Special Edition: Tracking AI diffusion in Q1 2026

0:00A lot of times people think about software developers or even computer scientists as people that need to know a programming language. And I think that that is the wrong approach. Like a software developer or someone that knows how to code is someone that can actually communicate computers to automate process. Whether that's in English or Python, it doesn't matter. So I think it's the wrong approach to think. It's like, oh, we will need less software developers because now you can code in English. No, no, you will have more software developers. it's going to become easier to code. It's already easier to code.

0:32That's Juan Levista-Ferez, Chief Data Scientist and Director of Microsoft's AI for Good Lab. Juan's team just dropped the latest Global AI Diffusion Report, our real-time look at how AI is spreading and scaling around the world. We talk about the recent surge of adoption in Asia, how AI is changing software development, and the role trust plays in adoption. The report shows real momentum, but also some troubling divides. AI Diffusion with Juan La Vista-Ferez, up next on Tools and Weapons. Juan, welcome back. I think it's an exciting time to sit down and talk. The AI for Good Lab has now published two more reports on AI Diffusion.

1:19You're starting to put these out every quarter. Yes. The Global Report shows accelerating growth for AI, shows a surge in Asia. It shows how AI for coding is impacting lots of different things. We'll get into all of that. And your new report, for the first time, a county-by-county usage report for the entire United States, shows that we have our own AI divide in the country between urban and rural counties. Before we tease that apart, can you say a little bit about how you do these calculations across the country, around the world? What gives the AI for Good Lab at Microsoft an ability that perhaps is unique to prepare these kinds of estimates?

2:07Yeah, that's a great question. Thank you for the invitation here. Within Microsoft, we have billions of devices out there. and we can collect in a very highly privacy-preserving way, very anonymous. We have data on how people are using these services. And we can report, based on the most common AI models, by country, by county, how many people are using. And then, of course, these are not the only devices out there, so we need to control for the Internet penetration on that county. We need to control for mobile versus PC. Once we adjust on all this information, we can report. And the great thing about this is that it allows us to do an apples-to-apples comparison with countries and also a trend over time.

2:56So the headline is at the end of the first quarter, the end of March, on a global basis, 17.8 % of the world's people or world's working age population was using AI. What's the definition of working age population? This is, we use the definition of the OECD that is between 15 and 65 years old. 15 and 65. Yes, yes. This is a good proxy on people that are using this. Majority of the people that we see using these PCs are within that range. And that was a point and a half over December. Yes. If it remains that way, we'd see 6 % growth over the course of this year. I think in some ways what was even more interesting is there's now 26 countries that have topped 30%.

3:44So we're seeing this steadily grow in a great many countries. And we always talk about who's the leader in the world. You want to share that? Yeah, so the leader remains the UAE. It's actually the first country that tops the 70 % of the working age population using. And what is amazing about the UAE is that by then we would have expected that there was some stopping growth, given that the majority of people are using it. And UAE remains not only the top one, but one of the ones that has the biggest growth for the last six months. So it has been incredible. One of the things I found interesting is how many people, especially in governments and in the tech sector, and the press are really following this report.

4:33Since the last report came, I've been in nine countries, and people talk about their number. Or late last year when I was in the UAE, I felt like everybody knew they were number one. And what's really interesting is they often know exactly what their percentage is. But as you say, in December, the UAE was at 64.0. Now it is 70.1. That is accelerating growth in so many ways. Yeah, certainly the UAE is doing something that is outstanding from that perspective, from the growth that we see. The U.S. is finally moving up the ladder. It went up three points. I think it's now 21st in the world instead of 24th.

5:17It's sort of remarkable in some ways that the United States has not yet broken into the top 20. but you're now seeing more movement. What do you make of that? Yeah, it's one of the countries that actually in this quarter has grown the highest, at least number of positions. We see a gap, and that's one of the reasons why we are doing a deep dive in the U.S. Within the U.S., we see also big disparities. You have places that are doing really well, places that are doing not as good. And I think we have that disparity between the rural population and the urban population that in the U.S. is big. I want to come to those gaps, both the north-south and the urban-rural.

