in France” to the correct SDMX-coded series such as “Consumer Price Index, All Items, Period Average”.

- Data ownership and governance:
  - Explicit ownership must be maintained throughout the data lifecycle to preserve attribution and accountability.
  - Embed ownership into metadata and API design: StatGPT should cite both the publisher and the data owner or creator.
  - AI systems should be directed to query the owner’s API whenever possible to ensure users receive up-to-date “first source” data.

- APIs and machine accessibility:
  - Static downloads, spreadsheets, and interactive human-facing tables limit reliable AI access.
  - APIs allow structured queries specifying indicator, geography, and time period to ensure the published figure is delivered every time.
  - Evidence: IMF’s World Economic Outlook API enables AI to retrieve authoritative numbers; many other agencies lack equivalent infrastructure.

### Strategic orientation, governance, and collaboration
- StatGPT is positioned as an integration tool and a freely available public good to be used, adapted, and extended by any organization engaged in information or data dissemination.
- The official statistics commons is a small subset of the broader global data commons; StatGPT’s strategy is integration, not competition.
- Successful AI application in official statistics requires collaboration:
  - Software providers must refine models and software to interpret, retrieve, and present data faithfully.
  - Producers must prioritize data quality, standardization, and machine-readable formats and models (e.g., SDMX).
- Leadership and sponsorship:
  - SDMX sponsoring organizations (BIS, ECB, Eurostat, IMF, OECD, UN, World Bank) are well placed to guide modernization.
  - A joint statement by SDMX sponsor organizations commits to provide strategic direction for AI-readiness of official statistics.
- To build a Global Trusted Data Commons:
  - Partner with national statistical offices, legal experts, technology leaders, and public and private data providers.
  - Implement collaborative governance, experiment with technologies such as model context protocols, and engage with the UN Statistical Commission’s AI-readiness agenda.

*IMF DEPARTMENTAL PAPERS • StatGPT: AI for Official Statistics (Executive Summary)*

### Executive Summary.................................................................................................v

### Executive Summary

### Context and problem statement
- Official statistics are fundamental for shaping evidence-based policies, monitoring economies, and fostering public trust (United Nations Statistics Division 2015).
- Despite substantial investment, official statistics remain difficult for many users to find, interpret, and apply (IMF, Statistics Department 2018).
- Users face challenges ranging from navigating multiple websites to understanding technical metadata, often leading to frustration, data misuse, or reliance on secondary sources.
- The emergence of generative artificial intelligence seems to promise a breakthrough: conversational access to official data.

### GenAI limitation identified
- Accuracy testing reveals a major flaw: Generative artificial intelligence returns inaccurate numerical results up to two-thirds of the time, eroding trust and risking the spread of misinformation.
- With numerical information, “close” is not acceptable: numerical information returned by GenAI applications is either correct or wrong; GenAI often returns a reasonable estimate that is not the published number, creating risk because users may not fact-check plausible but incorrect figures.

### StatGPT—approach and architecture
- The IMF Statistics Department is developing StatGPT, a new approach that redefines how AI engages with official statistics.
- StatGPT interprets user prompts and translates them into structured queries against application programming interfaces (APIs) maintained by official statistical agencies.
- StatGPT does not generate statistics; it retrieves and presents the precise published estimates.
- This architecture combines AI’s language capabilities with the precision of retrieving official statistics from national statistical authorities, addressing the trust deficit.

### Required actions for official statistical producers (three critical steps)
- Open their data and provide API access. Without robust APIs, AI tools cannot reliably retrieve the “published” number, leading to guesswork and inaccuracies.
- Enrich data models and metadata—making them AI ready. Standards like Statistical Data and Metadata Exchange provide a strong foundation, but more detailed, harmonized metadata are needed to ensure AI systems can map natural language requests to the correct indicators.
- Strengthen data governance and ownership. Clear attribution must be maintained, with AI systems directed to retrieve data from the primary owner whenever possible.

