Thank you for rejoining us. Our keynote speaker this afternoon is Governor Andrew Bailey of the Bank of England. Governor Bailey has served in his role since March 2020. He is also the chair of the Financial Stability Board. Before taking office, he was chief executive of the Financial Conduct Authority and earlier served as deputy governor for prudential regulation and chief executive of the Prudential Regulation Authority. Please join me in welcoming Governor Bailey.
Well, thanks. Oops. Thank you for that introduction. It's a great pleasure to be here today in Saudi. I have to confess that it's my first time here, which really should have been taken care of a long time ago. So I will definitely promise to come back, as it clearly is, even from the things I've seen so far, very beautiful.
I'm going to start with what I see as the key current judgments on financial stability, but I'm also going to touch on monetary policy because you can't, I think, divorce the two things. And then I'm going to look back at issues underpinning those judgments for many of us and what they give rise to.
I'm going to draw for the first part on the Bank of England's latest financial stability assessment, which we published just over a week ago.
The continuing impact of the conflict in the Middle East has created major uncertainty around economic growth, inflation and the path of interest rates in many countries. Sorry, my... And there has been and continues to be an intensification of the risk that already existing vulnerabilities in sovereign debt markets, asset valuations and credit markets will crystallize. At the same time,
the rapid increase in debt related to artificial intelligence has increased exposures related to those developments. At the same time, recent incidents in frontier AI have drawn further focus to the pace of that development and associated vulnerabilities, including cyber and operational risk.
We turn to the financial system and its structure. The story, I think, so far has been one of resilience. The financial system as a whole has been resilient in response to the increases in sovereign bond yields, and market adjustments have been mostly gradual. Banking systems generally appear to be appropriately capitalized, and nor are we seeing stress in funding.
It's important, I think, to point out that this level of resilience in banks and the benefits that economies get from that is in very good part due to the reforms undertaken since the global financial crisis.
At the same time, during that period, there has been a pronounced and sustained increase in non-bank financial activity, something I'm going to come on to later in my remarks.
But so far, the system has dealt with the impact of these economic shocks without a major impairment of trading conditions and liquidity.
Now you may say, well, markets are quite volatile though, and they're moving a lot, and that's true. Market movements are some way from normal, but we're not seeing stressed conditions. We're not seeing illiquidity, and we're not seeing disruption to two-way markets.
But I will add two important qualifications to this. We can't in any sense be relaxed about the current state of financial markets. We must be vigilant at all times.
We can in no sense think that what we've seen so far is a guide to what we will inevitably see in the future.
The second thing I'll say is that even against a background of overall open trading conditions, open two-way trading conditions, there will at times be individual events. They may not be systemic events, but they will of course require dealing with. I'm not going to comment on gilt, but I know you've had one recently. We had one in the UK a few years ago as well. And my experience has taught me that however stable the general environment in which we operate, there will always be these issues and they have to be dealt with.
There is one thing I think that we learned though, we learned it from the global financial crisis the hard way, that when you have to deal with these specific issues, it is much easier done against a background of generalized stability, and it's much harder if you don't have that situation.
So as a general matter, yes, we are seeing volatile markets. We are seeing shocks, substantial economic shocks, but markets are operating.
But we can't leave it. I'm not going to leave it, I'm afraid, at that point. That is not the overall conclusion, I'm afraid, because we do face very real and serious challenges.
And I'm going to use most of the rest of my remarks to lay these challenges out as I see them and some thoughts on what we do about the situation. And it's very timely because these are certainly thoughts that I'll be taking with me to Washington next week for the annual meetings of the IMF and the World Bank.
So, let me start with the big picture. Since the global financial crisis, I think in many countries, we've seen two big economic developments which are linked, and it's certainly true of the UK.
The first is a lower potential growth rate mainly due to lower productivity growth. Now, I should note that I realize that not all countries are in this position.
But that lower growth rate and that lower productivity growth matters a lot, quite simply because in many settings productivity growth drives living standards.
The second big development of recent times has been the frequency of very large negative supply side shocks.
COVID, Russia's invasion of Ukraine and of course now the conflict in the Middle East, to name obviously some of the most prominent ones.
I think the two military conflicts are classic negative supply side shocks. COVID, I would argue, was both a negative supply and a negative demand shock, with the balance between the two changing over time.
But the key point is that these have been very big shocks to our economies. We've moved into an era of very big shocks.
Moreover, the frequency of their incidence has been unprecedented in recent history. That's not true if you go further back in history, but it's been unprecedented in recent history.
And a couple of weeks ago, I was doing a conference and I was asked if we should treat this frequency of big shocks as the new normal.
My initial reaction to the question was a mixture of, well, we could talk all day about that question, and as I'm a central banker, not a political soothsayer.
But I added that a political economist may answer that we now live in a multipolar world in which the poles are competing on economic territory. And if you look through that lens, you may say actually, it's not unreasonable to think that these sorts of shocks will occur.
But we have to then be very clear what this accumulation of shocks and their frequency means for overall economic policy.
So let me start with monetary policy, the obvious place for a central bank to start.
A negative supply shock pushes inflation and output in opposite directions, whereas a demand shock pushes them in the same direction.
Supply shocks therefore create what we, certainly in our regime, call a trade-off between inflation and output.
The textbook response to a single negative supply shock is to look through a one-off price level rise, since monetary policy acts with long legs and moving interest rates will not solve the underlying problem. It will not cause oil and gas supplies to increase.
But there's a big but. This look-through approach only works if inflation expectations stay anchored to the target.
If households and firms start building the shock into their wage and price setting behavior, a temporary effect quickly turns into a persistent effect and generalized inflation, something we tend to call second round effects.
These second round effects aren't necessarily easy to identify in the here and now. There are lags involved, ones I mentioned earlier. So as a central bank, we have to make a forward-looking judgment here and we have to keep refreshing that judgment every time we meet to make a decision on interest rates. Of course, that's our job and we must do it.
