The following are the outputs of the captioning taken during an IGF intervention. Although it is largely accurate, in some cases it may be incomplete or inaccurate due to inaudible passages or transcription errors. It is posted as an aid, but should not be treated as an authoritative record.
***
>> Ladies and gentlemen, the programme will start shortly. Please find your seats.
>> All right. Welcome, everyone. Thank you so much for coming. I think we are going to have a wonderful, wonderful session here today. And at first, I would like to give the floor to our wonderful people at Google. So, please.
>> Perfect. Thank you so much. So, while I'm not working for Google, it's the I just Secretariat here. You may know me from the various exchanges that we've had, basically to give you some background information also on how we've been collaborating, actually, also with Google and also with partners of Google. The idea came from last year's Internet Governance Forum that we had also, actually starting with the Central Asian IGF and then at the global IGF, where we really saw the need of capacity building and also requests from members of Parliament to have some capacity‑building workshops. And basically, the idea is that this workshop is really hands on. It will be interactive. It will be very concrete around the topic of artificial intelligence and cloud policy. So, this is also something that we're doing together with the Inter‑Parliamentary Union, and this is why also Andy will say a few words. But again, thank you so much to Google also for being part of the IGF here and also providing this very hands‑on training for the next two sessions. Thank you so much.
>> ANDY RICHARDSON: Thank you, Celine, and good morning. Thank you for being here today. On behalf of the IPU, we are very happy to be associated with this workshop. Two things. One is that we hear a lot of times from Parliaments that this field of AI is very much new and emerging and that it's very important to get information about exactly what this is so that Parliamentarians can think about how to approach the subject. And second, we've also heard through the Parliamentary Track about the importance of multistakeholder collaboration, the ongoing dialogue between all of the different actors, including parliaments, the technical community, the private sector, civil society, and all of the components of the Internet ecosystem. I'm glad today that we're able to give a practical character to that multistakeholder collaboration with this workshop. Thank you. And I give the floor to Olga from Google.
>> Olga: Thank you so much, Celine and Andy for this opportunity to partner on this very practical and I hope very useful session in a workshop style around AI and cloud policies. I just wanted to say a few words on how we at Google think about AI for Parliamentarians and public officials. We do believe that AI is a technology, is a transformative, once‑in‑a‑generation opportunity for all of us. It will definitely change the way ‑‑ and it's already changing the way we work, the way we live, but also the way we govern and the way we legislate.
I also ‑‑ I'm cognizant that there are many representatives of parliaments from emerging markets in this room, and definitely, what we see is that people in emerging markets, your constituencies, they are really embracing AI, and we have our research that we published a couple of months ago that shows that users in emerging markets are actually much more excited in adopting AI. And definitely, this is where capacity building for parliamentarians ‑‑ those people who represent their constituencies ‑‑ really top of mind and urgently needed. So, we're very thankful for this opportunity to provide this workshop and hands ‑‑ share hands‑on knowledge and how actually AI can be used by parliamentarians in their day‑to‑day life.
Just a couple of, you know, notes to highlight how, actually, AI can be helpful for parliamentarians. Definitely, AI can help drive efficiency and can help you prepare for parliamentarian debates, do research when you're working on legislation, but also free up your time by automating routine tasks.
We also see great examples of how AI actually helps or changes the way parliamentarians can work with their constituencies. I just wanted to share one recent example. It comes from the U.S., but it can be applicable everywhere. Where a mayor of a small council decided to run an online town hall as he was working on a strategy for the council up to 2050. And with the help of Google AI, 1 million ideas out of council with a population of 178,000 people was collected. People were really, really eager to share their ideas of how they see the future of their country. And then the challenge was how to actually make sense of this 1 million ideas, because if you put these ideas on secret notes, this would be enough to fill up two football fields. And so, with the help of AI, in one day, they managed to actually make sense of this feedback, structure it in an actionable way and then hand over this as an input to the committee that was working, actually, on a strategy, while still keeping this authentic voice of the community. This is just one of the examples how AI can be helpful in strengthening this connection with your constituencies.
I don't want to take too much of our time from the workshop but really hope that this will be helpful. And we as a company, we're committed to developing, also deploying AI technology, boldly, responsibly, and together, with multiple stakeholders, in a multistakeholder approach, which is the model of IGF. And thank you so much, again, for this opportunity. With this, I will hand it over to our trainer for today. Thank you.
>> ALEKAI: Thank you so much. All right. My name is Aleksai, and I started my journey in AI ten years ago, studying machine learning. And now, for the past few years, I've helped companies and governments to take advantage of AI and help them understand what are the capabilities of LLMs and generative AI. And this is something we're going to deep dive today.