### Broader strategy and infrastructure
- The StatGPT strategy calls for the establishment of the Global Trusted Data Commons, a global index of AI-ready official statistics designed to serve as a bridge between GenAI applications and authoritative data sources.
- The backbone of this Commons is the Statistical Data and Metadata Exchange (SDMX) standard, ensuring harmonization and interoperability.
- It is envisioned that all GenAI applications, including StatGPT, ChatGPT, and Gemini, will be able to interface directly with the Global Trusted Data Commons to reliably access and present official statistics.

### Expected outcomes and call for collaboration
- By addressing the identified dissemination and governance challenges, StatGPT has the potential to transform access to official statistics, making them more discoverable, understandable, and trustworthy.
- The approach seeks to safeguard the authority of the agencies that produce official statistics and to ensure official statistics remain the definitive source of truth in an increasingly AI-driven information landscape.
- The paper positions StatGPT as both a technological innovation and a call for global collaboration among data producers, standard bodies, and AI developers.

*IMF DEPARTMENTAL PAPERS • StatGPT: AI for Official Statistics (Executive Summary)*

### 3. The Potential of GenAI to Address

### 3. The Potential of GenAI to Address These Challenges

### GenAI as a data access solution
- GenAI can retrieve specified data from specified sources in user-specified formats via prompting.
- Example: ChatGPT (4o model) produced a table of economic growth rates for the G7 from the IMF’s World Economic Outlook in 5 seconds (OpenAI 2025a).
- Traditional search, discover, and extract methods took much longer; a stepwise manual retrieval path from Google to IMF Data Mapper took 15 minutes; direct navigation to data.imf.org and using the IMF Data Explorer took 2 minutes and 40 seconds (Table 1).

### Time-to-data comparisons (from Table 1)
- Time when initiating the search from Google: 15 minutes (total manual path: 18 minutes).
- Time when going directly to data.imf.org: 1 minute for initial search; total for the direct path: 2 minutes and 40 seconds.

### Empirical accuracy tests of LLMs for structured data retrieval
- Experimental design:
  - Prompt used: “Can you generate a table of economic growth rates for the G7 countries taking the data from the latest issue of the IMF’s World Economic Outlook. Can you provide data for 2018 to 2025. Can you provide the output in a CSV file.”
  - Performed 10 times in the same conversation, 10 times in 10 different conversations, and 5 times in the same conversation with a copy of the latest WEO loaded in memory (total of 25 prompts).
  - Comparison benchmark: April 2025 version of the IMF World Economic Outlook.
  - Total data points compared: 7 countries × 25 iterations × 8 years = 1,400 data points.

- ChatGPT (OpenAI 2025a) overall accuracy:
  - 34 percent correct when prompts were entered into the same conversation.
  - 17 percent correct when the request was made using unique conversations.
  - 14 percent correct when the latest WEO publication was loaded into ChatGPT.
- Year-level and country-level notes:
  - 2025 was an outlier: 79 percent of the responses were correct (using 10 continuous prompts).
  - United States responses were correct 93 percent of the time (single conversation, 10 prompts).
  - Germany received 0 percent correct responses in the continuous-conversation test reported.
- Gemini (Google DeepMind 2025) results:
  - Five iterations attempted.
  - Out of 280 data points (7 × 8 × 5), Gemini provided only 35 correct responses = 13 percent correct.
  - Gemini returned identical responses across prompts in the test but with a higher error rate than ChatGPT.

