Trabajo 4dc400733420

Vista previa (transcripcion.txt)

[00:00:07] So, welcome everyone to the global network webinars. My name is Alexander Lushki and I will be moderating this webinar today together with my colleague Sean Lob. Today we have with us Malcolm Marini from the IMF who will talk about how AI
[00:00:27] is transforming the dissemination of official statistics. This is the fifth and last of our five webinars in our AI for official statistics series. Artificial intelligence is rapidly reshaping how official statistics are
[00:00:46] discovered, assessed, accessed, and used. Yet, the promise of AI powered dissemination comes with a critical challenge. ensuring that the data users receive through AI tools are accurate, traceable, and grounded in
[00:01:08] authorative sources. This presentation today provides an overview of AI powered tools and methods that are already in use across international organizations and national statistical offices to address this
[00:01:26] challenge and modernize the dissemination of official statistics. The presentation today will also discuss the machine readable metadata like SDMX in making data truly AI ready. Let me introduce now our speaker. So
[00:01:50] Marco is division chief in the statistics department of the IMF. He leads and coordinates a wide range of projects and activities related to data governance and data analytics for the fund including oversight of the IMF big
[00:02:08] data center. He's the author of several scientific publications on topics related to economic measurement including national accounts benchmarking, data revisions, forecasting and seasonal adjustment.
[00:02:24] Prior to joining the IMF, he was a senior senior researcher at ISAT Statistical Institute. He holds a PhD in statistical methods for business and economics from the University of Rome Roma. And now before I hand over to Marco, as
[00:02:46] usual, some practical remarks. [snorts and clears throat] So you can write your comments and questions into the meeting chat at any time. Um but during the Q&A session it would be really nice if you could ask the
[00:03:00] question yourself. For that you can then raise your hand and we will give you in the floor. As always, this global network webinar is being recorded and will
[00:03:17] in. So if you haven't joined the new global network yet, we invite you to follow the link which I will share in the chat in a few moments and to join the global network group on LinkedIn. Also, please sign up for our newsletter
[00:03:35] and so we can keep you updated about future events. And now over to you, Marco. The floor is all yours. Thank you. >> Thank you. Uh thank you, Alexander. Um let me just uh uh share the screen. Um okay so can you confirm you see um
[00:04:04] >> yes all good >> okay great um hello hello hello everyone uh I see uh over 150 u people connected that's that's great thank you for for joining from uh uh wherever you are um I want to thank also the US statistic
[00:04:21] division uh for organizing this AI spin series which is very uh relevant and also clearly for for for inviting me. Um so my name is Marco Minimas as Alexander said I'm I'm in the MF statistics department and today what I want to talk
[00:04:38] about is what I think is is one of the most consequential shifts happening in uh in official statistics right now which is not not in how we produce data but in how people uh access it. So my aim is just to give you a concrete
[00:04:54] picture of what's uh already happening and and show you real tools uh that are live or or almost live and leave you with a sense of where where the statistical community uh uh needs to go together. So I'll I'll try to speak for
[00:05:10] about 30 minutes and then and then have some time for for questions. Um
[00:05:20] okay so uh uh let me start with why this is urgent. The the way people get uh information has fundamentally changed with AI. Uh users now expect to ask a question um in plain language and and get an answer uh instantly. They don't
[00:05:40] want to navigate data portals that lot CSVs or decode a table and a tools like chasht Gemini become really the front door for for millions of people uh seeking data. So this creates a risk in the statistical community because if
[00:05:59] statistics are not accessible to uh AI systems then we will simply not appear in the answers people receive. But there's also an opportunity clearly official statistics have something that AI generated content cannot easily uh
[00:06:15] replicate uh which is institutional trust uh methodological rigor and also longtime series. So so in a sense we can reposition uh ourself as as a trusted backbone of of AI. Um and and the core challenging here is is
[00:06:34] trust. So uh again AI can democratize uh access to data. Uh but today's general purpose systems have have serious weaknesses. They elucinate numbers. They are very good with text but when it comes to numbers uh they're not perfect
[00:06:50] and and you know being precise with numbers is critical for organizations in the statistical community. Um and and the key question for this talk is how how do we make uh uh official statistics accessible through AI while preserving
[00:07:08] uh exactly the things that make them trustworthy which is accuracy uh provenence um comparability and and methodological clarity and so this this is the question that essentially shapes uh everything that I'll cover today
[00:07:25] um to make this concrete uh Um so let let's look at you know what happens when you ask a a standard ALM statistics question today. So you typically get an answer but no source attribution no link to a data set no explanation to the
[00:07:42] methodology and and and numbers may be outdated um and some some LLM you know have knowledge cut off so you don't get the the the the actual current numbers and also the coverage is is uneven so well if you look for G7 indicators then
[00:08:01] you can find good answers but then the moment you ask about smaller economies or or niche series then the quality drops uh sharply and the result is that official statistics become just another like random input out there and no more
[00:08:18] authoritative than than sort of a Wikipedia article or any like blog post. So what what we actually need is is structured machine readable uh um pipelines that connect AI system directly to official data at source. And
[00:08:36] this is the technical challenge that this presentation tries to uh address. So let let me let me just again repeat the frame. It's the shift that we are uh going through. So, so the traditional model for for accessing official
[00:08:51] statistics is you know through websites uh static tables uh sometimes PDF or even more like a fixed APIs and this is really was built for for people navigating screens and and it work well with for for train data users but it
[00:09:07] does not work for for for systems. Um the emerging model is is one where data uh is is delivered through chatboards um copilots um AI generated dashboards and automated reports and and the user ask a question in in natural language and the
[00:09:24] system reads structured data and applies context and and and generates a response. So, so there is a fundamental shift from uh searching for data to to asking questions of data and and this sounds simple but it requires a
[00:09:41] completely different infrastructure on our side and and then and the message at the bottom of this slide is the one I want you to take away from the next few slides because that shift is possible uh uh the standard uh already exists and
[00:09:54] some of us are already uh uh doing work in this direction. So what does it actually mean for data to be AI ready? You you often hear this u um sort of expression although there are different forms depending context.