And even though our focus today is somewhat on the government side, I think this is going to be really helpful for everyone, because we are going to really take hands‑on approach so that we are going to take a look at the different tools we have available, and then I'm also going to let you guys get hands on with a few of the tools.
But what we are really going to do today... So, there's a few chapters we're going to go through. So, first, we're going to take a look at the basis of AI data and the cloud. And I think it's a wonderful opportunity, because as you all know, the media, it's so full of this AI hype, and it's really difficult to keep up, and sometimes, we can be quite overwhelmed with all the information. So, this is really a great opportunity to really understand the basics of AI. So, that's where we're going to start. So, we look at the data, we look at the cloud and AI and how these are related.
After that, we go more hands on. We're going to look at some examples of how we can use these tools; what's the difference between predictive and generative AI, and we're really trying to get a good understanding of how we can take advantage of AI.
Then, we are going to go even more hands on, and we are going to look at some of the tools, such as Google Gemini and Notebook LM, and then we're going to try them out and see what they are capable of.
And finally, we're going to discuss some of the challenges and strategies related to AI. And we're going to have some time in the end so that we can discuss about AI, about principles, about regulation, and then we also have a Q&A. So, whatever you have in mind in terms of AI and generative AI. This is a great place to ask questions. So, that's roughly the timetable for today. So, first, we go through these phases, and then after that, we are going to have this discussion, and then this Q&A session.
So, let's dive into today's topic, and let's really look at how the AI, the Cloud, and data, how they are related. Let's first define some terminology. So, what is data? So, data is basically just digital information. But why is it so important when it comes to AI and the Cloud? And the reason is that the data is what we use to train the AI models. So, without the data, we cannot train the models. So, data, it's crucially important, in order for us to build these AI systems.
All right. How about the cloud? Or actually, let's look at a few examples before we move into Cloud, let's take a few examples of different types of data. So, I think this is actually a really nice way to look at the different types we have, and this is, especially from the government perspective, something that is really important.
So, first of all, we have personal data, which is something like names and addresses and ID numbers and so on. Then, another class of data, a subset of personal data, is the sensitive data. So, especially when it comes to governments, it is really important to think about these things because this sensitive health data, for example, is something, we really want to secure that. And this is something that these cloud providers can help us with. Then we have the non‑personal data. So, data that we cannot connect back to any individual. And here, one noteworthy thing is that you really want to make sure that the link between the persons and today that they are so interrelated that it's not possible to link the data back to individuals, something to really keep in mind when designing these kinds of systems that take advantage, for example, non‑personal data.
And then, of course, we have the public data. So, data that is freely available. And I think the public data is something that holds massive opportunities, because this is ‑‑ if we give people access and companies access to a lot of high‑quality public data, the companies, they would take advantage of that and build amazing applications and AI systems on top of the data.
Now, I was eagerly jumping to Cloud, so let's do it now. So, what is the goal of Cloud, and what is Cloud? So, we can think of the Cloud as just a huge amount of remote computers. With cloud, we can take advantage of the most capable computers in the world. So, it's quite amazing to think about that. Me as an individual, I have the similar capabilities compared to huge corporations because of the cloud and access to computers.
So, Cloud is this network of remote computers, and we can use it to process the data.
All right, then we come to AI. So, what is AI, and how should we think about AI and how these things are related? First of all, the definition is AI is mimicking problem solving. So, with AI, we can solve problems that we usually think that would require human kind of intelligence. So, I think that's quite nice definition and way to think about AI.
Okay, so, how are all these related? So, we can think about the data ‑‑ it's necessary, like the building block. So, without the data ‑‑ we need the data, in order to train the AI. But then for us, in order to train the AI, we also need some place where to hold the data, so this is where the Cloud comes in. So, the Cloud is like the infrastructure layer we can use to hold the data securely. And then we have the engine, the AI. We can think of the AI as an engine on top of the Cloud. So, these are all really crucial elements and these are the main building blocks of AI, and maybe something we often don't think about enough. So, I think it's really good to remind ourselves that when we look at the applications of AI, like chatbots, we often forget the building blocks. It is important to remind us that, in order for us to have these amazing chat applications, for example, we need the building blocks to be in place.
How Cloud is enabling AI is that, with Cloud, as I mentioned this, the amazing thing about the Cloud is that it enables me, as an individual, or some small government in some distant country to have amazing capabilities, similar capabilities to the world's largest corporations. And one of the nice, like, features of the Cloud is that they are very efficient. And what we can do with the Cloud is that they scale. So, when we don't ‑‑ let's say we have ‑‑ we want to take advantage of some supercomputer, we can use that for one day, but for the next day, we can use it and there will be no cost associated with it. So, I think that's the amazing thing in terms of cloud computing that we can really pay only for what we need.