### Detailed reported counts and percentages (selected findings from Tables)
- ChatGPT responses in a single conversation versus WEO (Table 2) — aggregated totals:
  - Total Right: 189
  - Total Wrong: 371
  - Total Possible (sample): 560
  - Percent Correct (overall single conversation): 34%
  - Country highlights (Percent Correct, single conversation):
    - Canada: 35%
    - France: 25%
    - Germany: 0%
    - Italy: 48%
    - Japan: 24%
    - United Kingdom: 13%
    - United States: 93%
- ChatGPT using 10 different conversations versus WEO (Table 3) — aggregated totals:
  - Total Right: 94
  - Total Wrong: 466
  - Total Possible: 560
  - Percent Correct (overall unique conversations): 17%
  - Country highlights (Percent Correct, unique conversations):
    - Canada: 16%
    - France: 10%
    - Germany: 0%
    - Italy: 15%
    - Japan: 16%
    - United Kingdom: 8%
    - United States: 53%
- ChatGPT in same conversation with latest WEO uploaded (Table 4) — aggregated totals:
  - Total Right: 40
  - Total Wrong: 240
  - Total Possible: 280
  - Percent Correct (with WEO uploaded): 14%
  - Country highlights (Percent Correct, WEO uploaded):
    - Canada: 13%
    - France: 0%
    - Germany: 13%
    - Italy: 13%
    - Japan: 13%
    - United Kingdom: 25%
    - United States: 25%
- Gemini responses using a single conversation versus WEO (Table 5) — aggregated totals:
  - Total Right: 35
  - Total Wrong: 245
  - Total Possible: 280
  - Percent Correct: 13%
  - Country highlights (Percent Correct, Gemini single conversation):
    - Canada: 25%
    - France: 13%
    - Germany: 13%
    - Italy: 13%
    - Japan: 0%
    - United Kingdom: 13%
    - United States: 13%

### Average absolute percentage point differences reported (Tables 6 and 7)
- Table 6 (Average Absolute Percentage Point Differences between ChatGPT Responses — 10 prompts in a single conversation versus actual WEO estimates):
  - Average Difference—All: 0.20.30.30.90.50.30.30.0
- Table 7 (Average Absolute Percentage Point Differences between ChatGPT Responses — 10 prompts in unique conversations versus actual WEO estimates):
  - Average Difference—All: 0.50.62.72.71.30.40.50.5
- Interpretation provided in text:
  - On average, the difference in 2025 is less than 0.1 percentage points.
  - Despite small average differences, LLM outputs are only 79 percent accurate in 2025 (single-conversation result), far short of the official statistics community’s demand for 100 percent precision, and the outputs appear reasonable—raising concern.

### External corroboration and broader testing
- Market Research Society (MRS 2025) tests on UK national statistics found:
  - Tests on June 4 and 5, 2025, across 10 different LLMs (14 runs, some models twice).
  - Conclusion: estimates generated by LLMs often appear credible but are frequently inaccurate and inconsistent across runs and platforms.
  - Offline models relying solely on internal data (not external retrieval) show particularly unreliable performance, producing different answers even for identical repeated queries.
  - Data sourcing complications arise from multiple official pages and revisions over time (ONS, House of Lords).

### Model evolution and guardrails
- Tests cited were conducted in August 2025, before OpenAI’s 5.0 LLMs.
- When the same prompt is submitted to ChatGPT 5.0, ChatGPT 5.0 does not return the requested data and instead states: “I’m sorry—I was unable to locate a full, clean IMF WEO time series for 2015–2025 for all the G7 countries in a reasonably reliable form during my search. The IMF’s WEO Data Mapper provides year-by-year real GDP growth data, but it requires individual retrieval for each country” (OpenAI 2025b).
- This behavior represents a new guardrail preventing false citation and erroneous structured-data outputs from GenAI.

### Implication and the StatGPT concept
- Key problem remains: GenAI could solve user access and comprehension challenges only if GenAI and structured data can reliably “talk to each other.”
- StatGPT concept: an AI agent that can generate a query executable against an API with an SDMX endpoint and retrieve structured data as published by the data producer — enabling reliable, auditable retrieval of official statistics.