[00:10:13] Uh but let let me try to see like what makes data ready like for for this context. So first machine readable formats is really the starting point. data that a computer can can uh read without human intervention is critical.
[00:10:28] But the the bigger differentiator uh uh especially in our field is metadata. So we we need rich structure metadata including concepts, definitions, units, time periods, geographic coverage is what allows an AI system to understand
[00:10:45] what a number means and not just what it is. And without metadata you have numbers without context. And semantic interoperability means that when like say IMF, World Bank, uh uh Euroat and others report GDP uh uh or CPI an AI
[00:11:05] system can understand that they are compatible because they share the same concept definitions and core standards. Um another uh aspect of readiness is a APIs. APIs uh enable real time material. So uh an
[00:11:21] AI answer can reflect the latest release and not something that has been sort of there for training data set for like you know years years before and and provenence and versioning ensure also that the you when a user gets an answer
[00:11:36] they uh can trace it back to a specific data set a specific vintage and a specific methodology. So the standard that delivers all of these together um is SDMX uh which is the uh topic of the next uh slide. So SDMX is the statistical data
[00:11:55] and metadata exchange standard is is really the foundation that makes all of this work. It's a global standard maintained by the SDMX sponsors the MF, BS, ECB, Euroat, OCD, UN World Bank and ILO. uh and SDMS covers both data
[00:12:13] structure and metadata uh common vocabularies, code list and also API based dissemination and for AI uh SDX matters uh in in three specific ways. So first uh it provides the structure format the the AI systems can actually
[00:12:34] uh uh parse and and reason over. Uh second it enables cross country cooperativity goes to the same code means same thing regardless of the source organizations and third it supports uh uh traceability
[00:12:46] you can always follow the data back to its to its origin. So you can think of SDMX as the infrastructure layer that sits between official statistics systems and AI applications and without it AI systems have to guess uh what the data
[00:13:02] means uh with with with it with SDMX they can now uh know. So so let let me make uh uh uh the SDMax uh AI connection a bit more more concrete. So, so with structure SDMX data sets you you can now flow directly
[00:13:20] into AI pipeline. So there is no need to manually formatting, no data wrangling. So the structure is already there and the meta data layer really does the uh heavy lifting for for these. So when a user ask about GDP, the AI needs to know
[00:13:38] well real or nominal market price or factor cost annual or quarterly. So these are concept that SDMX definitions resolve because they it's it's in their structure. Um an SDM SDMX API uh API also enable real time retrieval which
[00:13:55] means an AI tool can fetch the latest data at the moment of the query rather than relying on old training data and and the outcome is a shift from uh documents designed for for human readers um to uh data system that AI can act on.
[00:14:12] So this is the transformation we are uh working on. Um now let me describe what the MF uh is building. So uh our approach combines three elements um like as clearly SDMX structured data sets um then APIs uh and and also AI uh on top
[00:14:34] clearly uh start GPT some of you may have heard of it is a is an IMF tool that uh we have just released is a natural language interface to uh official IMF data sets. Um users can ask questions in plain lage like you know
[00:14:52] show me inflation for in emerging markets for the last 5 years and receive accurate structural response that are backed by uh actual IMF data uh that IMF uh uh collect from uh uh its member countries. Uh so the key design
[00:15:08] principle is that every response from m subject V must be accurate tra traceable back to it source and linked to authoritative IMF data. So we deliberately engineer trust into the systems from from the start
[00:15:24] um and and here I show uh what statip is uh it has been uh published uh in June 2026. It is live at portal stat.org. org. Um, and what you're seeing is is it goes even beyond such we um sort of branded this as a global trusted data
[00:15:48] commons. Um, it's essentially like the next step. It's not just tools but also kind of giving trust to the data that uh are searched from tools. And so when you go into this portal, you you can navigate a search across almost 10,000
[00:16:05] official indicators from the not only from the IMF but also from from from BIS who also uh uh sort of agree to uh include the the data sets into this uh data common uh and also uh uh uh over 130 uh uh national stat data from over
[00:16:24] 130 national statistic is a available through the national data summary pages also uh um uh supported by MF but in what you see there is that there is a conversational interface there's no like browsing hierarchy or stuff like that so
[00:16:41] you ask about GDP like uh inflation trade uh or any economic indicators you also have some prompts that help you uh navigate this uh u sort of database uh data sets and and and critically the answers that you receive here are
[00:16:58] grounded in official SDMX data. This is not a like a a wrapper around the general purpose LLM uh that guesses the numbers but every every figure figure is retrieved from a structured uh authoritative source and and we are sort
[00:17:16] of uh expanding this uh data common. We have a an on boarding strategy which is a tiered. So we are starting with what we call tier zero uh which essentially includes uh all the SDMX sponsors. So clearly supports SDMX and also national
[00:17:35] statistical offices that have SDMX ready data and and this will sort of be progressively uh uh like upgraded with on on boarding of trusted official statistics by year. So the idea is to expand cover coverage um carefully uh
[00:17:54] and maintaining quality and prominence at uh at every step. So so I I encourage everyone here to go to uh this site portal.org and try it after the session. Um let me uh maybe just go a little bit broader than just the MF and and say
[00:18:11] what is happening out there. Um, beyond SDMX APIs, there is also a new technical standard that is emerging that is worth highlighting which is the model context protocol or or MCP. Um, MCP is a standardized way for AI models to
[00:18:30] connect to external data services in in real time and think of it as a really universal plug for AI. So instead of each AI system having to be custom integrated with each data source, MCP provides a common interface. So this MCP
[00:18:45] is like HTTP, right? When we started with websites, HTTP was was developed, MCP is kind of expected to go in that direction. Although not all companies are still uh aligned, but it it seems that it's it's going into that
[00:19:02] direction. and and for for official statistics the the benefits of these um sort of are significant because MCT allows this real time access uh um and responses would reflect the latest release and so uh grounding the AI in in
[00:19:19] structure retrieve data dramatically reduce uh risk of aluc hallucination. I I want to sort of give um credit to to to India because the India statistics office is an early mover here. Um they developed an MCP uh that is integrated
[00:19:36] into the national uh statistics office website and I just read an article today that MOSI actually want to go beyond the the data. I want they want to build a common data platform as the authority source for all government data in India.