But as a further complication, I think as a feature of the present context, a recent history of such shocks and thus a recent history of higher inflation can make it more likely that the public will raise its expectations of future inflation.
So the textbook look-through approach is much harder to sustain in an environment where the shocks pile up in rapid succession.
And that's exactly what happened with the Ukraine war breaking out, following on so rapidly from COVID.
In the UK at present, we're seeing the first round effect of higher energy prices and inflation.
And we're conscious that the history of recent years points to a risk of elevated inflation expectations.
That said, the evidence of pass-through to wider inflation, the second round effects, is currently quite subdued.
But it's early days and the longer the pattern of higher energy prices persists, the more difficult this situation becomes and the risk of stronger second round effects increases.
So we have to watch this extremely carefully.
Let me turn now to financial stability. Negative supply shocks can require monetary policy to be tightened alongside a weakening of economic activity.
A combination which puts stress on balance sheets by raising debt servicing costs at a time when household and corporate earnings can be under pressure.
Likewise, such a combination can cause assets to reprice as market values of corporates fall, and this can translate into lower values for the collateral which is used to secure credit supply.
And in this type of situation, a tension can break out between monetary policy and financial stability policy.
But it doesn't have to be like this. The pre-existing resilience of the financial system and of balance sheets across the economy, households, corporates and governments will come even more into play at this point.
If the starting point is basically sound, this element of the tension that can arise in the event of negative supply shocks can be managed successfully.
Now, as I said earlier, I think we do have quite a lot of basic resilience in the financial system to draw upon, which we've built up since the crisis.
I can speak to the UK here. We have a lot of resilience in household and corporate balance sheets as well.
This almost naturally brings me to fiscal policy. Now, to be very clear, I'm a central banker, so I'm going to tread very carefully at this point. It's not my job to pronounce on what fiscal policy should be. So what I will say, you may think falls into the category of stating the obvious, but doing this sometimes is important given the conditions we're seeing in government bond markets at present. I think it's more important than ever.
Lower growth rates and negative supply shocks have the effect of squeezing fiscal policy in two ways.
They tend to weaken the previously set course of fiscal policy while also leading to pressure to cushion the impact of shocks.
Higher borrowing costs and weaker growth raise debt to GDP levels, and at the same time calls for spending to increase are louder.
And if markets doubt the fiscal path, yields rise further, which then feeds back into monetary and financial conditions.
And the sequencing of negative shocks is also relevant in this context.
Classically, fiscal policy can operate as a countercyclical tool of policy. In other words, governments can step in to cushion negative shocks and then switch back and reverse in the better times.
With such an unexpected and severe shock, and let's take COVID as an example, people can expect governments to come in and provide support, and in doing so use the government balance sheet. And many governments did that, including our own, not unreasonably.
But if the negative shocks come at higher frequency and underlying growth in the economy is lower, countercyclical policy becomes very difficult to follow successfully.
Now I think the government bond market strains we're seeing at the moment have a strong element of the impact of the overall succession of negative supply shocks and low underlying growth rates and this difficulty of operating countercyclical policy.
Now I'm very aware that that diagnosis does not... it's good to diagnose, it's always good to diagnose, but it doesn't solve the problem. But I do think it's important that we lay out the assessment carefully.
I do think that on the whole we get better responses in terms of policy if we do so.
So this brings me for the remainder of my time to come to the response to this situation that we find ourselves in, this situation of lower growth, negative supply shocks and the consequences of them. But before I do that, I want to add one other element of the situation which I mentioned briefly earlier, and that's market structures and financial markets.
Many government debt markets, the UK included, have changed profoundly in recent years.
They've moved from being dominated by longer-term investors, often known as real money, fairly price insensitive, and instead we now have markets that are much more dominated by elevated leverage, an absence of traditional buyers, and particularly traditional buyers of longer duration debt.
This last development, certainly in the UK, is associated with another fundamental development in many societies: the impact of aging populations on pension fund accrual.
So there are a number of important consequences of this transformation in government bond markets.
First, and I will be quite direct in saying this, it has enabled governments of all descriptions to issue more debt.
The structure has shown greater power of absorption, a function I would argue in good part of the greater leverage in it.
But I'm afraid that this greater absorption is into a structure that does not look as robust in a fundamental sense.
Yes, trading conditions continued to be orderly in the face of the most recent shocks, but that was not the case in the early days of COVID. And as I mentioned earlier, in the UK in 2022, we had a particular experience that we had to deal with, something we dealt with, but again, it was an illustration of what can happen in these much changed market conditions.
And the reason is that heavy leverage positioning is inherently more fragile than the older structure of longer-term investments.
And I don't think we can ignore market structure and market conditions.
The risk is that we can face conditions where heavily leveraged positioning shifts rapidly to sudden deleveraging flows, leading to further dynamic losses as models reprice and stop-outs are triggered.
And I would add one further point here. The fundamental change in market structure has also led to a rise in the interconnectedness of financial markets across countries.
Put simply, the same funds, and a relatively small number at that, are often taking big leverage positions in several markets. This creates the risk that sudden deleveraging in one market could transfer to others. This is something that we're very focused on in the Financial Stability Board because we do not have good enough data and it's not transparent enough at the moment about that situation. And I have to be honest with you, I have said to my colleagues involved in this work that we are not probably going to leave the room until we have made progress on this. It's too important.
So what else should we do about the situation? Well, let me give some thoughts, but I'm also going to add one spoiler at the beginning. I haven't included bringing about world peace. It's definitely outside the remit of a central banker, but it would be a good thing. So I will limit myself to pointing out that working to solve the causes of the negative supply shocks and limit the potential for further shocks in the future has the added benefit of helping to deal with the economic challenges we face.
That's a sort of statement of the obvious, but the obvious often has to be said.
Next, we need to take all the actions we can to raise the sustainable rate of growth in our economies and thereby improve supply side conditions.
I've talked a lot about negative supply shocks. Of course, there are such things as positive supply shocks. Many things influence growth in our economies and we should pursue all of those things across a broad front. But there isn't, and there isn't a single answer. It would be a mistake to assume that there is a single answer to the growth question.