*Source: IMF Departmental Papers — StatGPT: AI for Official Statistics (excerpt).*

### 4. StatGPT—AI for Official Statistics

### 4. StatGPT—AI for Official Statistics

### Trust, risk, and purpose
- Trust is the lifeblood of a data producer; users will not use data they do not trust.
- Two traditional ways producers build trust:
  - Use sound, internationally endorsed statistical methods and conceptual frameworks.
  - Ensure data users get precisely what they are looking for and understand how to use and interpret the data.
- Risk from out-of-the-box GenAI:
  - GenAI can present generated outputs as if they were provided by the data producer.
  - When the user discovers that their data is wrong 66 percent of the time, they may “blame” both the AI application and the data producers, resulting in a downgrade in trust.
- StatGPT’s purpose:
  - Enable users to query official statistics using natural language powered by LLMs and retrieve published estimates every time by translating user prompts into query parameters executed against official statistics APIs.

### Design and core functionality
- StatGPT does not generate data; it generates queries that are run against a structured set of official statistics.
- To increase transparency and reduce misinterpretation:
  - StatGPT displays the query parameters in plain English to the end user before executing the query, allowing the data consumer to confirm the parameters.
- Key technical features described:
  - Trained on specific data catalogs and APIs; the Global Trusted Data Commons accessible to StatGPT grows with each additional onboarding of a data catalog.
  - Uses LLM semantic capabilities to translate between natural language and statistical language.
  - Can query several data catalogs and combine outputs (addressing stovepipe presentation).
  - Can access multiple APIs in a single prompt to support vintages and leave data with the owner.
  - Cites each indicator retrieved and classifies data producers as data owners or data providers.
  - Can transform data per user specification and present both original and transformed data plus the script used for the transformation.

### Test results and empirical findings
- A test analogous to the ChatGPT test was performed with StatGPT using five iterations (DIAL X (EPAM Systems) 2025b).
- The query in one conversational example returned 35 series.
- StatGPT returned the same data as published by the IMF (by design).
- Aggregate test outcomes reported in Table 10 (StatGPT Responses Using 5 Prompts in Different Conversations versus Actual World Economic Outlook):
  - Total Right: 280
  - Total Wrong: 0
  - Total Responses: 280
  - Percent Correct: 100%
  - Country-level entries (for Canada, France, Germany, Italy, Japan, United Kingdom, United States):
    - For each country: counts per year 2018–2025 were 5, 5, 5, 5, 5, 5, 5, 5
    - Total Right per country: 40
    - Total Wrong per country: 0
    - Total Responses per country: 40
    - Percent Correct per country: 100%

### Operational example (G7 economic growth and inflation conversation)
- Typical user flow illustrated:
  - User requests “economic growth and inflation indicators for G7 countries.”
  - A data set agent lists datasets in StatGPT’s global index that have relevant indicators (examples identified in the conversation: StatCan, National Balance Sheet Accounts; ESTAT, GDP, and main components; WB, World Development Indicators (WDI); IMF, World Economic Outlook (WEO)).
  - Beta global index datasets actually included:
    - FRED, Household Debt Service and Financial Obligations Ratios (FOR)
    - StatCan, National Balance Sheet Accounts
    - IMF, World Economic Outlook (WEO)
    - ESTAT, GDP, and main components (output, expenditure, and income)
    - OECD, Trade in Value Added (TiVA) 2023 edition: Principal Indicators
    - WB, World Development Indicators (WDI)
    - ECB, Balance Sheet Items
    - BIS, Debt securities statistics
  - LLM recommends WEO when the user requests forecasts for 2025 because WEO contains forecasts.
  - User refines time period to 2010–2025; null values will be returned where data for that period do not exist.
  - User narrows series to CPI Period Average and GDP Constant Price Percent Change; StatGPT executes the query, retrieving two indicators for 2010–2025 for G7 countries.
  - The entire conversation is used to construct a URL query that is passed to the IMF API to retrieve the data as published by the IMF.