6:01But before we do, let's talk about a place that seems to be closing the gap. One of the things highlighted by the AI for Good Lab in this report is the growth in Asia. Yes. And you've really dissected that, but it's fascinating. I was in Japan and Thailand, two of the three countries in Asia that, together with South Korea, have been growing the most this past quarter. And you do a deep dive on Japan to try to understand what is driving that growth. What is the single biggest factor? We think that the biggest factor, similar to what we saw in South Korea, is language. in the sense that a lot of these models that were not doing as well in Japanese a year ago, two years ago, now we are seeing that that gap, the performance in these models in English versus in Japanese, for example, they're almost on par, which means that now these models are much, it's much easier for people to use these models and feel that they're useful.

7:02So I think that that's one of the main reasons why people start using these models and now they can solve problems that before they couldn't solve, and they become more as a user versus before. What is contributing to the improvement? Well, I think it's the improvements. I think that a lot of these companies are investing in these markets, are improving the models similar to what we saw in South Korea. Clearly, they realize there's a gap, and I think companies like Omni.ai, like Anthropic, are improving their performance in those languages, and that is clear from the results in some of the tests that we run.

7:41I know you do other work in this space. I mean, the AI for Good Lab and other parts of Microsoft have been working on linguistic capabilities really around the world. What do you see as the biggest barrier? Is it just a shortage of data in a local language or is it something else? It starts with a shortage of data. If you are in a country like the U.S. where English is an 80-language, which is like 50 % of the content on the web is in English, makes it relatively easy to train a very good model. And similar happens even with French or German. Once you pass certain languages, that's no longer true.

8:19And we still have, we live in a world where you have countries where they don't have any access to these language models because these models are not trained on those languages. That was not the case in South Korea and Japan, but they are not low-resource languages because they still have a good portion of the web, but it's not on par with what you see in English or German, for example, or Spanish. So now there is an investment to make sure that these models can do well in those languages, and that is happening, and we see that in the results. And you talk in the report about data, improving performance, and then as performance improves, demand increases, and you start seeing these countries really scaling up deployment.

9:06Can you say a little bit more about that? Yeah. Clearly, what we see is a very good correlation between the performance in these languages from these language models and the people using these models. And I think we also saw it in the early days, like even GPT-35, when GPT-35 was a very good model, but it wasn't what we see in GPT-4 or GPT-5. Yes, you could use it for editing things, but you couldn't solve many of the problems that you could solve today. I think that as soon as people see the power of these models and they are doing, they can start using it for multiple other purposes. And that brings more adoption.

9:48Well, I saw one of these interesting examples of scaled deployment when we were in Thailand in April. And as you show in the report, Thailand is one of the three countries in Asia where AI grew the most in the first quarter. Thailand has recently applied to join the OECD. And when you do that kind of application to join, you have to show how you're going to conform your domestic laws to OECD standards. That's typically a process that takes three to five years just to prepare the application if you are not a native English-speaking country, and Thailand's not. So what they had to do was take 70 ,000 laws in Thai and translate them to English and then compare them to about 270 different OECD standards.

10:39In this case, it didn't take them three years or five years. It took them three months. with a team of five people, five lawyers who work for the government. And the difference was AI. They used AI to translate 70 ,000 laws from Thai to English and then used AI to do the comparisons, compare these laws to these standards, where are their gaps, where are their examples of what other countries have taken steps to then close those gaps to meet the OECD's requirements. I thought it was fascinating because in so many ways, it not just accelerated a process for the benefit of everyone in the country, but it eliminated a lot of what we, I think, rightly think of as drudgery.