That said, I do think that over history, discrete major advances in technology, often known as general-purpose technologies, have made very important contributions to lifting growth rates in a sustained way.
That happened before the global financial crisis with the contribution of ICT and the internet. In growth terms, that contribution ran its course and had probably begun to ease off before the crisis got underway.
But this obviously begs the question: what is the next big technological advance that might create this positive supply effect?
Well, I think the answer is probably fairly obvious. It's a combination of artificial intelligence and robotics.
These technologies have the potential to drive scientific discovery, enhance productivity, and contribute substantially to prosperity.
Now, I'm not naive. There are important issues to be dealt with on how to realize the benefits safely, and we're working, certainly with our governments in the UK, very hard on this.
That's another, I think that's a subject for another day, but it's an important one.
AI is also playing an important part in financial markets in terms of capital raising to finance the development and deployment. That's natural. The rapid increase in AI-related debt issuance broadens the exposure of capital markets to those developments in AI.
And I think an important point here is that as these financial dependencies increase, so the linkage between AI security and safety and financial stability will take on another dimension. That's another dimension as well as the cyber risk type dimension that we're already dealing with.
Because if market valuations and conditions become more sensitive to the overall standing of and trust in AI, the connection to financial stability will grow in prominence.
Now, so far equity markets in aggregate have been resilient to pretty much all the economic news, including events in the Middle East. There was a quite sharp but temporary sell-off in AI-related and semiconductor stocks in July. We saw evidence of some deleveraging and portfolio rebalancing. But I would say that looking across the whole of the period, we are still seeing resilient markets.
And I think the lesson here is that financing investment in technology and thus growth is critical.
But we must remain focused on and test for the consequences of a possible significant shock to earnings and growth expectations, reflecting concerns around the pace of AI development and adoption.
Growth prospects and fiscal outlooks depend in good part on expectations that AI development and adoption can generate substantial and significant productivity gains. We may get bumps along the road, but safely deployed, I would put AI very much in the set of things to pursue to help alleviate the current economic challenges that we face.
The next set of actions we should take to tackle these challenges brings me back to the core of macroeconomic policy, monetary and fiscal policy.
It's hard to produce a single answer which explains the elevation of bond yields at present.
But the contributory factors are not that hard to identify. Markets are displaying elevated risk premia.
One contributor to this pattern is likely to be the uncertainty around the duration and scale of the current supply shock. How long will the conflict go on for? How serious will be the consequences of it looking forwards?
But it also points to the need to ensure uncertainty around the course of policy is kept to a minimum.
For monetary policy, the objective is to bring inflation back to its target and to be fully committed to do so, which we are. Our credibility is based on that commitment.
As I set out earlier, the recent frequency of shocks makes it important to watch for the evidence of embedding of expectations of inflation, i.e. that policy will not return to target in the time span that we think it should. And communication is important here, backing up the course of policy actions themselves.
Now, I have to say I'm somewhat in the skeptics club when it comes to unconditional forward guidance on future interest rate moves. I think the world is just simply too uncertain for that.
But I'm a strong supporter of explaining ourselves conditionally in terms of what we will do if the economy and thus the course of inflation evolves in a particular range of plausible directions. If you like, telling stories which help to explain in conditional terms our actions.
But of course with that of greatest importance is that we reinforce our commitment to the inflation target.
Now as I said earlier, it's not for me as a central banker to comment on fiscal policy and thereby cross the line beyond central banking. So I'm only going to make one point which I don't think does cross that line, which parallels I think the commitment to the inflation target.
That is, whatever the stance of fiscal policy is, it must be credible and directed at stability and be seen to be such by markets.
A continued commitment to the policy stance can then help to control the emergence of risk premia, something that is likely to happen otherwise when a major conflict such as the current one in the Middle East breaks out.
In other words, such commitments are needed more than ever when these negative shocks occur.
Fiscal rules can have an important role to play in securing this outcome, just like inflation targets.
My last area of action concerns market structures, particularly in response to the major changes in the structure of many government bond markets, because it is important that this leverage in core financial markets is well managed.
In the UK, we will be taking forward work on measures to improve resilience in the gilt repo market as an important part of managing risks from the increase in market leverage.
There is a broader dimension to this work including consideration of market-based measures such as greater central clearing and minimum haircuts.
As I noted earlier, so far we can say that the consequences of the latest negative supply shock, the Middle East conflict, have not included disorderly financial markets, but they are volatile and we have to watch them very carefully. And the risks from large and concentrated positions and the potential for greater interconnections across markets to trigger wider disruption mean that we cannot be relaxed at all on this front.
So let me conclude quite briefly. We face difficult conditions.
The proximate shocks are large. The duration of them is highly uncertain.
Moreover, the cumulative impact of the successive shocks is a serious issue.
There isn't a simple quick fix, but there are important steps that we must take to raise growth rates, to ensure our macroeconomic policy frameworks are fully committed to achieve the stable conditions that can best support growth, and to take action to limit the scope for wider disruption within financial markets.
These are very important issues in today's world. Thank you.
Thank you, Governor Bailey, for your remarks. Our final session today considers artificial intelligence and the global economy. AI may influence productivity, potential output, and labor markets. The policy challenge is to encourage innovation while maintaining sound governance, accountability, and public trust. The session will be moderated by Deputy Governor Fatih Karahan of the Central Bank of the Republic of Turkey. She'll be joined by Governor Jasmina Selimovich of the Central Bank of Bosnia and Herzegovina, Chief Executive Eddie Yue of the Hong Kong Monetary Authority, and Deputy Chairman Nodir Aulov of the Central Bank of the Republic of Uzbekistan. Please welcome the moderator and the panelists.
Hi. Hi. Welcome to our session on artificial intelligence and the global economy. I have three great panelists with me today: Jasmina Selimovich from Bosnia and Herzegovina, Eddie Yue from the Hong Kong Monetary Authority, and Nodir Aulov from Uzbekistan.