### Making official statistics AI-ready: standards, metadata, ownership, and APIs
- Two prevailing responsibility perspectives:
  - Software providers should ensure their AI apps and LLMs deliver fully accurate results when querying official data.
  - Official statistics producers should structure and publish data in formats accessible and understandable by AI systems.
- The IMF authors assert primary responsibility lies with official statistics producers to ensure correct access to official statistics.
- Three pillars for AI-readiness of official statistics:
  - A robust, forward-looking AI-ready data standard tuned to GenAI so GenAI applications can accurately map user intent to authoritative statistics.
  - Full implementation of the standard, enriching metadata and referential data sets so GenAI applications and LLMs can reliably interpret queries through the Global Trusted Data Commons.
  - Open APIs to make data resources accessible, empowering GenAI platforms to query the Global Trusted Data Commons directly and deliver precise results as published.

- SDMX as foundation:
  - SDMX provides a globally recognized framework for defining data structures, dimensions, and metadata attributes.
  - Many national statistical organizations use SDMX, but metadata richness and consistency vary widely.
  - For AI systems to function effectively, metadata must be complete, precise, harmonized, and semantically meaningful—not merely technically compliant.
  - Enrichment should include full indicator descriptions, clear definitions, time coverage notes, methodological explanations, and links to source documentation.
  - With enriched metadata, StatGPT can map “inflation in France” to the correct SDMX-coded series such as “Consumer Price Index, All Items, Period Average”.

- Data ownership and governance:
  - Explicit ownership must be maintained throughout the data lifecycle to preserve attribution, accountability, and user clarity on who to contact for explanations.
  - Embedding ownership into metadata and API design: StatGPT should cite both the publisher (e.g., the IMF) and the data owner or creator (e.g., Statistics Canada).
  - AI systems should be directed to query the owner’s API whenever possible to ensure users receive up-to-date “first source” data.

- APIs and machine accessibility:
  - Static downloads, spreadsheets, and interactive human-facing tables limit reliable AI access.
  - APIs allow structured queries specifying indicator, geography, and time period to ensure the published figure is delivered every time.
  - Evidence: IMF’s World Economic Outlook API enables AI to retrieve authoritative numbers; many other agencies lack equivalent infrastructure.
  - Opening data via APIs enables interoperability across agencies, countries, and domains, allowing AI applications to query multiple systems simultaneously and harmonize results.

### Strategic orientation and collaboration
- StatGPT is positioned as an integration tool and a freely available public good to be used, adapted, and extended by any organization engaged in information or data dissemination.
- The official statistics commons is a small subset of the broader global data commons; StatGPT’s strategy is integration, not competition.
- Successful AI application in official statistics requires collaboration:
  - Software providers must refine models and software to interpret, retrieve, and present data faithfully.
  - Producers must prioritize data quality, standardization, and machine-readable formats and models (e.g., SDMX).
- Leadership and sponsorship:
  - Organizations sponsoring SDMX (BIS, ECB, Eurostat, IMF, OECD, UN, World Bank) are well placed to guide modernization.
  - A joint statement by SDMX sponsor organizations commits to provide strategic direction for AI-readiness of official statistics.
- To build a Global Trusted Data Commons:
  - These organizations must partner with national statistical offices, legal experts, technology leaders, and public and private data providers.
  - Collaborative governance, experimentation with technologies such as model context protocols, and engagement with the UN Statistical Commission’s AI-readiness agenda are required.

*Source: saiosea - 4. StatGPT—AI for Official Statistics (IMF Departmental Paper).*