[00:19:50] So that's that's very ambitious but definitely something to follow. Um but the the the broader message here is that we are um moving toward like a plugandplay ecosystem where AI applications can connect to official
[00:20:05] statistic data this the same way they might connect to any other resource could be a calendar or search engine uh in a sense seamlessly and in real time and and and we provenence which are important elements for trusted
[00:20:20] statistics. Um also want to say that uh there are other AI tools in the ecosystem. Um ST GBT as I mentioned is is not an isolated example. There are uh many other examples across international organizations and NSOs
[00:20:39] uh that are testing prototyping a whole ecosystem of AI tools. There are conversational chatbots for the access. Uh this is the most pro most probably the most uh uh visible category uh that lets users query official data
[00:20:56] natural language. Uh there are also AI generated dashboards and visualization that are increasingly common system that not only like retrieve data but uh say automatically choose the right chart type and and generate comments. Uh there
[00:21:13] are all also um uh experimentation to uh to create automated press releases and report uh graphics um AI draft the statistical commentary and and the human reviews and publishes. This is like becoming more and more common
[00:21:32] and there's also the documentation uh uh uh and manual summarization tools. So at Bul Palace and us and navigate the vast methodological literature that surround uh official statistics. So the key point
[00:21:45] that I want to make is that we we sort of we are moving past the prototype phase. There are some of these tools are operational are in production uh with the real organizations be behind it. So the the question is no longer like
[00:21:59] whether this will happen but it's really how fast this um will will scale up. Um I also want to highlight other initiatives that are led by uh SDMX sponsor agencies and also one from IMF um that I think represent sort of what
[00:22:19] what AI can can really help uh help us uh uh kind of going forward. First so the BIS has developed an SDMX AI assistant which is available sdmx.io. This is a chatbot uh built on on the BS uh infrastructure that gives um access
[00:22:41] to the full uh SDMX library documentation, guidelines, tools. Um it went public on SDMX.io in in May 2026. I believe it will be announced on sdmx.org also soon. So for anyone working with SDMX, given that we talk about I talk
[00:22:58] about SDMs a lot, it's really really useful. You can ask uh you can go there now on SDMX.io and ask any question like you know explain a data structure definition retrieve a code list or walk you through a specific guideline and and
[00:23:12] and you will get the very good answer. Um and second sort of uh the the OCD's work to the uh CIS community uh they they built a national language pilot on top of the data explorer using an MCP service and also they are doing um work
[00:23:33] on on defining AI readiness as part of the uh at a uh the eye level group on modernization of of a fish statistics. they're co-leading this AI dissemination work stream. So this is also an example of of sort of AI integration that needs
[00:23:50] to be embedded into institutional strategy not just treated as a as a side project. And and finally uh there is another tool that I has has has developed and published externally is called talk to talk to manuals. Uh this
[00:24:05] is a different tool. We're not u navigating data but we are actually navigating statistical manuals like a national accounts uh system of national accounts the balance of payments man manual the government final statistics
[00:24:18] manual um so when a compiler uh or data user ask a methodological question the tool retrieves the the relevant data from these manners the reant metadata the relevant set of sentences paragraphs from this manner and provides a
[00:24:33] reference answer. So if you want to test this, it's available through the MF compiler app. You see the link here and uh you have to register and then you can test it out. um tools to tools are necessary uh they
[00:24:50] represent the technical sort of backbone but it's also important uh that governance around these tools is is clear and and AI readiness require requires a whole ecosystem to move institutions standards body and capacity
[00:25:07] so here I listed a few of the initiatives that I'm aware of I'm sure there are more uh that are sort dealing with sort of the AI and and data dissemination for official statistics. Um Paris 21 um has established a task
[00:25:22] team on AI for official statistics. I presented last week then uh they're focusing particularly on the needs of of low and middle inome countries and ensuring that the benefits of AI are accessible and sustainable for all
[00:25:37] national statistics system. I mentioned the HLG MOS uh and the UNC uh they're also developing an AI modernization frameworks that can guide NSOs through this transition. uh there is an important initiative led by the Swiss
[00:25:53] federal statistical office which is a trusted data observatory uh which is exploring how to link official data with a trust mechanism uh uh ensuring that data uh uh so is accompanied by with with trusted
[00:26:10] metadata. Then there is a giggalli group which is working under the UN statistic commission is another important forum for dialog dialogue on these issues particularly relevant for national statistics systems uh in in
[00:26:26] lower income countries and finally the SDMX AI AI task team which is uh the one that I'm mostly involved uh it has been established by the SDMax sponsors in early 2026 um is a is a body that essentially is is
[00:26:44] working at the intersection of SDMX and AI. Uh uh and it works of course in conjunction with the other bodies of SDMX in particular the technical working group which is dealing with the more technical aspects of how to uh uh so
[00:27:00] upgrade or make SDMX work with AI assistant. So the message that I want to um just leave deep leave deep leave deep leave deep leave deep leave deep leave deep leave deep leave deep leave deep leave uh with you on this slide is