Another great benefit is if we were to have this ‑‑ let's say we do some ‑‑ we build systems related to medical, for example, or we have some kind of systems that are really sensitive in terms of the data they have. So, if we were to build our own servers, we would need a lot of expertise in terms of security, for example. But with Cloud, we get that security out of the shell. So, also a really nice feature.
So, accessibility, efficiency. You only pay what you need. I think Cloud is something maybe we should be more enthusiastic about. Everybody's talking about AI, but I think there's still a lot we can accomplish with the Cloud.
So, let's switch gears a little bit and let's talk about AI. So, as I mentioned, the whole media thing around AI is ‑‑ I think it has caused a lot of false beliefs, so let's do some demystifying.
Okay, myth number 1: AI will replace human jobs. And what I think the reality is that AI, what it's really doing is it's freeing up the capacity of humans to do more interesting stuff. So, with AI ‑‑ when AI is automating the boring stuff, we can focus on more creative problem solving. So, I think that's quite wonderful, and I think the future of human work is looking really, really bright.
Okay, when we look at movies related to AI, I think we often get this feeling that, okay, these AI systems, they are human‑like, and in the movies, like, they think and feel like humans, but the reality is that AI systems, they are just statistical models, and they don't have any feelings. And they don't have emotions. Often, AI systems, they are not very capable of detecting our emotions. So, that's still what we need to do.
The third one is that AI systems are always right. And that's totally not true. So, AI systems can make mistakes, and now that we looked into the data, there's oftentimes the mistakes of AI actually come back to data. So, our AI models are only as good as our data. So, if we have, like some data that is not representing what we are trying to achieve, or there are some biases in the data, they will end up in our RA systems. So, our systems are only as good as the data.
And then also, when it comes to generative AI, we also have the hallucinations. And we are going to talk about hallucinations more later on, but what they mean is that AI can sometimes produce plausible, but incorrect, outcomes. So, you ask something to AI chat, and it sounds about right, but actually, when you dig deeper, you figure out, okay, this is not correct at all.
I think what we should do here, and what I really encourage you all to do is to discuss about AI principles. And this is what we are seeing happening now across the globe. For example, in the European Union, in a lot of private‑sector companies, there's a lot of discussion going on in terms of AI principles. And I think that is great.
For example, I have an example here. These are the AI principles of Google. So, let's take a quick look at those. So, here are the AI principles of Google. So, bold innovation, responsible development and deployment and collaborative progress. And I think I used Google's AI tools daily for a few years, and I really think Google has done a nice job at applying those principles, so I think that's something that is also really important that we just not only just write the principles down, but we really live up to those principles.
All right. Let's switch gears and move on to the second phase of this presentation. So, streamlining your work with AI. So, what kind of things we could automate with AI? Here's a few examples. So, how AI and Cloud could help you in your role is that they can help you in routine task automation, such as filling forms, writing emails, data analysis, especially the latest reasoning models from AI Labs, they are amazing at doing data analysis.
Then, we can use this as personal assistance; we can do predictive analytics with more traditional machine learning; we can do policy formulation and evaluation; and then, we can also use AI for public engagement. And we have a few examples of those, so let's take a closer look.
All right. So, automation of routine tasks is something the current generative AI models are really, really good at. And this is ‑‑ an example of this, there are many tax administration assistants. So, what they have done is they use AI to fill the forms, so repetitive, boring human job is automated with generative AI.
Then, of course, we have, well, you can write emails, you can create documents, do research. There's a lot of different kinds of routine tasks you could automate.
Then we have the data analysis and insights. It's been something that's been possible for quite some time. So, with machine learning, we've been able to do data analysis and gather insights from different kinds of transactional data for decades. But the recent advancements, they've made it more accessible. So, now it's quite feasible to any small government, or even a small company, to take advantage of the data analysis capabilities of these models.
An example of this is that these models are actually, they are really good at analysing images. For example, Ray scan images. AI is actually, in most cases, it's as good as humans and sometimes even better than humans at doing this kind of image analysis. So, amazing progress there. And also, the nice feature of this is that these are actually ‑‑ nowadays, you don't have to necessarily train your own models, but you can just use the Cloud and pretrained models, so these are quite accessible for everyone.
All right, then we have the personal assistant. I think a lot of you are quite familiar with. So, you can chat with AI and ask questions. And a really nice example of that is Estonia's Burokratt Chat, which is basically a Q&A platform for everything government related. Whether you need a new ID or you need to file some government‑related information, you can ask for the chat, "Okay, how do I do that?" And the chat, it's always available. So, nice, nice example of personal assistant.
Then, we have the predictive analytics. So, now there's a lot of interest toward generative AI, but it's good to remember that we have also this predictive analytics side, and there's a lot we can do with machine learning and predictive analytics. Such as how to better plant crops. And we've got some nice video from India, where they are optimizing planting the crops with AI.