As discussed earlier today and also as mentioned by Andrew a few minutes ago, as the global economy we have been going through various shocks within the last decade, along with an ongoing decline in potential growth, which can also be described as the speed limit for economies. And the question has been if this decline could be reversible through various policies, right? And now that we have an accelerated AI process globally, its impact on potential growth is increasingly questioned, and this naturally connects with monetary policy and central banks. So on this panel, we're going to discuss the issue of AI from a holistic perspective. And my first question goes to Eddie. Eddie, do you think artificial intelligence would change the economy's speed limits and potential growth? And if so, how might this affect the central bank's inflation targeting frameworks? And do you think eventually these changes would be incorporated into monetary policy rules?
Right. Thank you. Happy to say, in the short term, of course, AI will have a very positive impact on cyclical growth. You know, we talk about the AI capital spend in the US. In Asia, we generally benefit a lot more from the trade on electronic goods, a lot of which are manufactured in Asia. Just to give you the example of Hong Kong, in our exports, about 70% of our exported goods are actually electronic and tech-related goods. So with this AI boom, we are actually seeing our trade figures going up, exports going up quite substantially. From an increase of 15% last year to this year, we've got an increase of 40% in our exports, mainly riding on the AI boom, and that actually got our first half economic growth to about 5% this year. So in the short term, there's absolutely a positive impact.
But the more long-term structural impact will be a little harder to tell. Of course, the automation, the saving, or the efficiency brought by AI and robotics will definitely boost productivity and growth in the long term. And of course, as to how much individual economies will gain from that will depend on the relative size of your cognitive-related sector and also the AI preparedness of your economy.
But what also happens on the flip side is about the displacement of labor. And if you displace labor from the market, there will also be impact on demand, on their consumption. Whether that will actually offset part of the growth or structural growth that you're expecting from AI is actually hard to tell at this stage.
And we've looked at the labor displacement issue in Hong Kong, very narrowly on the banking sector. We expected that quite a large number of our banking practitioners could potentially be displaced. We're not seeing that yet. Maybe it's just the start of the use of AI, or maybe banks, upon our advice, are actually starting to reskill and redeploy their current practitioners who are subject to risk of AI.
But we also have taken a look at the new jobs, especially for the lower-level or middle-level jobs. In the last year or so, it's actually very clearly declining. So fresh graduates are getting harder to find entry-level jobs, and that means that for them to accumulate the needed human capital to boost future productivity and future growth will be a lot more limited. So this is an area where there's a lot of uncertainty on whether this weaker human capital formation can actually potentially reduce our potential growth.
And on the monetary policy operations, I would just say that just like all the uncertainties that we see, we actually don't know the extent of productivity gains coming from AI and when that will come in. So there's got to be a lot more flexibility in thinking about monetary policy operations, distinguishing between the short-term, probably more growth-inducive impact and inflationary impact from the longer-term structural, more disinflationary impact, without knowing where or when the impact will come in. So flexibility, clear communications is probably the key word.
But I'll just end here by saying that monetary policy aside, we might actually need a bit more coordination with both the fiscal side and also the education and manpower side to think about the labor displacement issue. We are doing our part on the financial sector by almost requiring banks to have manpower planning in the light of technology and implement that. The government is doing a lot on education. And I think that will be a big issue that we need to face.
Thank you. Obviously, AI is likely to affect monetary policy through multiple channels, but we need some time to see the real effects. Well, AI is a strategic instrument for our economies, obviously, and admittedly, innovations of such magnitude rarely occur, and when they do, they also bring some great challenges. So in this respect, I'd like to continue with Nodir, a similar question. And how do you see AI affecting the macroeconomy, and what does that mean for monetary policy? Do you foresee any challenges?
Yeah. Thank you very much, first of all, for inviting us to this first international economic forum of the Silk Road, the city where all these developments and the regionals are connected, and Silk Road country as Uzbekistan. It's very kind of the connecting history and connecting the partnerships of Istanbul is very important. And yeah, your question is very relevant in today's world because nowadays changing overall environment where the agents are working on it in a way that it's becoming faster to change prices.
For example, we also observe in our grocery price big companies and also the fintech companies are very sensitive to the changes of the economic policy, very sensitive to changes of the other big shocks because they have developed their tools, their analytical models based on AI and their information. And at the same time, it kind of faster their reaction functions, price-setting functions. And this is one part of the question when the overall AI influences the price-setting behavior of the agents.
Second, it comes from the, yes, productivity. Maybe there are a lot of investments, and also that we're witnessing in other countries that big fintech companies are investing a lot. And also they're not just investing, they also accumulate the very good kind of talents of the economy in one place and paying very high salaries and investing in the overall infrastructure, servers. And these investments also in the short term may have some kind of influence to prices, cost shocks and demand shocks.
But at the same time, we should exactly don't know yet the productivity when it comes to these sectors because of this, if it will be some kind of gap between them, it can influence also the credit market. It can influence also their overall pricing behavior.
And another very important issue, yes, we should also get ready for this kind of changes from the central bank perspective to kind of formulate tools, to formulate the instruments to be capable to answer these questions for a lot of positive changes also coming to the economy. For example, stablecoins are very popular nowadays, and we have also implementing some private ones with the private companies, also CBDCs, also these kind of all changes tries us and also to push us to more investment in investigating these kind of relationships at the same time, also to bring additional knowledge from our foreign partners, from the global experience at the same time.
Other big companies doing accumulate also the very strong potential, kind of very good minds of the human capital to the central bank in order to prepare all preparedness for the future chances. But at the same time, if in the short term we can see some kind of steepness of the Phillips curve because of the elasticity of the price changes at the same time from another side.
As this optimization comes, it will be less pressure from the wage side. It can be some kind of not push from the cost push side to the Phillips curve, and we should analyze this potential, which comes first, which comes later, and it can be from both sides. And that's why there are a long way to go, and also we should go this way with very good instruments and capacity.