[00:27:09] simple that AI is not just technical but really requires ecosystem readiness. So like a system of standards, governance, capacity and coordination. They work together and as you see there are there may be some overlap from this group but
[00:27:24] uh we know that we are connected and and sort of we we try to uh uh address risk of doing the same thing by just communicating and and and working together. uh in this direction though there is another uh kind of working together
[00:27:41] issue that I want to alllight which is working with technology companies working with the private sector. Um so uh many of us think that we need to actively engage with technology companies that build AI systems. So AI
[00:27:57] assistants are are becoming the primary gateway to information for for a large and growing share of the population. Um and as I said if official statistic are not correctly represented in those systems uh we have a problem that no
[00:28:12] amount of internal data quality work will fix. So there is already good good experience especially with Google. Um UNSD has developed the data commons of of UN agencies with with the partnership with Google. Uh IMF uh OCD u
[00:28:32] I believe also the BS had some uh relationship with with Google as well and this collaboration has been has been very productive. Uh we should do more. um the major LLM providers um are sh are shaping like how
[00:28:49] hundreds of millions of people u can of find information. So it's important that we we engage with with those players and and and I think this is important for the community. It's not just to ensure that our data uh
[00:29:06] appears in a answers but also to ensure that appears correctly with attribution with proence and and with an methodological context that make data trustworthy. So uh let me bring uh together uh sort of like five things
[00:29:23] that I want to like emphasize at the end of this presentation. Uh so so first AI is transforming dissemination not not someday but now um the trust is the differentiator so official this is something that no AI model can can
[00:29:38] manufacture um SDMX is foundational for uh AI readiness if you're not if your organization is not investing in SDMX you should consider uh and tools like SPT show what is already possible and and finally
[00:29:54] collaboration uh is really not optional. There's no single organization that can build this ecosystem alone. Uh it requires many to uh work together and I just want to close on this uh I want to call a call to action. So there been a
[00:30:10] lot of experiment experimentation but the the there are now tools in production. The tools work the standardiz so it really time to move from experimentation to implementation. uh uh and then means that there is a
[00:30:25] need to like for for investment uh in three things in particular in standards especially SDMax that's what we propose in metadata quality and also interpolability across systems and organization and it means really kind of
[00:30:40] working together not just within this network but as I said also with technology companies um and and officials this is a really public good and sort of keeping them private uh visible and and correctly represented
[00:30:54] in the age of AI is one of the most important things uh our community can do. So I'll stop here and thank you again for your um uh for attending and I look forward to your question. Over to you Alexander.
[00:31:08] Thanks Marco and thanks for this overview of what is already being done in this um realm of using AI for the dissemination of official statistics. I would like now to open up the floor for questions and comments. So uh I have
[00:31:27] enabled the microphones and the cameras. So please if you have a question uh raise your hand and um maybe ask it live. Otherwise you can always uh write it into the meeting chat. Um so I saw in the in the in the chat
[00:31:46] actually a number of comments uh regarding the the cost actually of um providing some of those AI services. So it was mentioned that in the case of chat uh stat GPT the IMF pays for the AI. So uh rate limiting must maybe is a
[00:32:08] is a must uh to keep the costs at check. Um so um well what is your view on that Marco? >> Uh so uh can you so is this a question in the chat? So can I >> it was a question by uh Kavier Martin but well he he was mentioning that the
[00:32:32] IMA so the MCP if you're using MCP the user pays for the AI uh that was his view and that uh in the case of stat GPT the AI uh costs are paid by the IMF because you run that platform and he was wondering whether uh this is actually a
[00:32:54] cost factor. than for the IMF of running this platform. >> Yeah. No, there is a um clearly for organizations that have um uh like stand up this AI tools in in its own server. There is a a cost implication for for tokens. But yeah,
[00:33:13] this these tokens are not like too expensive. Of course, you need to factor that in. uh but um considering also that the type of uh like traffic that is generated you know in our uh world is not the one that sort of you would
[00:33:30] imagine like on platform like load or chd for general purpose kind of questions right so I think yeah there there is a um a cost sort of to sustain and and this is something that MF when we develop such we factor in right so uh
[00:33:46] but again it's it's not something that would uh would uh like say would be determinant now on MCM on MCP um yeah that that's a different type of model I um yeah I I don't know I guess that that is also something that would would need
[00:34:02] to sort of whether like a paid version perhaps that's that's the cost that you would need to sustain to access this MCP I guess like a free version probably would not allow you to do that but I I yeah don't yeah don't think that is a
[00:34:16] like a something that you would need to pay for for like by question right so again but I'm I'm not sort of familiar with how MCP works for for the case of India but yeah that the cost sort of factor of course will
[00:34:29] need to be factor in when uh this decision are made >> yeah thank you so anyone else wants to ask a question please raise your hand um and then [snorts] >> I've got a question the floor >> yeah please come in Sorry. Sure.