11:25All of that translation, all of that laborious looking at one thing, looking at another and comparing it. These are things that AI is very good at. Especially these generative models that before, like AI is not necessarily new, but it was not doing well on dealing with text. That's the majority of the human knowledges. Thanks to these large numbers, now you can do that. And solving something like that, before it would have been impossible without a lot of effort from humans. Now we can do using these models. So Asia's growth is good news. Yes. The U.S. ratcheting up three steps. That's good news for the United States.

12:07The UAE's leadership, great news for the UAE. But there's some big gaps around the world. Let's talk first about the north-south divide. What did the first quarter bring in results on that score? When we look at the report, what we see is that the global north continues to grow faster than the global south. So we saw the global north went up in this quarter of 2.8 points. The global south was less than half of that. It's 1.3 points, which means that the gap between the north and the south continues to grow. And once you do a deep dive, and we actually included that in the report, it's like some of the drivers, is that access to the Internet, access to electricity, access to skills, we see that not only this will likely continue to accelerate, there is a point where some of these countries will hit a wall.

12:58I don't think it's happening yet, but I don't think some of these countries are not that far from that wall, which means that the difference between the global south and the global north will continue to increase, and that's something that, yeah, it's unfortunate. And the report does an interesting job, I think, of breaking it down into the different layers of technology that matters. And as you show here, when it just starts with access to electricity, the global north is more than 98%. The global south is at 88.9. So you get that gap. Then you look at access to the Internet. The global north is at 90.

13:37The global south, 65.7. So that's a 25-point gap. You look at access to digital skills. The global north is at 70. The global south is at 48.2. So there's another sort of 22-point gap. I think it shows that closing this AI gap actually will require that we close lots of gaps. The electricity gap, the Internet access gap, the skilling gap. It just goes to show how much effort it's going to take. And the bigger problem there is that once you have the infrastructure in place, adoption is not difficult. And this is why explains why less than three years, a bit more than three years that this started, we see a significant portion of the world already using it.

14:27Because as long as your LLM can speak your language, it's relatively easy. Having access to the internet, having access to digital skills, having access to electricity, the investment that these countries will need to do is significantly higher than having that one. That part is going to be the difficult part of adoption. Yeah. That gap persists. It's even getting wider. Your U.S. report actually shows, interestingly enough, a gap of similar magnitude within the United States between urban counties and rural counties. The urban counties are how much more than the rural counties? Yeah. So we see similar to what we see between the global north and the global south.

15:08interesting enough, almost in a very similar ratio, the urban areas in the U.S. have around half of the AI diffusion that you see in metropolitan areas. Rural is half of urban. Rural is half of urban. Yeah. And when we look at some, a lot of countries in the U.S., they have lower AI diffusion than a lot of countries in sub-Saharan Africa. So, and this is not because of access for electricity. Like, they actually have electricity. They have access to the Internet. there is that gap, and this is something that I think is worth studying. Right now, we can observe the data. I don't think we know why this is happening, but we clearly see that divide in the U.S.

15:50between rural and urban. We have seen that in other countries. The rural and urban divide is not something that is unique to the U.S., but clearly when you look at the map in the U.S., it's clear. The one thing that we do know, even though there's more that we need to learn, is that we can look at a similar map of counties in the United States. And trust in AI tends to correlate with usage of AI. Trust is higher in urban areas. It is lower in rural areas. We've long been saying as a company, both internally and externally, that people will only use technology that they trust. So is it fair to say that's one hypothesis we're going to have to go test now and see if we can learn some more about what it is about trust in AI that may be part of this story?

16:41Yeah, we see, like you said, the correlation is clear. Like the rural areas in the U.S. have much less trust, significantly less trust. There's almost a very good correlation between the inverse correlation between trust and AI adoption. That's a very good hypothesis. The thing that is noteworthy about, in part, in my view, is that the uses of AI in rural counties are, I think, so compelling. I mean, you just take the health care challenge. You and I have been looking at that recently, and there's almost 2 ,000 rural counties in the United States, and yet 45 % of them have five or fewer doctors.

