 In many countries, the unbanked are partly unbanked because they're at a huge physical distance from the banks. Telecommunications networks which are being used by the AI are covering that distance. That's the first thing to reach them. But the more artificial intelligent aspect of it that matters is about how data is being used because the unbanked don't have that much data about them available for decisions to be made about lending to them, for instance. They have not got a pattern of lending transactions behind them that can be used for decisions to lend to them. However, mobile phones are ubiquitous in even now the poorest countries in the world and these leave behind, recorded with the telecom operators, a lot of data about people, how much they spend, how often they top their phones up, how wide is their social network, whether they make calls only at the end of the month or have a regular pattern. And this is the location of whether they're going regularly to one place for a job, for example. You can devise a profile of people by analysing this data and pooling it with data of others using training data and as a result reduce what we call the asymmetry of information between digital financial service providers and people because the providers don't know that much about the people, they can't take a risk on them. However, if they have the data and are able to profile them well, then they can actually take a risk on them. And so this is used increasingly to bring unbanked people into digital financial services market, which enables them to begin to have a credit risk background. In addition to that, artificial intelligence is being used with, for example, the location data to predict when people are carrying out transactions in unexpected locations, which can be a signal for potential fraud and so they're able to be used to identify potential fraudulent transactions and again reduce the cost of transactions to them. That, just by reducing the cost improves the prospects of extending these services to the population of unbanked. The third area really is also just using artificial intelligence through detecting people's behaviour patterns to figure out what sorts of services they might actually need, they might have some demand for so that these providers can better target the service that they want to offer to the customer. So we're seeing a lot of examples of this across Africa and Asia. It's rolling out increasingly widely worldwide. With artificial intelligence, the challenge about discrimination or bias, there is a risk always that garbage in, garbage out. The algorithms can only process the data and come up with as good conclusions as the data that comes in. And if the data that comes into the system reflects existing social bias in the country, then the outputs may well reflect that as well. So for instance, if you have population groups, perhaps religious, perhaps ethnic, which have historically had much lower income levels or lower use of telecom services, for instance, you may find that they are less likely to get a green light for credit decisions simply because the algorithms have not been able to reach the same positive conclusions about them as they might for other population groups. And this can reinforce the disadvantages that these population groups face. There are other policy issues such as if you're an individual and you're not happy with a decision that has been made by an automated decision-making system, not to grant you a loan, for example. How do you, what sort of recourse do you have? How do you find a human that you can actually appeal to as opposed to just being stuck with this electronic decision about you? So bringing a human into the loop in a recourse system can be important. But lastly, there's a whole question around explainability of automated decision-making that use algorithms because the more precise they are, the more layers to which machine learning has gone and down in the depths of analyzing behavior patterns and data, the harder it actually is to explain how the decisions were reached. And so there is sort of a balance, a trade-off between making the technology more explainable on the one hand and making it more precise on the other. So policymakers need to grapple with all of these issues as we move forward in this area. Well, one is simply to contribute to the overall efforts to promote responsible use of AI through development of standards. I think, for example, all the work on the digital credit standard that was done by the Smart Campaign, I think about the Singapore Monetary Authority's work on the principles to promote fairness ethics, accountability, transparency. In the financial sector, I think the ITU can contribute through these sorts of standards. But to get more technical, which is where I think the ITU really adds its value, one of the challenges that we face is not even in the detailed operation of the AI, but it's really in the access to the data itself. Too often, it is difficult to get access to data that can be used well for AI. The data is held in silos in different companies, each of which may hold it in different forms. It's held in different sectors under different regulations, and very often companies are just not ready to yield that data up or provide access to it for this sort of use. They face a regulatory risk if they do provide the access, and it may well be, for example, if it's a telecom operator, the revenue they stand to make from this is actually not that great compared to their core business. So what sort of incentives can we set up to enable, to encourage telecom operators in particular to make that data available? And one area I think the ITU can really add value would be around working on standards that would do perhaps three things. One would be to have a sort of a convention around what data really is the most valuable to use, because maybe 90% is not that valuable, 10% is, and so let's focus really on the valuable stuff, and have some conventions around who can get access, how they can get access to that data, what they're allowed to do with it, and what must be done with it after they have used it. Having some conventions on that I think could be quite helpful. It might also reduce the regulatory risk that the operators face in providing access to it and make them ready to do so. Another area where standards could be useful would be in providing for unique identifiers that would be used across different firms in the economy to link to the data, which would then allow that to be processed at scale, because it's really the processing at scale that makes the big difference. And then thirdly, standards could have some sort of consumer redress mechanism, enabling consumers to question the data that's been used to make a decision about them so that they can correct it if it's wrong and have another go at applying for, in this case, a load. So I think the ITU could convene telecom operators, policymakers, financial sector players around work on these kinds of standards down in the nitty-gritty, which I think could socialize, could create value for all of society than using this data by providing access to it.