[00:34:48] >> I um yeah, Marco. So, hi. I'm Marco. I'm one of I'm Sean, one of the co-hosts. I just wanted to ask um typically when users want to find out about official statistics. So, you talked about the paradigm shift. They had to go navigate
[00:35:01] data portals, look up definitions, metadata. Once they kind of figured out what they wanted, they would query the data, pull it, and then kind of do do this manual we would call context merging where you take all the stuff
[00:35:11] you've just learned and you pull it into your own context. sort of implicit in this. There's this mental task and you know translation and understanding and this and that. Now, now it's interesting if you fast forward to 2026, we have
[00:35:23] this new pattern where sort of a lot of that um cognitive effort is offloaded to the LLM and you can kind of come with a very specific very simple prompt and sort of obtain an answer. But I just wondered if what you what your thoughts
[00:35:37] were about this this gets talked about as cognitive offloading. It's this idea that as tasks become simpler, some of the essential learning required to make good use of, you know, this data becomes lost. So, I just wanted to know if you
[00:35:52] had any thoughts about this issue of a cognitive offloading as people start getting simpler and simpler ways to access data. Are we losing out some of the [snorts] the benefits of costly searching of grappling with definitions
[00:36:05] of terms? When you look, you try to find CPI as a home appraiser, then you have to understand how CPI is measured. You have to understand the nuances of the metadata a little bit to make sufficient use. But with an AI, you can just ask a
[00:36:17] naive question and get a naive answer. Is there any anything lost in this transformation in your opinion? >> Well, uh I think uh there may there may be some loss in sort of in the um kind of human uh kind of factor human
[00:36:36] thinking, right? When you start like the okay, what what do I connect? what do I integrate? Uh however there's a there's a lot of gain in my view like uh when when when you um when you ask like a like integration of different data sets
[00:36:52] and again suppose you're using statute and so you you build something that is connecting data say from from BS from a country from IMF uh that that work uh takes like seconds right so if you do that uh it would take hours and and and
[00:37:11] the result I think would >> may not be as as uh you know good as the the one that I provides because you know l linking this uh data sometimes you know humans take shortcuts right so because they you know they they don't
[00:37:26] want to spend much time on data wrangling and say okay so let's let's let's do this so >> so so that's the biggest gain that I see of course it's the speed uh but also that the gain that I see is also the Um
[00:37:41] the fact that you know by not spending time on doing these things you spend time on you know thinking about you know once you have your like integrated data set so how can I assess this how can I analyze and AI can also help with that
[00:37:54] right so many times you ask questions about okay so I have this data set now I have this idea but can you suggest how can I you know address this question and and sometimes I is you know very good it's it's Not because it's AI but it's
[00:38:10] it's essentially AI is the word knowledge right there. So it's essentially you know it's picking up the brain of uh of uh uh billions of people that has written content out there. So, so again I I see a lot of gains. I I
[00:38:26] know what you mean that sometimes this integration you you you could lose sort of there's a human touch >> but uh again you can also direct uh the the the AI systems to to to tell you you know to do what you want right so for me
[00:38:43] it's more a matter of like speeding up the sort of those uh kind of manual task and uh and and and rather and dedicating more time for var sort of task Yeah, that's a great response. I I just wanted to um probe and see what your
[00:38:59] thoughts were. There are other uh questions in the chat. I I can ask. Bob Taylor asks, "What does it take to convert existing public data repositories like CCAM to something like STMX?" So, you talked about STMX being
[00:39:11] sort of a great sort of AI native uh data format. And so if if people are sitting here on top of data sets that are potentially not SDMX ready, um do you have any thoughts on sort of the mental model, what what shifts um and
[00:39:25] how how they should how they can easily go about doing that? >> Yeah. >> Yeah, that's a that's a great question. Um uh so uh there's a lot of resources on STMX uh uh out there. So uh what needs to be done essentially is uh is
[00:39:42] kind of a number of things but the very first thing is what we call data modeling right so you have our data set uh it's a and it's a spreadsheet it has you know certain dimensions that's what we call SDMX you have dimensions you
[00:39:55] have attributes so the data modeling is essentially take your data uh I don't know it's a times here quarterly data you know concept provenence and and and create like a artifacts that essentially create the
[00:40:11] structure around your data and and this is not like a like just a as simple as that. So it requires definite knowledge um as I mentioned and now there's also these resources out there that the uh use AI to help you model the data. So
[00:40:28] like uh the SDMX AI assistant that B just released, you go there and you type in what you want to do and you you get a very good answer or you know how to approach a problem and uh um uh there's also uh tools that uh uh so AI is also
[00:40:45] now be used to to model the data. So it's again structured data set is you know if you do it manually and if it's a complex data set it takes uh it takes it can take a lot of time. But now there are also AI modeler. So you can actually
[00:41:00] use AI to model your data. Um so and then you know I want also maybe promote the work of the IMF. Uh if if you are a statistics agency or a central bank we also have a capacity building pro program that really helps organization
[00:41:20] um move to to SDMX. Uh so we are working with several already. you know we have our uh data standards initiatives we have a transition uh over a 100 countries to SDMX over the past 15 20 years so we do that uh all the time I'm
[00:41:38] sure also you know UNSD OCD and other sponsor agencies have their own capacity building program that help uh uh its constituency um again for for for people out there uh that are not perhaps in the statistics
[00:41:55] community but want to model their data. There are available resources. There also I think training out there that are accessible for uh uh um SDMX uh sort of newbies. There is a LinkedIn work group uh LinkedIn group that can also uh be
[00:42:12] useful. There's a YouTube channel. So there are a number of like resources that we we we can uh give access to and uh so you know for for Bob like if you want to maybe more specific answer you invite him to reach out to me and maybe
[00:42:27] I I can be more precise. >> Wonderful. Thank thanks Marco. I see uh a hand up by Isabella um from UN women. Isabella you want to come in?