Then, we have the policy formulation and evaluation. This is more related to recent developments. So, now with these generative AI tools, what we can do is that we can create different scenarios. So, if we were to implement this kind of policy X, give me three different scenarios, what could happen? How could this affect? So, quite amazing. We can use AI as this kind of strategy, discussing and brainstorming partner.
And then we have an example of enhanced public engagement. So, what we can do with AI, we can analyse, for example, the news, and we can analyse the media related to, for example, some policy. And then, instead of just guessing, okay, how people might feel about this, or instead of doing some questionnaire that would require a lot of labor, we can use AI more easily to detect, okay, this is how people are feeling about this new policy.
All right. Now that we've gone through the examples, let's do this little comparison of humans and AI. So, I want you to think about this. Okay, what could be our strengths and what kind of strengths AI could have. And I think maybe you can already guess most of what I'm about to say.
For humans, emotional intelligence. This is a huge one. So, machines, they don't have, as we now all know, they don't have emotions. So, therefore, they can mimic, and we can sometimes think that they can, like, pretend to understand human emotions, but really, they can't. So, this is something that, at least currently, is only a skill of humans. Then another huge one, creativity and innovation. So, the models, even though they are extremely capable and we can use them to brainstorm the ideas, I still strongly believe that the best ideas and the real creativity, it comes from us humans. So, don't rely too much on AI when it comes to innovation.
Then we have the ethical and moral reasoning, something you really don't want to give these statistical models, but you want to keep it yourselves. Then also, adaptability and flexibility. So, even though the GenAI systems are very capable, they are not as flexible as humans.
Complex decision making, also another area where you still want to rely a lot of human intelligence. Interpersonal interactions, strategic planning, ethical governance. These also are kinds of areas where you really want to put a lot of emphasis on humans, and you want to rely on humans to make these kinds of decisions. Something you don't want to outsource to an LLM.
All right. Now that we have taken a look at the ‑‑ where the humans shine, let's take a look at what are the greatest strengths of AI. So, the first huge one: Speed and efficiency. So, it's amazing how fast these models are. A great example of this is something I highly recommend you to try out, is Google's DeepResearch, which is this new kind of tool that you can do research with. And what's amazing about it is, if you were to do that same amount of research yourself, I would say that it would often take you a few days, at least a day, a full day of work. But with Google DeepResearch, you can go through more than 100 resources often in a short time as 10 to 15 minutes, and it will write your research report sometimes as long as 50 pages. So, speed and efficiency is something where the AI really shines.
Another example of this is AI is really good at going through a lot, huge amounts of data. For AI, it really doesn't matter whether you have a gigabit or terabyte of data. It really doesn't matter.
Scalability, really nice feature of AI, as mentioned. Doesn't matter whether you have huge amounts of data or just a bit of data, usually you can get the outcome in quite similar times.
And then, we have consistency and precision. So, really nice feature of especially these AI systems that are using more traditional machine learning. They are very predictable and there is a lot of consistency. So, if you run it multiple times, you get the same answer.
Then we have the availability. As mentioned in the Estonia Burokratt example. So, instead of just asking, or there's the few persons in some office taking calls and responding to messages, the Burokratt, it's always open, so 24/7 availability.
AI is very good at detecting patterns from data. It's good at automation of routine tasks, as we've seen. Then we have the predictive analytics capabilities. And now, since the rise of generative AI and LLMs, we also have this natural language capabilities that are quite, quite amazing.
All right. I think now that we have, like, gone through the fundamentals, so we look at how data, cloud, and AI are related, and then we've also looked at some examples of how different governments are taking advantage of AI and also look at what are the greatest strengths of AI systems and what are the greatest strengths of us humans. I think now it's a great time to move on and get really hands on with AI.
So, in this section, what we're going to do is that we are going to look at the ‑‑ in more detail about the differences of predictive AI and generative AI, and then, you will also ‑‑ I will give you some time to try out Google Gemini and also another tool from Google called Notebook LM. And then, we are also going to discuss about, okay, how ‑‑ what is a good prompt and what are the efficient ways of using these tools. So, let's dive in.
Okay, what is predictive AI and how is it different from generative AI? So, predictive AI, it's the more traditional machine learning, something I started with ten years ago. And I think it's really important to keep the capabilities of predictive AI in mind, because now there's so much hype around generative AI that I see in a lot of companies that they are not taking the full advantage of AI, because they are neglecting the predictive AI part.
So, what is predictive AI? It's, when we have transactional, numerical data, with predictive AI, we can predict, okay, what will happen next? So, one classical example of this is the weather prediction. So, we have a lot of data, how the weather was in the past, and based on that, we can predict, okay, how is the weather going to be tomorrow? Really difficult task, but we've made a lot of progress with that in the recent years.