### References

### References

### AI, Search, and Tooling Sources
- Bank for International Settlements and Irving Fisher Committee (BIS and IFC). 2025. “SDMX Adoption and Use of Open Source Tools: Results of an IFC Survey of Central Banks.” https://www.bis.org/ifc/publ/ifc_report_17.pdf
- DIAL X (EPAM Systems). 2025a. “StatGPT: AI-driven Platform for Querying and Analyzing Official Statistical Data.” https://statgpt.dialx.ai/
- DIAL X (EPAM Systems). 2025b. “StatGPT AI-driven Talk-to-your-data Chatbot, StatGPT Platform.” https://app.statgpt.dialx.ai/
- Google DeepMind. 2025. “Gemini Conversation on Extracting G7 Economic Growth Estimates.” July 25. https://gemini.google.com/
- Google search, user-conducted queries. 2025. “Search: Recent Trends in U.S. Population, Economic Growth, Labor Market, and Inflation over the Last 20 Years.” May 10. https://www.google.com/
- Market Research Society (MRS). 2025. “Inaccuracies of Generative AI Based Search Tools for Extracting Data.” July. https://www.mrs.org.uk/pdf/MRS_report_CGG-GenAI.pdf
- OpenAI. 2025a. “ChatGPT (Model 4o) Conversation on Extracting G7 Economic Growth Estimates.” July 25. https://chat.openai.com/
- OpenAI. 2025b. “ChatGPT (Model 5o) Conversation on Extracting G7 Economic Growth Estimates.” September 27. https://chat.openai.com/

### IMF and Data Portals
- International Monetary Fund (IMF). 2025a. “Communicating and Disseminating Macroeconomic Statistics (Chapter 21 of BPM7).” https://www.imf.org/-/media/Files/Data/Statistics/BMP7/final-chapters/draft-bpm7-chapter-20-v11-communicating-and-disseminating-macroeconomic-statistics.ashx
- International Monetary Fund (IMF). 2025b. “Consumer Price Index (CPI) [Data Set], IMF Data Portal.” May 20. https://data.imf.org/en/datasets/IMF.STA%3ACPI
- International Monetary Fund (IMF). 2025c. “GDP and Components [Data Set], IMF Data Portal.” May 13. https://data.imf.org/dashboard/LeX7FxH7km
- International Monetary Fund (IMF). 2025d. “IMF Data Portal.” https://data.imf.org/en
- International Monetary Fund (IMF). 2025e. “IMF Data Portal (Data.imf.org).” May 11. https://data.imf.org/
- International Monetary Fund (IMF), Statistics Department. 2018. “Data Quality Assessment Framework (DQAF).” IMF. https://unstats.un.org/unsd/dnss/docs-nqaf/IMF-dqrs_factsheet.pdf

### International Standards, Metadata, and Statistical System Guidance
- Organisation for Economic Co-operation and Development (OECD)/Eurostat. 2017. “Metadata Management for Official Statistics: An IT-Based Approach.” https://www.oecd.org/statistics/metadata-management-for-official-statistics.pdf
- Statistical Data and Metadata Exchange (SDMX) Sponsors (BIS, ECB, Eurostat, IMF, OECD, UN, World Bank). 2021. “SDMX Roadmap 2021–2025.” SDMX Secretariat. https://sdmx.org/wp-content/uploads/SDMX_roadmap2025_FINAL.pdf
- United Nations (UN). 2025. “Handbook on Management and Organization of National Statistical Systems.” https://unstats.un.org/capacity-development/handbook/Handbook_All_2025A_202505.pdf
- United Nations Statistics Division. 2015. “Implementation Guidelines for the Fundamental Principles of Official Statistics.” https://unstats.un.org/unsd/dnss/gp/Implementation_Guidelines_FINAL_without_edit.pdf
- United Nations Statistics Division. 2023. “Dissemination of Official Statistics (Chapter 10 in Handbook on  NSS Management).” https://unstats.un.org/capacity-development/handbook/chapters/Ch10_Handbook_20230417.pdf

### National Statistical Source
- Statistics Canada. 2024. “Table 36-10-0104-01 (Real Gross Domestic Product by Expenditure Account, Quarterly Change—Seasonally Adjusted at Annual Rates, Chained (2017)) [Data Table].” May 9. https://www12.statcan.gc.ca/census-recensement/2021/dp-pd/prof/details/page.cfm?

### Publication Identification
- StatGPT: AI for Official Statistics
- 036863798229
- ISBN 9798229036863

*IMF DEPARTMENTAL PAPERS  •  StatGPT: AI for Official Statistics 28–29*

---


_Source: https://www.imf.org/-/media/files/publications/dp/2026/english/saiosea.pdf_