Thank you. So apparently, as central banks, we should expect some challenges regarding monetary policy, but we should also take some advantages of artificial intelligence in central banking. For example, at the CBRT, we have an artificial intelligence commission, and we are leading various AI projects at the bank in this commission. And we have a simple framework based on three pillars: mind, pools, and core.
Mind is for macroeconomic insights and data for deriving deep economic understanding with enriched data. And pools is for payments, unusual leaks, and security is, you know, is critical to monitor the heartbeat of the economy through the payment system. And core is the third leg of the system, is for coordination and efficiency, mostly the backbone of central bank operations for, you know, efficient and reliable internal operations. And this is just our case.
But in this sense, I'd like to shift the discussion towards the use of AI in central banks. And my question is for Jasmina. From a central banker's perspective, how do you assess the benefits of using artificial intelligence in central banking operations? And I was wondering where do you see the greatest value: economic foresight, forecasting, or real-time monitoring, policy analysis? Which one of these? And have you adopted AI in any of these areas in your bank? And one last related thing is, you are welcome to mention the risks you find important as well.
Deputy Governor, thank you very much for your questions. First, let me thank the Governor of the Central Bank of Turkey for organizing an event like this. It's a pleasure to see good friends and colleagues, governors, here.
Let me start with one question that we think in my bank very frequently, and that's actually a dilemma: is it a greater risk in using AI or not to understand it properly and not to understand it on time? So honestly speaking, I don't know the answer. The only thing which is definite is that we are on a big frontier when it comes to central banking in terms of artificial intelligence or superintelligence that was announced two or three days ago.
I believe from our perspective, benefits and risks, they cannot be treated separately. Those are practically the same metal, but just two sides of the same metal. The benefits of AI, they are very clear, and they are coming from its ability to process a huge number of unstructured data at a high speed. But at the same time, that is something what is provoking high risks.
We know that AI is definitely suffering from hallucination. We can experience that easily in our everyday life when we type something in, it will probably give us the preferred answer, something that we like to see. At the same time, it has embedded bias, definitely. We have a low number of providers. So when we combine everything, we need to be not to choose, should I do this or that, either-or. It's not either-or exercise. I believe it's much more about pricing everything equally, so if you, or not equally, adequately.
So if you're asking me what are the biggest areas where we can use it, probably you mentioned the economic analysis, so information advantage. It's definitely one of the major areas for using AI. We know that we come up with the decisions under the uncertainty that information is always, not always, but usually late information. And then when we use the AI, the timing between actual event and our understanding of the event is definitely shortened.
The second thing you mentioned, the payment systems and the supervision. So that's another, so practically operational intelligence in a way, that's another area where we can, I believe, have huge benefits. When we have anomaly detection with AI, we can find the irregular patterns that regular rule-based systems cannot identify.
And something what is, I believe, very much used but somehow under the radar, it's something that has to do with organizational or institutional memory, because we have huge piles of data of analysis that become easily approached and we can easily analyze them using AI. Otherwise, we know what is the alternative.
So in the Central Bank of Bosnia and Herzegovina, to be very direct, we don't have predictive AI models that we use in production. Between 2025, at the end of 2025, we started with, we set up the team for AI. It was a multidisciplinary team, and actually we included colleagues from different departments in order to have as good output as possible.
They drafted the strategy with a very clear roadmap what to do, and we already started for the chief economist office, that was the first pilot to do some proofs of concept for collecting the data for the prices developments analysis. And our traditional models currently are relying on econometric models.
That's the way it should be. But now the question is, with AI, does AI replace the current model or it sets aside? So is it something that goes in parallel or it is complete something else? We identified the area for some sort of hybrid. So we have the AI model or AI output which we then analyze within our office and then we benchmark it to the classic one.
So in that sense, we can find out what is the difference. Because at the end of the day, AI is here not to replace us, not to think about instead of us, not to come up with the decisions instead of us, but to help us to do the due diligence and to help us actually to profile deeper when we want to reach some decision.
Thank you, Governor Jasmina. I think it's quite inspiring for all of us here to hear about the AI projects and work undertaken by other central banks. So let me continue with you within the same scope. Another bank, the central bank of, has been one of the most active banks in the region in employing AI applications, and I understand that you have an in-house application assistant called CBUGPT. So could you walk us through that journey? What tools have you built so far?
Okay, yeah, as a trend, we also invested in the overall capacity of the calculations, central bank and infrastructure servers, and formulated with the professional CBUGPT kind of platform and overall the model Central Bank GPT. It helps in very many dimensions. First of all, it's like monetary policy research, and it starts from the, like, as a research advisor, research assistant, it first helps to calculate, helps to collect information from the data sources, from our sources.
And we're inspired by the State GPT program of the IMF, the State Department, and they just officially used these GPT technologies on their data sets. And this first part of this research program is to be the official research assistant for economists, and then it turns to the research advisor. It helps to find the kind of the patterns. It helps to advise the young economists how to do analysis.
But at the same time, this workflow well coordinated with the kind of the peer economist, with the division chief and others in order to not go two ways. It's not necessarily. And it's another big portion of that we're conducting kind of the AI lab in the Central Bank of Uzbekistan. So this AI lab, the guys, the staff formulating synthetic personas for the, for conducting the surveys, inflation expectations.
So they feed these agents, synthetic agencies. We have nowadays one hundred and fifty synthetic agents. So we fed up with these agencies information, and they also read many data sets. They'll read every news from the, from the overall the available sources, and monthly they give us their inflation expectations estimations, and we always compare these synthetic personas' inflation expectations with actual ones and find kind of the disparities.
At the same time, we have another element of it. We can invest, we can just teach them to pay attention more central bank news, more central bank web pages, and maybe we can direct them how to do this, and we can kind of be flexible in teaching these agencies.
Another one big kind of direction of the work is like we are formulating with the agents digital twin of the economy. What was problem for us, for example, in running DSGE models, QPM models to find the exact kind of the calibrations, exact numbers of the coefficients. Now we have the AI, they can find the behavioral patterns of the agents and put in one big model. So this is kind of new for us as well.