But important to keep in mind, this has nothing to do with the generative AI, so it's a different form. And how you could use, and when you should use predictive AI versus generative AI is something really important to think about when you start applying AI in your processes.
Okay, so, predictive AI is all about predicting what the future will be when we have the transactional data.
Okay, how about generative AI? Generative AI, as the name implies, it's about creating something new, so generating. So, with generative AI, we are generating ‑‑ it could be text. Oftentimes, the most familiar example is these chatbots that are highly capable of generating text, but it can also be images, it can be code, it can be PDF files, it can be transactional data files, anything, basically, you can imagine. You can generate with these Generative AI models.
I think one word you heard often is LLMs. So, large language models. How are they related to Generative AI? And what are they? So, LLMs ‑‑ large language models ‑‑ they are specific AI systems built on top of huge neural networks that are trained, basically, with the whole data of the open Internet, and they are really capable of predicting the next word.
So, when you, for example, go to basically any of these chat AI applications, what you are interacting with there is these LLMs. So large language models. But yeah, that is often what is behind the Generative AI. The technology behind Generative AI is most often LLMs, large language models.
A few more ‑‑ just to recap, once again, the differences. So, predictive AI ‑‑ we try to anticipate the future trends, so what will happen. We have some ‑‑ we want to derive some data‑driven insight. So, if you have a lot of transactional data, no matter what transactional data, but if you have a lot of numeric data, what you want to use is predictive analytics.
So, examples: Policy forecasting, resource allocation, infrastructure planning, stuff like that.
Okay, how about Generative AI? So, when you want to apply it is when you want to create new content. Its emphasis is on creativity and innovation, so you can create emails, you can fill out forms, you can generate images, and so on.
All right. Now that there is so much focus and hype around Generative AI, this presentation wouldn't be a full one if we wouldn't dive somewhat deep into GenAI, so let's do that now, and let's look at some of the use cases of GenAI more closely.
So, with GenAI, you can really generate anything, as we've already discussed, and now we are going to dive deeper into some examples. One of which is the new search. It's quite remarkable how generative AI and LLMs are changing the way we gather information. So, now the Generative AI is really helping us. I would say it's not only about speed, but it's more about the nuance and the details. So, with this, I think really the key benefit of this Generative AI search is that now instead of the blue links, what you get is a brief overview of what you have asked. And it's ‑‑ in addition, we have the links. So, then, if we want to dig deeper, then we can, like ‑‑ and what we often should do is that we should really take a look at the resources and verify ourselves that, okay, this is exactly what I need.
But yeah, this is really fascinating, and I think it has made a huge impact on the way we can find information and gather information.
Okay, another example is the email writing. So, in Gmail, I think it's especially lifesaver for us non‑native English speakers. So, before GenAI, I think probably like two hours of my working days went when I was writing emails, because ‑‑ and I was always somewhat worried, okay, whether this is ‑‑ is this appropriate enough, or is my language, like, suitable for this kind of discussion? Now, I think my time management in terms of writing emails, it has gone from like two hours to maybe 15 minutes. So, something I ‑‑ if you don't do this, I really recommend you to try it out. It's a huge, huge lifesaver.
And how does it work? And how do I use it is that I just weekly type out what I need, and then I let the AI to be my, like, English assistant so that the AI makes my emails more formalized and makes sure they are grammarly correct and perfect.
AI can also help you write documents and slides.
>> AUDIENCE: (Off microphone).
>> Alecsi PAAVOLA: Great question. We will have a different Q&A session in the end, but I will repeat that and take it now. So, the question is, how can we take advantage of this tool, and is it already ‑‑ is it a separate tool or is it already in Gmail? And at least in my Gmail, it's already there.
I think the amazing people from Google will address this later today, whether it's globally available, or is it still in beta. Maybe we have a more clear answer here.
>> Olga: Yeah, just very quickly. We are rolling this out from region to region. And I think Your Excellency, you're from Egypt. What we're also doing is that, specifically for Arabic, which is a language of many dialects, we're making sure that our system, they speak different dialects. And I think right now we're covering about 20 dialects in Arabic. And actually, we see that Arabic is one of the highest ‑‑ is one of the languages that are most used, in our systems that can already perform those tasks in about 45 languages. So, just a short note, but it should be embedded in your Gmail already. If not yet, this is coming.
>> ALEKSI PAAVOLA: Thank you so much. All right, yeah, we can take advantage of AI in writing documents and slides. This is something you can do inside, in your Google documents and slides, or you can use Gemini Chat, which is what I currently prefer. But that's something we're going to try soon.