17:27There's 198 of them that have no doctors. And we already see AI, including AI services from Microsoft, from Nuance, being used by doctors to be much more productive. The capture and the essence of a conversation between a doctor and patient, to free the doctor up to see more patients. The more acute the doctor shortage, I think the more compelling the need is to put AI to work to help doctors see more patients. But obviously, that hasn't necessarily translated into more trust in AI at this point. Yeah. Healthcare in general, but particularly in areas where they don't have any other solution, I think that is clearly a game changer.

18:13similar happens to areas like agriculture, like using AI to help on the agriculture. For example, use the reliance on fertilizers, make it more efficient. We see a lot, many more use cases that are extremely compelling for rural America. But yeah, we still don't see the adoption of this technology. Well, yeah, and the other one that you and I have been talking about that the AI for Good Lab has been working on is fighting wildfires. I mean, the ability of these AI-enhanced cameras that you've been directly involved in in California to be able to detect and identify smoke patterns that quickly show wildfires.

18:54And as you have been showing me, last year in the United States, wildfires destroyed an amount of acreage equal to the state of Massachusetts in size. So, you know, this is a great example of where AI can put out fires, save homes, save lives. It feels like we have an opportunity for a broader conversation about how AI can be put to work in ways that will genuinely serve the needs of rural communities in this country. Yeah, wildfires is a great example on the conversations that we need to have. And I think we just started on that. Clearly, the case of California, we would love to actually bring that to other states, too.

19:39I think that California is kind of ground zero for wildfires. But wildfires is being affecting, like, a significant amount of other states, too. And other countries. Other countries, yeah. It's a global issue. As we continue to mature this report, really develop it, we're not only now doing quarterly reports, global as well as county by county in the United States, hopefully county by county and some other countries too. But you're now starting to get to the point where you're also analyzing different sectors. You started with software coding. Why did you choose that as the first? Well, software coding is clearly one of the areas that AI is already seeing a huge improvement.

20:20I see it in my team, where everybody in my team now is using these tools to help them do software development. we move, I think that what the world observed in November 2022 with ChatGPT, we have a very similar moment that happened in December 2025, where suddenly these models, either the entropic models or the OpenAI models, through technology like GitHub Copilot, it allowed now the software developers to start coding in their own language, whether that's English, Spanish, or Mandarin. and I'm using it myself and the improvements in productivity that we see is huge. So we wanted to look at that data as like, hey, clearly we are seeing an amazing moment for software development.

21:05Is that being translated into code? And we, through the GitHub data that is an amazing data source that we have, we already observing that, correct? So we see huge improvements in the amount of repositories. These are projects in GitHub for you to be aware. Like in the last six months, we see more repositories that were created in GitHub than in the almost first 10 years of GitHub. So clearly we are seeing a huge increase in productivity in the software developer side. So people I'm sure are familiar with or have heard about Anthropics model, Claude, and people using it for coding OpenAI's model.

21:44But part of what you point to is just the evolution of GitHub. It's no longer just a place where people store code and make it available to a team in a repository or repo. Even GitHub Copilot isn't just a tool that people are using for AI to write code. Talk a little bit about the evolution of our own GitHub service and what that means for this. Now, for the first time, and I think that that moment happened to me last December using GitHub Copilot, was that I no longer needed to actually write code in Python or C or C Sharp. I was starting to use English. And I think that that completely changes the dynamics of coding.

22:30And I think that, especially for the people that maybe their job was not coding, I think it's changing discipline. Instead of writing a spec, you're building a whole prototype just by coding it in English. This is, I think, is going to change the dynamics of the projects from idea to bringing ideas to life. I think we're just starting to see that impact in society, and I think that impact is going to be huge. You look at all of the tools that are coming together on a service like GitHub. what does that tell you about what the future of a software developer job is starting to look like? I would argue we still even need more software developers.