We are learning this digital twin of our economy, representing every agent's behavior with a well-informed, well-analytical background. At the same time, this model also develops itself and generates additional kinds of root behaviors between the agents. This is also one issue. Another issue, another direction is we have the 14 regions and we are also generating the regional AIs, regional agents for each region, representing one region for us. For example, in our decision-making processes, we can ask from that agent regional patterns of the development from the perspective of that region, and these regional agents answer our questions and give us the overall directions of how the region is developing and how our overall decisions, overall economic changes are happening in the regions. This is another one of the platform, and also we are working on applying machine learning techniques to create risk assessments. So we are formulating an alternative scoring model for the entrepreneurs. It's like a startup project. We will then give this product to the market. So from Visava, we kind of give data to this agent, and also we help to collect data series from other agencies, police, settlements, and usage of telephone mobiles, and we anonymize this and give it to this model. Now this model can give entrepreneurs alternative scoring. Why we developed this model? We conducted surveys among banks, and they said the most important factor in their pricing of the overall loan was to investigate the creditworthiness of their claims, and they said they're paying a lot for that. And then we coordinated with our consultants, and now we are providing this instrument to the banks. They can use this instrument to score their clients. This is free nowadays from our awareness, and it can help for the overall banks and overall credit agencies, overall institutions, to first of all estimate the credit for their clients. This is also one direction. Overall, together with this kind of synthetic lab, digital twin of the economy, also the research advisory assistance, we are kind of helping ourselves to more secure data from the sources, at the same time to find partners, and at the same time to educate our young staff to work with this kind of new generational models.
Okay, thank you for the details. So apparently AI comes with all its benefits, but it also includes many risks that we'll continue to discuss here. I just want to jump right here into a related issue, which is governance. At least from a central banking perspective, I think we must lead by example in setting responsible AI standards in the area of finance and ensure that AI systems serve the public interest. But let's try to find some answers to this question, and I'd like to continue with you, Eddie. So AI is developing super fast, right, and regulatory frameworks may struggle to match this pace. And we have our responsibilities for financial stability and regulation. And given this, do you think the time lag between technological advancements and regulatory responses to this could pose a risk to the financial system? And if so, which areas do you consider most vulnerable?
Right. Thank you. The banks in Hong Kong are actually very fast adopters of AI. In fact, we did a survey: about three quarters of our banks are already using or piloting Gen AI use cases. And what we see them doing is that most of them use a phased approach by trying out low-risk areas like the back-office administrative functions instead of jumping right into the customer-facing or more complex areas. And this is an approach that we actually encourage because this will allow them to validate their models, to strengthen their risk management before they actually expand the use, and this will also avoid the cases of uncontrolled or unexplainable AI, thereby affecting their reputation or client relationships. But of course, in the adoption of AI, it also comes with different risks, and we've talked a lot about hallucination, model risk, or correlated trading risk. But the risks that I'm most concerned about are two: one is cyber risk, second is third-party service provider dependence. On cyber risk, it's very clear from all the international discussions that it has actually been aggravated by the appearance of the frontier AI models. It does not change the basic cyber hygiene or cyber defense mechanisms that the banks need, but it does mean that the attacks will come more frequently and in a much bigger scale than before. And what's needed for the banks is that they really need — we actually, not the banks, but we actually need to ensure that the collective defense system of Hong Kong's banking sector is there. It's not just about the bigger banks; it's also about the medium-sized and smaller banks. We want everybody to be on the same boat in terms of raising the cyber defense level, because otherwise, once you attack, the financial system as a whole could be at risk. So we've created this AI cyber risk task force that focuses on sharing not just information about attacks but also practices and experiences. Of course, it's all about detection, defense, patching, and response and recovery. That's the basic framework. But we get banks to share with each other, and we're going to publish very shortly a set of sound practices so that for the more advanced banks, they can share the basic cyber hygiene and the defense practices that they have. And we will also put a catalog of AI use cases for cyber defense. We'll put all these out before the year end so that for the medium-sized and smaller banks, they can actually learn how to use AI to defend against frontier AI models. And I thought that sharing that collective ability, not just domestically in our banking sector but probably cross-border, will be extremely important. The other thing that we've done in terms of strengthening cyber defense is to issue guidance in response to the appearance of the frontier AI model, especially about vulnerability and patching management. So we've got suggestions about how you detect vulnerabilities, how to triage and also prioritize the vulnerabilities when you do the patching, how you can have a much speedier approval for patching for critical functions, and how you should phase out your non-critical patching to avoid unintended operational disruptions, because if you patch too fast, there could be operational failure, and we need to manage that as well. And the other area that we are worried about is the third-party service provider, the concentration, the common reliance of the banks. And here we actually, of course, we've got guidance on the deployment of third-party service providers, the governance, the recovery if anything goes wrong. But in the last two years, we actually did a useful exercise, which is recommended by the IMF, cyber mapping of the whole financial sector. So we joined hands with the other financial regulators to really visualize what kind of connections and common dependence our financial institutions across different financial sectors have, both in terms of their business and also the technology links. And with that visualization, it's actually a very good risk management tool. So if in future an incident like CrowdStrike happens again, we will exactly know the first-round and second-round impact of the links to different institutions. We will actually know how the contagion path will go, and that will actually be extremely useful in thinking about how we should coordinate a more effective response on a sector basis. So I thought these two areas are the main risks that we are concerned about when there's more adoption of AI, and these are the actions that we are taking to address them.
Thank you for your perspective. In fact, the governance issue, as you just implied, is not only within central banks but also nationwide and also globally. So at this point, I would like to turn to Governor Jasmina, since Bosnia and Herzegovina recently signed the Council of Europe Framework Convention on Artificial Intelligence, Human Rights, Democracy, and the Rule of Law, and this is an important legal commitment. But the key question is how these principles will be put into practice. So what impact do you expect from the convention on the financial sector and the use of AI, obviously, in your jurisdiction? And under this convention, how can AI innovation be supported when you try to protect human rights and ensure accountability? Is it a hard job?