Then something I'm really excited is that the AI's ability to help us in generating these Google Sheets or Excel formulas. So, it's simply amazing how good these tools are, and I find it difficult when I'm giving these kinds of trainings related to writing these Sheets or Excel formulas ‑‑ I find it difficult to find examples that the participants cannot immediately solve. So, this is also something, if you have maybe shied away from Google Sheets or Excel, so numbers are really not your thing, now it's a great time to re‑evaluate and try it out. So, it's amazing how you can just describe the function you need, and AI will take care of it.
All right. And then, we have the NotebookLM, so something we are also going to try out today. This is an amazing tool for more deeper work, and when you have the need to upload a lot of your own resources, it's very helpful. And then, inside the NotebookLM, there's also a really nice property of this called Audio Overviews, which is something I really encourage you to try out. It's quite amazing that what you can do with it, you can upload your own document or research paper, and NotebookLM, it will generate a podcast for that document. And the quality of the podcasts, it's totally amazing.
Okay. Then, something that we always have to think about and something I get a lot of questions is how about the privacy and confidentiality? And yeah, the nice property of Google is that they've really taken that into account. For example, when it comes to NotebookLM, Google is not using your data to train the models, so you can ‑‑ and they are not reviewed by any human, so it's really something that I really value when using this.
You can also connect your Google Workspace to NotebookLM, something I also really encourage you to try out. And what it enables is that, after you have made this connection, you can basically ask and chat with all the documents you have in your Google Workspace. So, if ‑‑ especially if you have a lot of data, it's amazing how much efficiency gains this feature can bring.
All right. It's time to test out the NotebookLM. So, you can scan the link or you can go directly to notebooklm.Google.com, and there you will find the NotebookLM. And before you start, or if you are really eager, you can start already, and if you are familiar. But what I'm going to do right now is we're going to switch gears and I'm going to show you a quick demo of how the NotebookLM works. Splashgts. So, this is ‑‑ when I went to notebooklm.Google.com, this is how it looks for me. I'm already signed in, and here I have a few Notebooks already.
But what we're going to do here is that I have this document. It's called AI Sprinters. So, it's in the document form what we are going through today. And what we're going to do here is I'm going to create a new ‑‑ I'm going to select here the "Create New." And we are going to upload this AI Sprinters document to NotebookLM. And it's going to take a while. And while we wait ‑‑ okay, I think ‑‑ yeah, I think we're done.
So, now we have the document uploaded, and what we're going to do next is that we're going to ask a question. So, the question we're going to ask ‑‑ I'm not typing it because it's going to take too long. So, I copy‑pasted the question: How can governments use the report to ethically implement AI in their department? So, that's the question we are going to ask.
And now, if you ask this same question to any AI chatbot, it's going to respond based on maybe a search, or based on the knowledge in the parameters of the model. But now, the great thing about NotebookLM is that now it's going to answer based on the document we uploaded. And we will see how it looks in a minute.
We also get a reference to the document, and we can really dive deep into the sources. So, this is a really nice feature. So, when we hover over the sources, we can see, okay, this part of the answer is actually coming from this part of the source document.
Yeah, but I will give you now a few minutes to try it out. You can use it with your app, or if you have a computer, you can use it with your computer. There's also an app you can download from iOS or Android store. But yeah, please, please try it.
>> Olga: Yeah, it's working. One of the special features about this product is that it can also understand at the same time documents written in different languages. For example, if you're representing a country where there is more than one language spoken and more than one language in which you receive, you know, requests from your constituency, you can upload directly in one NotebookLM documents in different languages, and then ask questions or ask to summarize these documents in the language that you use, and the system will be able to understand these different languages and give you output in the language that you prefer. Something that I wanted to highlight, knowing that we do have lots of people from countries where more than one language is in use.
>> ALEKSI PAAVOLA: Okay, let's do one more minute, and here just to show you the podcast I mentioned here in my screen on the right side. You can see this deep‑dive conversation here. If I were to generate, it would generate a podcast from the uploaded report. And that's something I really, really think you should try. You will be quite amazed.
Can you come again?
>> AUDIENCE: (Off microphone).
>> ALEKSI PAAVOLA: Yes, great, great question. Is the NotebookLM able to summarize documents? And the answer is, yes. You can upload a document, and then, instead of asking a specific question, you can ask, "Okay, please provide me a short summary" of that document. Yes, you can do that there as well.
All right. Can we please switch back to slides? All right. Thank you so much. Let's continue. You can then, after the session, you can continue playing with NotebookLM. All right.
Now, when we're talking about generative AI, I think it's really important to think about how we interact with it. So, the responses we get are only going to be as good as the prompts or the instructions we write. And what I've seen in the trainings, in the recent, let's say year or two, when I've worked with a lot of people and with these generative AI systems, I see a huge difference between what you can get out of these AI systems from people to people.
And what's the difference is that if you really ‑‑ this is really a skill. It's a skill you need to master. So, when you start interacting with these generative AI systems, you will not get everything out of them immediately, but it's a skill, it takes a lot of practice, but with practice, you can become very good and efficient with these GenAI systems. So, I think this is, like, probably one of the most important skills you can have at this time.