[00:42:49] Um, hello. Can you hear me? >> Yeah, we can hear you. >> I just changed system so I was worried I was lost. Um, >> so yes, my question is actually, you know, related. It's excellent work that you're doing. I'm so happy to hear that
[00:43:06] technology is being embraced so actively at um, IMF. I'm I'm thinking in the direction of uh the singularity is near um and I'm looking at national statistical offices. I'm looking at statistical offices in
[00:43:20] the region of Africa where I work. [snorts] Um I've heard comments in one that um they want to actually restrict access of the statisticians to um AI um in some of these statistical offices. I just feel this is a very sensitive
[00:43:41] transition period even though in the IMF this is being done systematically with enough resources with enough skilled staff etc. What I'm flagging is uh you know there are statistical offices where maybe
[00:43:58] there isn't such a systematic approach and where this transision can pose risks in terms of introducing um introducing flaws in a system that over time developed to be systematic when you look at all the quality assurance processes
[00:44:16] etc. Now you suddenly need to have completely new systems. Do you have any thoughts on this? Do you have any advice to statistical officers that are actually not well resourced and where the statisticians themselves may
[00:44:30] feel extremely vulnerable? Um where quite often the government is the main employer in a country.
[00:44:45] >> Thank you Isabelle Marco. Uh yeah, I mean the issue of uh uh resources um definitely I mean it's it's it's it's important for for all countries and and organization. So I don't I don't want to uh uh sort of un understate that. Um
[00:45:10] however again it's a for me again there are as I said in the first slide there are challenges and opportunities and I mean what what you focus on of course are the challenges how you get the you know sort of the buying from from from
[00:45:25] leadership also the national context and definitely there is a there is a for me when when it comes to AI because the benefits are so widespread uh the um the buy should come from from earlier. What what what we are seeing in our
[00:45:40] organization at least from MF is that when you when you propose something related to AI, you you get buy in from from say um the sort of top of the organization through the uh you know the kind of employees at all levels, managers,
[00:46:00] mid-level managers because it's such a powerful tool that uh really it's a addressing the risk together. uh using AI respon responsibly um is a u is is really sort of the the way I guess this should operate. So, so I guess and I
[00:46:20] talk about resources but in terms of you know how how to address the challenges I guess is sort of to work together at all level organization perhaps to build something that is called like a responsible use of AI and so that's
[00:46:32] that's done by many organization we have done our own sort of work at the MAP again because it's it's not that you know you put out there LLM tools and you upload this uh say confidential data that you in
[00:46:48] organization and and and then there's a risk of course as this data get get leaked out there right so I think that's that's important it's also an important uh individual component in this so uh each each um sort of employee or
[00:47:02] organization should be responsible how this this uh system uh are used um so and then of course as I said working together I guess is important so being being open to to participate in this initi atives
[00:47:18] understand what's going on out there. I think organizations like MF also in Africa we have we are partnering with the African Development Bank. Uh they have this fantastic platform called open uh data for Africa uh who's also like
[00:47:36] based on SDMX will be of course integrated with AI um in the near future. So th those are like so elements that perhaps reduce barriers like to to uh accessing uh AI uh even in low uh say statistical capacity systems.
[00:47:56] Um so again it's I want to sort of go back to my last point in the slide right so working together reaching out and and see also uh how uh kind of partners can uh uh or peers can also uh help in this transition. I hope this answer the
[00:48:11] question. Thank you. Thank you, Marco. Um, I hadn't I saw in the chat uh another question that I would like to raise here. So, are these examples of the benefits of using AI to broaden the reach and use of statistical
[00:48:35] data? especially is there evidence that this data can then be used by a less data literate or non-technical audience. Uh that's a good question. I I don't think there is evidence. I think if there is evidence is anecdotal evidence.
[00:48:59] Um look look if you um one one of the uh I think uh maybe non-tactical audience that uh for sure is using official data through uh generic LLMs uh is um is students you know you know when you look at the use of chachu pt or
[00:49:24] code um there is a peak uh during school years like when school is on and then like the do usage data drops during summer months. uh so that that that that means that you know it's it's essentially that
[00:49:43] students not using these LM tools and I would imagine that students um especially maybe college students um especially those working in economics I would imagine that they get their data you know very easily right from this
[00:50:01] system so so I guess this is anecdotal there's no evidence um I'd say maybe also going the the media maybe you know journalists can can also use this uh uh sources for for writing articles there's also some
[00:50:20] anecdotal evidence of um research papers that are essentially written through AI uh I remember a paper actually that identified I don't know it was a like a a a very high percentage of papers that were uh submitted to a journal uh uh
[00:50:39] that was tested like these like AI detectors and and the percentage was very high was like over 30%. And and you know there's also an increase in this production of working papers uh because AI makes so easy to
[00:50:55] write and so again I would imagine that those working papers not only generate text but they can also generate data. So uh again it's anecdotal evidence it's not you know factual evidence but this is actually an area that would need to
[00:51:11] be also explore like what do we have evidence about uh kind of uh uses accessing statistical data through uh generic LLMs that could actually uh you know strengthen the case for something like stat. Thank you for that question.
[00:51:29] >> Yeah, thank you Marco. Uh Sean wanted to come in once more. Please go ahead. >> Yeah, I got a I got a question. I just wanted because I I can sense from some some of the questions uh I think Isabella asked something of this nature.