So in our case, the signed document is actually the first legally binding document in the area of artificial intelligence according to international standards, and actually that's primarily the reason why it matters to us at this great time. In that sense, we are moving from the area where we were thinking what is possible to another area which is much more important, which says what is permissible. And now we go, I believe we should start in the right direction. At the same time, just signing the document, that's the beginning. The principles will not protect anyone if we don't have strong institutions. So now we need to develop the institutions, to have skilled people inside, to have tools, people equipped with skills and tools, to have institutional capacity to support it. In the case of Bosnia and Herzegovina, what we expect is that those principles should be translated into supervisory expectations throughout the set of documents, and then after that into operational controls inside every single institution that we have. And the constitution of the country, we have on the state level, and then underneath, within the two entities in the country, we have on the entity level institutions. The central bank is doing the coordination when it comes to the financial sector, more particularly the banking part. When I think about the second part of your question, when you were referring to where will the banking sector feel the most the results of this...
I can try. I believe wherever AI has something to do with creditworthiness assessment or AML, that will be the area to deal with this document. If I translate this to the question, I believe that if I think, can a bank refuse a loan based on an algorithm that it cannot explain according to this document that you just signed? The answer is no. If I go further, practically every single question will be no. So we need to be able to understand the background. We cannot come up with decisions only on the basis of algorithms that we cannot understand or we cannot reconstruct some decisions. So based on that, I believe that the financial sector will need to make another step, that's the following part, to strengthen their capacities in order to be able to implement the principles that we have signed. In the case of Bosnia and Herzegovina, now we are not members of the European Union. However, I don't believe that now we are having two parallel roads. I believe this is one bridge to the European Union. So we are trying to do something which is accelerated integration, to do as much homework as possible prior to entering the European Union, and I believe this activity is definitely one of those.
Yeah, it totally makes sense. So inspired by this convention, I would like to bring up a sensitive issue for central banks, because central banks increasingly rely on sensitive data sets, and protecting data is vital for all of us. And AI is not fully trustworthy yet, right? And especially for an area of central banking which has huge responsibilities, we have to strike a good balance here. And let's elaborate on this issue based on a specific question, and for that I'll turn to another bank. Central banks, as I said, hold some of the most confidential information in the economy, from supervisory data to household survey responses. And at the same time, many AI services, as mentioned by Eddie before, depend on a small number of global providers. And how do we make sure AI is used without putting sensitive data at risk at your bank?
Yeah. Right. In this sense, first, we kind of try to rely on our own infrastructure in the beginning stage: data collection, data storage, and kind of the overall beginning stages of the research or beginning stage of usage of this data. When it comes to using the services of global providers, we have strict rules, data governance. So sensitive data, secret data, regulation, and very important data should not be used with AI. And if there is a need to use some kind of bigger data series, less sensitive, it should be anonymized and used with exact bridges. So our own infrastructure stores data information, and then with the bridges, anonymization, generalization, or some kind of finding general patterns, provided to the global providers. This is a certain rule for the central bank as well as we ask from banks also not to disseminate the sensitive data to the global providers. And this is our main focus. And the most important third element of this process is human oversight from the beginning of this process of using these technologies for the overall research, overall monetary policy analysis, and this kind of issues. Human should be the overall last resort of the decision maker, and an economist should review all the processes and make the last decision, and also to decide whether it can be further given to the agents or the providers, whether it should be more classified as strict information to be shared. In this discussion, we have the common understanding with the market participants, also the researchers and other participants of the overall economic activity, that this should go through like this: sensitive, very important information in our own infrastructure with certain rules and very strict confidentiality, anonymization, and then we can provide this information to the global providers just for getting more sense and getting more kind of leverage potential of the global providers, but with certain human oversight.
Thank you so much. Now let's focus on recent developments and the prospects as central banks. We all find the reliability and security of payment ecosystems highly critical, and what is next for the area of payments is the use of AI, which brings us to agentic payments, meaning AI agents can actually plan and initiate a payment under delegated authority. In fact, for years, the idea of AI that can think and act on our behalf has been imagined, even through movies. Some of you might remember Jarvis from Marvel movies. You know, Just A Rather Very Intelligent System. I use this example because it's quite familiar to make the connection. And this imagination is now moving from movies to reality, and Jarvis is coming to life. It's underway for payments as well. And this would create some opportunities and some challenges as well. And Governor Jasmina, my question is for you, in the sense we are increasingly talking about agentic payments these days. So how do you foresee its future in terms of reliability and trust?
I believe when we talk about payments, for centuries we are aware of the fact that payments were resting on a very easy premise. It was practically a human intention that was authenticated properly. And now when we talk about AI and the agents, that premise is challenged significantly. So we have a list of questions, but the first one can be definitely whose intention it is right now. So it's definitely not human anymore. Who is doing the authentication? And maybe the most interesting one, if things go wrong, who is bearing the risk and the consequences? So when we talk about it, I believe we can answer through at least two lenses. The first one, of course, is regulation. So how to set regulation in the future, because it's definitely the case that the framework that we are having today is not ready for the current situation. So the authentication for now is relying on humans. If we talk about the intention, it is relying on humans as well. So practically, if we have the agents, it means that we might have misled the agent, or the agent is misusing or wrongly using the mandate which was determined for the agent. And then if we execute a number of those transactions, it can become a serious issue. So in terms of regulation, that's probably the first thing that should be mentioned. The second thing, and in that sense, the governance, know your customer, know your agent exercise, limited liabilities, that is something that we need to exercise in the future, and actually that's something that we need to improve significantly. The second thing which you mentioned is trust. If we go back in the history, we know that 50 years ago when we had an ATM machine, it was treated like a hole in the wall, and people were a bit suspicious when they were supposed to put the money in that. 20 years ago, when we needed to type something in the cell phone and to give someone the number of the card, it was, wow, should we do that or not? Today we just take the phone, put it close to something, and the money just goes away or is being received, whatever. Which means that practically the receipt is the same. At the beginning we have a fear, then we finally adopt something somehow to the current situation, we adopt it, and then we'll live with that. So I believe in the case of AI agents, the situation will be the same. But everything that is being done repeatedly, everything that is an automated process, everything that is an exercise that is not irreversible can be given to agents. Everything that should be based on judgment, that should be given to humans. That's at least how I view that. In the case of the central bank, I believe this is not a matter of preference. I believe this is our obligation, considering the fact what is the mandate of the central bank, considering trust as the only currency that cannot be devalued. So this is something that we need to prove and to take care of that very, very carefully.