So, we're going to briefly look at that. So, prompts, the instructions we give to generative AI, AI models. So, let's discuss this, and what is a good prompt?
So, one good way to approach it is that we split the prompt into these four steps. So, in our prompt, we give persona, we give a task, we give context, and we give the format. So, for example, you are a public engagement officer, or you are an experienced financial director or marketing director, or whatever you need. So, first you give the persona. Then you give the task. Okay, write an email. Create a summary. And so on. And then you give some context. So, write a summary ‑‑ so, you are a marketing director of a large company. Write a summary of... and then you give context, right? Of the attached PDF document.
And then, the last thing you give the output format. So, I want the output to be a really nice business email, or I want bullet points, or you want Python code. It can be anything. But this is a good way to approach it. So, when you are writing prompts, always think about it that way. So, okay, you give the persona, you give the task, you give the context, and then, what is the output format you're looking for?
Then some more prompt writing tips. Use natural language. I've seen that, you know, maybe some of you, you're really good at doing traditional Google searches, and this can be somewhat difficult for you because you're used to interacting with Google with this, you know, you're very good at figuring out the key terms. But when we are using generative AI, we want to use natural language. So, think about it like talking to your friend or colleague.
Then, also, be specific, so be very specific. What is it you are trying to do? So, if your prompts are really short, probably you are not specific enough. And then iterate. So, no matter how good you are with these systems, it's often the case that you need to, like, try a few times, or multiple times, and iterate.
Avoid complexity. So, think about it like giving ‑‑ how would you give some task to your colleague, for example? And make it conversational. So, remember that you don't have to get it exactly right in the first try, but then you can make corrections, similar you could correct your colleague when he's responding in some way you are not looking for.
And remember this, that this is probably the thing I see the most or something that people are not taking full advantage of when using generative AI, is that they are not providing their own documents. So, try to provide your own documents, and the quality of the answers will be completely different. A lot better.
All right. Now we are going to try Gemini. So, if you head over to Gemini.Google.com, or scan the QR code, and can we switch the screen? Thank you so much. So, here you can see my Gemini account. And here, if you have a page subscription, you can select the model. We are going to use the 2.5 Pro. This is the model I really like to use, really capable model for different kinds of tasks.
And then, here is our example question. So, what we're going to do is we're going to just paste this question into the chat. So, "I would like to write a report on the best use of AI in medical supply logistics in Kenya. Please write an outline for this report."
Let's hit Enter and let Gemini work its magic, and it's going to provide us a nice outline for this AI in medical supply logistics in Kenya. Actually, what it did ‑‑ because I have this advanced subscription ‑‑ it's now actually suggesting us to do a deep research. This will take like ten minutes, so I'm not going to run it here. But something I really encourage you all to try.
But what you could do now is that, just go to Gemini.com and take a new chat and ask anything and get a little feel about how Gemini works. And remember the four suggestions about how to write good prompts. So, maybe that's something you can implement when you try it out. But yeah, just head over to Gemini.Google.com and take it for a spin, and I'm going to give you a few minutes to try it out.
All right, can we please have a switch? Thank you. This is something you can also continue after the presentation.
Yeah, but let's move forward, and a few words about hallucinations. So, probably a term you've come across somewhere. It's something that is often discussed in the media when it comes to GenAI. And what it basically means is that the AI, generative AI is giving you plausible sounding answers, which are not correct. And this is ‑‑ it's really good to keep in mind that this can happen. And the models, we've come a long way, let's say. A few years ago, the issue related to hallucination was much, much, much bigger, and now it's actually quite rare with these modern systems to see this. But it's still something that it's good to keep in mind.
And there's ways to mitigate that. So, when you give your own files, your own context, this is probably one of the greatest ways to mitigate hallucinations. And then, another thing is when the model uses Search, so that's also a way to sort of like ground the answers for AI so that it, instead of answering based on the knowledge in the parameters of the model, it answers based on what it found on Google.
But yeah, something to keep in mind. Hopefully, we'll get rid of this some day. And maybe just to give some perspective about this is that, I don't know about you, but when I interact with my colleagues, they are not always right, you know? So, this is something that happens with humans as well. So, I figure ‑‑ it has happened to me multiple times that my colleague is giving me a plausible sounding answer, which is actually not true, so it happens with humans also.
But nevertheless, remember to check for accuracy. So, when you are using the generative AI models, check that ‑‑ always check with the human that the output is what you are looking for. You can ask the AI to provide sources. You can use multiple tools. There are a lot of ways to verify the content.
And also, ensure your safety. So, always use the human insight. Follow the data rules of your organization and be aware that sometimes these AIs can produce harmful outputs.