[00:51:42] I wanted to get your opinion say in five five to 10 years from now we look back on today. What what were some of the biggest risks facing official statistics and the use of AI that we didn't see or that we are overly optimistic about? You
[00:51:56] know, so what are some what are some of the potential not knowing we can never do this perfectly, but how how where should our eyes be more skeptical as we start using officials AI and official statistics for instance the you know the
[00:52:09] nondeterministic nature of AI systems. Is this something uh or is it is it is it an underuse? Is is the risk that we don't fully adopt AI fast enough? Um because there are lots of ways this could play out and no one really knows.
[00:52:22] But I'm just as as an expert from a leading international organization, I'm curious what your thoughts are about what these risks are that we might not be fully seeing. >> Yeah, I I think the biggest risk for me
[00:52:35] is not doing anything. So if you just uh let others uh take the lead and and not being involved in these conversations and just do our work um like we used we are used to. uh we uh yeah we in 5 years time um you know the uh
[00:52:57] like access to our data to official statistics data through uh national website would drop dramatically to a point where nobody would go there. So uh again it's a I think experience like um say MF GPD but also in the MCP at
[00:53:14] testing out things uh and see how it works. So it put ourself into into the picture right. So I think the statistical community should take this as as a serious uh um sort of task going forward uh because AI is a pervading every
[00:53:33] aspects of life right so it's not only just statistics statistics but everything and and look we we there's also there's been already the big data revolution that uh know it's a it's putting statist official
[00:53:47] this is a risk uh we have facing declining response rates in in the major surveys in the world like you know labor force survey or enterprise service household survey so um again there are questions in advanced countries like the
[00:54:05] UK US about like quality of this kind of survey data so so again that for me the big the biggest risk is is not not acting um there's also a risk of uh not working together uh which is again kind of experimenting and and and doing
[00:54:24] things by by like itself without collaborating there's a risk of dis diserting dispersing energies and money uh unnecessarily. So this is an area where and this why it's important to have this congregations
[00:54:44] of uh like groups that work together and also both within groups but across groups uh because that's where the synergies are but I always say that working together has a cost right so uh you know if if if one day you wake up
[00:55:00] and say I want to do this you do it uh there's no cost except your your sort of time but if you want to say okay I want to talk to Sean, I want to talk to Alexander. Let me send an email and then I wait for your response and then we
[00:55:13] meet and then a week later we haven't done anything. So there's always a cost in collaborating but I think for things like this collaboration eventually will uh bring benefits even if it takes perhaps a little longer.
[00:55:29] >> Thanks Marco. Um, I would like now to to hand over to Ammon, the branch chief of the data innovation and capacity branch of UNSD for some closing remarks. >> Thank you. Thank you very much, Alex. And thank you and your team for steering
[00:55:50] and moderating so heavily this webinar series. And thank you, Marco. Thank you for joining us and thank you for your time and for sharing your insights. really appreciate it. Uh I think you uh you kept uh our global network webinar
[00:56:06] series on AI today. All right. Uh which run across five presentations during the whole months of June. All right. And I think uh we really appreciate obviously the time of all those like you who invested their time to share their
[00:56:23] insights. Now uh when we tried to conceive this global network webinar series on AI uh what we try to do as I said at the beginning we were really hoping and trying to move from a kind of a general conversation on AI to
[00:56:43] a kind of telling a real story moving a little bit closer to implementation. uh having kind of a coherent approach and in doing so we were trying and and hoping as well to demystify AI a little bit for a lot of the statisticians, the
[00:57:04] statistical community, the community, the NS >> in general and obviously particularly the NSO's in the south uh which have some resource constraints as we heard from Isabella and others during the course of the u of the series.
[00:57:21] >> Now, it's difficult for me to obviously try to wrap up in five minutes [clears throat] what has been discussed in five uh weeks and in five hours. But let me try to venture a little bit to get some of the takeaways and obviously
[00:57:38] uh myself and my team we're going to go back to the board and do a more in-depth reading of all this uh presentations and come back hopefully to you uh with a more meaningful summary that I'm making now. But I think one of the key
[00:57:54] takeaways that uh at least intrigued me of course and that came strongly from the beginning is that good governance is key for the use of AI. And when I speak about governance I speak about two aspects of governance and they are
[00:58:11] different in a way. One is the institutional governance and the other one is the data governance which are obviously difference and totally uh kind of uh but they come together and they complement each other. When I speak
[00:58:25] about institutional governance, as uh we as we learned from the NC presentation which kind of started this series by Ken Post and Condrant, uh we learned that having a strategic approach like NC had through a strategic
[00:58:43] committee all right for policy pilots and technology choices and that have an authority over all these aspects is key and it's key to ensure what Isabella has mentioned as a systematic approach. All right. And a more strategic approach to
[00:58:58] the use of AI in the institutions. All right. And underlying this kind of strategic committee, we had kind of another enabling structure which enabled constant interaction between different communities, the technology community,
[00:59:17] the statisticians, the data scientists etc. All right. and constant interaction were happening and were conducive through another enabling structure. Another thing is that the focus of innovation was continuous as part of
[00:59:32] this kind of structural approach as well. Uh NC as we know had two innovation teams in 2017. They offered us the Onexia platform which we are thankful for which we are using in the UN global platform and they continue to
[00:59:47] have their AI lab. All right. Which continue to experiment and really uh to focus on testing and experimentation as they move forward. So uh so that's I think a key take uh in terms of the institutional uh governance on the data