23:21I would say the majority of the people will become software developers, correct? In the sense that no matter what your job, whether you're a lawyer, an architect, or an accountant, your interaction with the computer will be through coding. You're not going to be coding in Python. You're going to be coding in English. But that notion will still be there. The ideas of building software, the idea of having ideas, and I think that's going to become even better, like more impactful. Even after 30, 40 years that people have been trying to make sure that society and the kids are coding, around 0.5 % of the population know how to code.

23:58So it's a very niche. I don't believe that's going to – sorry, I believe that's going to change dramatically. There is obviously a big debate with wide-ranging views across the population, across different experts, about the impact that AI will have on jobs. And yet right now, we're still seeing growth in software development jobs in the United States. In fact, the report points out that the U.S. Department of Commerce last year reported record growth, 8.5 percent increase, 2.2 million people in the United States employed as software developers, even while the growth of AI for coding was exploding.

24:44You talk about the different economic factors that you see at work, at least at this point in the development of AI. You have a strong background in economics as well as a strong background in code and AI. How do you analyze the economics so far? I have economists on the team and we've been discussing this because the first perception that a lot of people are thinking, well, now everybody can become a software developer. The productivity increases significantly. We will need less software developers. That's the first, I think, perception that a lot of people have. And that is true if you have a market that is fixed.

25:26For example, that's what happens in agriculture, correct? So you have kind of a fixed amount of land. You increase productivity back in the early 20th century. You increase productivity. The amount of jobs that went into farm actually start decreasing. But in the case of software development, the sky is the limit in the sense that you can grow. And when we look at the last 30, 40 years, this is not the first time that we see a big increase in productivity in software developers. Like we used to code in assembly. We used to code in Fortran and Cobalt. Every time that we saw an improvement in productivity, we also saw more jobs in the area.

26:07because the fact that you can build more stuff makes it more compelling for people to use these tools. And so as long as we think that's elastic, this should actually translate into more jobs. And that's what at least the data is showing. We don't know what's going to happen in the next five years. Right now we are seeing growth. And at least according to the economy, as long as this part is elastic, the fact that more improvements in productivity might drive more jobs that happened in the past. And it will be interesting to see. Obviously, it's one of these classic questions. Time will tell. I think it's interesting when you look back at history and, you know, in terms that almost anybody, I think, can appreciate the invention of the washing machine massively reduced the amount of time to wash clothes.

26:56Before the washing machine, it would take about six hours to, in effect, clean what we now think of as a load of laundry. And then that eventually fell to about 30 minutes. And most of it was time where people could put the laundry in the washing machine and walk away and do something else. But it's the exact same point you're making. The first thing it did was improve demand for clean clothes. Yes. People's whole expectation was that they would have clothes that would be washed or cleaned more often, whereas there was an explosion, about a tripling of the washing of clothes. As software is cheaper to produce, there is then an opportunity to use more software.

27:42I think one of the things that will be interesting to see is how this translates not only into the number of people employed as software developers, but first, as you point out, the nature of the work. And we're seeing that across our industry, but also the types of companies that people work in. And, you know, we're seeing that in some ways software development jobs are migrating to some degree out of the largest companies to more companies and not just tech companies. And this has been going on for decades, but even more perhaps now where, as we have long said, every company is a software company.

28:24company I think right now might be experiencing an increase in demand and an increased ability to hire people because AIs reduced the barrier to entry who are creating the next generation of software. One decision that I think is interesting, Brad, is that a lot of times people think about software developers or even computer scientists as people that need to know a programming language. And I think that is the wrong approach. which is like a software developer or someone that knows how to code is someone that can actually communicate computers to automate process. Whether that's in English or Python, it doesn't matter.

29:00So I think it's the wrong approach to think. It's like, oh, we will need less software developers because now you can code in English. No, no, you will have more software developers. It's going to become easier to code. It's already easier to code. I think it's a really interesting point that you make because it's not just people who are working full-time to develop software. Maybe a great many of us, many white-collar workers of all kinds of backgrounds and professions will be, as you say, communicating with a computer and, in effect, creating something that is manifested in code that we put to work, especially in an agentic AI, to help us do our jobs.