Thank you very much. So here, accountability and liability are two important issues, and AI oversight emerges as an issue to work on as well. Eddie, as the HKMA, you have had some reports on the experience of Gen AI sandboxes, highlighting some advice for financial institutions on key challenges across the AI implementation life cycle. Right. An interesting point there, as I noticed, is the AI versus AI approach. And I just mentioned Jarvis from Marvel movies, and we have seen some stories on the AI versus AI theme there, but it would be great to hear some real reflections from you on this approach. How do you see the greatest potential for AI to provide effective oversight of other AI applications?
Well, it's an approach that's piloted by our banks because when you think about it, as you mentioned, the broader use of AI depends on trust, and trust is built on accuracy, consistency, and explainability of AI. These are important elements, and for accuracy and consistency, when banks are increasingly using AI, especially for high-volume responses, it's getting more difficult for humans to intervene and manually check every work output or every response that the AI makes. So it actually makes sense to think about whether we can deploy AI to monitor, challenge, and control the outcomes of other AI agents or AI systems. And we already have some use cases that are put into our third cohort of the sandbox. And here we are talking about continuous, scalable, and embedded AI versus AI workflow. And I'll just give you one or two examples. For example, there's one use case that got an AI agent to respond to product inquiries by customers, and there are many of them, not only product inquiries but service requests, complaints, etc. And before the message is sent, an independent AI will check the output of this AI agent. And what they check is several things, like whether the approved product information is there, whether the disclosure is there, whether the AI agent makes unrealistic promises or recommendations. It is also actually able to flag up anomalous activities, data leakage, and unauthorized use of tools. And in another case, which is the use of an AI corporate treasury assistant in another bank, not only can the AI checker check the factual correctness of the information, the logic, but this AI checker can actually check the tone and the helpfulness of the AI agent when it delivers the output. So these are what technology can bring in terms of high volume and also operational process, but this also brings risk, obviously. There's the correlation risk because if the two AI models are trained using the same methodology, the logic, then they will come up with the same answers. So if they want an AI checker to be there, we need them to be separately trained with different logic so that they can check on each other. And the other one is diluted accountability because if there's an AI checker in the system, the human will actually think that if the AI checker thinks that it's okay, then it's got to be okay. It's more clever than I am. But is that the case? We need to think about where to pull in the human oversight intervention, and importantly, we want to make sure that for every AI use and every AI checker use, there will be an accountable human owner. So ultimately, the owner has to be human. And the third one is, of course, security risk. There's always the possibility that there could be AI models coming in to hack the system, changing the logic of the AI checker, manipulating the data, or even affecting the calibration of the checking system. So all these are actually new risks that come with the new use of AI checkers. And the way that we think about it is that this is very new, and many of them are still in the sandbox, but if they want to roll out for use, just like the phased approach that I talked about earlier, we advise the financial institutions to take what we call a layered and risk-based approach. So for lower-risk repetitive processes, you can actually use AI checkers to check the AI agents, but there's also got to be hard limits, things like the access rights, financial thresholds, and critical calculations. So there's got to be limits on how much the AI checker can do. And for higher-risk cases, we will actually require the financial institution to have independent challenge and also more meaningful human oversight. And I thought the purpose of the pilot is actually not to replace human oversight by AI oversight. It's really a way to strengthen the AI officer or strengthen governance by having the AI augment the human job in overseeing the AI agents. And in the coming cohort that will come very soon, we are actually receiving more proposals about agentic workflows, whether there could be AI also checking the agentic workflows, which actually comes to another level on top of responding to inquiries or being the research assistant, because agentic workflow means a lot more autonomy in the AI response, and we need an even more robust AI checker to make sure that the agentic workflow is robust. And of course, the whole idea of this sandbox is to share the experience. So it's not like it will be rolled out very soon or by only individual institutions. The whole experience will be shared with the industry, and we will also look into it and see whether we need to provide sufficient guidance to make sure that the AI checker or AI versus AI is doable and can be more widely rolled out. I would like to just sort of add a word about explainability. The AI versus AI is more about accuracy and consistency, but it does not solve the problem of explainability. And it's actually a very difficult area. Without that, it will be very difficult to push out wider use of AI with sufficient trust. We do have a project together with the BIS Innovation Hub to think about whether we can come up with a prototype that can help both banks and also regulators to understand the logic of the different AI models, thereby sort of making it more explainable, and we can also see where the linkages are in case anything happens. And this project is done in collaboration not only with the Innovation Hub but a few central banks, SAMA, and also the FCA. And we do expect that the prototype will come out at the end of the year or early next year. It is a public good that all regulators and all banks will be welcome to use it, and I hope that that will go some way in addressing the unexplainability of the AIs.
Thank you, Eddie. Quite interesting points. So we still have some time for getting questions from the audience. If you have any questions, we'll be happy to get them for our panelists. Questions for global economy or monetary policy or specific applications? Oh, please. Can we have a microphone here? To the front seat.
Thank you for the discussions. Very quick question about the use case of AI at central banks. It's mostly all good, and definitely we hope that it's going to increase our forecasting capacity at the central banks. In your own experience at your institutions, did you so far see concrete, tangible positive outcomes in terms of forecasting capacity? It may be the inflation forecasting or GDP forecast. Thank you.