[01:00:05] governance and this is really when we speak we speak about the foundations of trust here and I think we heard uh we heard very clearly uh from our colleagues uh Matias [clears throat] from Netherlands Dave from openminded
[01:00:20] Rafael from Switzerland Benjamin from Canada on the privacy enhancing technology and AI I and we knew that we learned from them clearly uh that privacy and utility are not a zero sum tradeoffs and there is a balance to be
[01:00:39] continuously sought between privacy and utility and as they recommended through their experience the starting point for many of them has became utility and adding layers of privacy as they move forward in terms of the data sensitivity
[01:00:55] and the need to protect the data. They provided us with a lot of tools that exist. All right. Synthetic data, differential privacy, confidential computing and federated learning that let us extract value while exactly
[01:01:09] protecting the data uh sensitivity in doing so. So I think this is another lessons learned and another main aspects of this data governance was the shift in mindset of uh to data and the approach to data
[01:01:26] sovereignty uh which we heard from Carla from the world economic forum and I think on that we learned that data sovereignty is not achieved by geography or by isolation but I think we should really think it in terms of operational
[01:01:41] control across the full data cycle and AI life cycle. And I think this change of mindset in is essential as we move forward uh to data governance and AI governance from who owns the data to who has which rights and under what
[01:01:58] conditions. So that's another uh kind of takeaway that I would like to emphasize. And then we moved to the practical use of AI in the production cycle of official statistics. And we heard fantastic presentation from Brendan from
[01:02:15] the Cso Island where again he presented so many use cases after kind of uh conducting really an AI maturity assessment. uh and these use cases which were related to data generated uh data generation, data analytics, data
[01:02:36] processing and also responding to concrete policy uh demands uh brought us to something that was kind of almost given. But as I said in that uh specific webinar, I think uh problem definition is key and what we are trying to achieve
[01:02:59] through a deep a use case of AI has to be thought of very strategically but also very concretely and with a costbenefit analysis that is real and I think that goes back to the good governance from the very beginning as
[01:03:13] well. And then we ended up with you Marco today about the uh AI as a dissemination and how AI is transforming dissemination. And I thank you again for your presentation today. Uh we know and we saw from your presentation how AI is
[01:03:29] kind of democratizing really access to data and that again uh the word trust that came several times through the webinar series. All right remains really the key distinction for the official statistics and this that's why we need
[01:03:46] to get uh our official statistics AI ready uh in a right manner because we owe it to the public uh and and we owe it to this democratization of the data as well. So uh I think with this I have to say and I'm glad to say that uh
[01:04:06] what's next what's next is obviously uh I have to say that we are happy that the UN statistical commission obviously took up this this important topic uh through its last uh Friday seminar uh before its UN
[01:04:25] commission 57 last March. uh they endorsed the establishment as you alluded Marco to a city group which is known as the Kali city group as it is led uh by the National Institute of Statistics in uh of Rwanda. All right.
[01:04:42] Uh they convened their first meeting as steering committee and they put their structure in place. uh so uh they will come back to the UN statistical commission next February March uh with clear hopefully of the consultative
[01:04:58] process dos and so on and I think the wealth of bodies that are involved in in in the AI use and AI dissemination and AI readiness as you mentioned them Marco in your presentation uh will all going to be obviously stakeholders and
[01:05:15] hopefully will plug in into the city group in a meaningful manner so that Leo statistical commission and the broader statistical community will be appraised of what's happening uh across the board and through different stakeholders in a
[01:05:27] coherent manner. Uh from that I'm also happy to say that uh the UND and my branch in particular through our data innovation section is involved in trying to make use of AI to the benefit of the NSOS and to the member states as much as
[01:05:44] possible. Uh congratulations Marco again for your stat GPT. It's it's an excellent example. All right. And I think uh our data innovation section is working exactly now to do similar things. All right. With the LDG's UN
[01:06:01] data all right to make it accessible exactly like ST GPT is doing to the uh LDG's uh data in through natural language and make it really interactive and so on. Same thing for the UN data commons uh where 26 UN agencies are
[01:06:18] involved in now. All right. And which could be uh one uh window shop for the uh trusty data. All right. Uh related to so many agencies uh and and I think that's going to be a our small contributions uh in this kind
[01:06:37] of AI story. uh more importantly we will have obviously to go back and consult and through the city group see how uh capacity building programs for AI can be done and capacity building activities of AI could be developed to support NSO in
[01:06:54] the south and so on and I think we're going to be taking up this as we move forward let me close up because I spoke too much a little bit by saying sharing with you some numbers all right we have more than or close to 900 people who
[01:07:10] actually lively uh followed uh our uh webinar series on AI. We also have 600 reviews for the recording of the four webinars alone and we have more than 710 different people registered for at least one of the five webinars including
[01:07:30] today's webinar and I think if anything that shows that there is still a lot of appetite uh for uh for the AI use and the AI application for official uh [clears throat] data. uh we will continue working with you and conversing
[01:07:43] with you and thank you very much for your engagement and uh the conversation will continue. Much appreciated up to you Alex. >> Thanks Summer for those uh closing remarks and just before we um all go um
[01:07:59] can I ask everyone please turn on your cameras and the microphones just for a few moments and then we give uh Marco a big round of applause. So, please join me in switching on the camera and having a nice applause for Marco. And uh
[01:08:17] as Am has mentioned uh while we had five webinars in the series, the links to the recordings are in the chat and also uh we will also post the slides and the recording of this webinar today on the global network and the link is in the
[01:08:36] chat. Thanks everyone for joining. It's also so nice to see so many uh faces here on the screen right now. Thank you. >> Bye. >> Thanks so much. >> Thank you. Bye. Bye.