We're going to skip this, as we just tried out the Gemini, but something you can do after this presentation is to try summarizing documents, either with Gemini or with NotebookLM. So, both of those tools are highly capable of summarizing documents.
Let's switch gears to the last section. So, let's talk briefly about challenges and strategies when it comes to AI. We have a question coming from the audience.
>> AUDIENCE: (Off microphone)
>> ALEKSI PAAVOLA: Yeah, we have a question related, which tool I would recommend. Should you use Gemini for summarizing or should you use NotebookLM? I would say you can use either one. I think both are very capable of doing summarization. So, I would say it doesn't really matter. You can use either one.
>> AUDIENCE: (Off microphone)
>> ALEKSI PAAVOLA: Yeah, I can repeat the question. The question was, when should I use Notebook LM, and when should I go to Gemini? And that's actually a great question. And I will say that when you want to provide a lot of your own resources, then the NotebookLM is the right tool. So, let's say you have lots of like government documents you have to work with, then you can upload the documents to NotebookLM. So, for this kind of, like, deeper research, the NotebookLM is the right tool.
But then, when you want to quickly want to, like ‑‑ let's say you want to create or brainstorm. You have some idea related to ‑‑ for example, we had the medical logistics in Kenya. You want to brainstorm. You want to create an outline. You want to create something quickly. And you don't have tens of PDFs for material. Then you would use ‑‑ or I would prefer Gemini for that. But the capabilities, they are, like, quite similar. Great, great question.
Yeah, but let's talk briefly about the challenges and strategies. So, we are going to take a look at the access to models, misuse, the data gap, and then also some infrastructure‑related challenges. Before we dive into challenges, let's briefly discuss the potential.
So, I think the consensus of the potential of AI, it looks something like this. So, it is estimated that the implications and the contribution of AI to digital economy could be $7.4 billion by 2033. So, huge opportunity. But what are some of the roadblocks ahead? So, let's take a look at that.
First of all, poor connectivity. So, in order for us to use the AI and the capabilities, we need good connectivity. It's a must, in order to take full advantage. And this is something I think there's still a lot of places where the connectivity is not good enough to really take advantage of AI.
Then we have the strict data localization policies. I always have one suitcase filled up and ready to go, if he who decides to make the policies so strict, that I cannot access the best cloud models. So, let's hope I can live in Finland in the future as well, but this is something that I'm actually quite, quite worried that there are places that have such strict data policies that people are unable to access the latest and greatest technology.
Then we have the misuse. It's unfortunate that this GenAI technology is also capable of producing a lot of harmful information at huge scale, and this is true for text and images and video. So, good to keep in mind that when you are interacting with someone online, maybe you are not interacting with a human. And if you see some video footage somewhere, you cannot be fully sure that it's something that has actually happened.
Then we have the potential of widening the economic divide. So, we already see that the AI ‑‑ the potential and the leverage AI can give is enormous. And I'm somewhat worried that there's ‑‑ I see two potential pathways forward. So, one of which is that, if everyone can get access to these AI tools, we are going to have a really amazing future, because the productivity of one person, we can multiply that easily. But on the other hand, if only some of the people get access to these models, then we are going to see this huge widening of the economic. So, something I'm somewhat, somewhat worried about.
Then we have the data gap. So, what it means is that, as we learned in the beginning of this session, data is the building block of AI. So, without the data, without the infrastructure, we cannot train the AI models. So, therefore, if we have logistics difficulties, infrastructure challenges, limited financial resources, political instability, or lack of standardized methods for collecting data, we cannot take full advantage of AI.
Okay. What are some of the roadblocks of collecting and managing data efficiently? So, when we talk about collecting data, we need to have the infrastructure in place. And we also need to be ethical about the collection of data. Then we have the processing of the data. Then we also need the infrastructure. We need a lot of skilled employees. So, we need technical expertise, we need training, we need education.
Okay, how can we then build strong data infrastructure? So, what are the building blocks? First of all, we need to establish the whole‑of‑government commitment to really take advantage and use the data, and we want to make the data publicly available. I've seen amazing things happen when governments make the data publicly available and let entrepreneurs build and use that data to train, for example, AI models.
Also important to facilitate that the data flows before different organizations, to avoid silos. And probably, the most critical thing is to improve the data infrastructure so that everyone can have access to high‑speed Internet and these extremely capable models that are becoming better and better.
So, important things. Regulatory oversight, continuous monitoring, stakeholder engagement. Before we dive into discussions and Q&A, let's quickly take a look at the few areas where I think AI's going to have a huge impact. Culture is one.
So, it's amazing what we can do with AI in terms of analysing the culture‑related data. Also, another great recent development is the language capabilities of the models. So, we have a lot of very small languages. And now, with AI, we can really understand and access these small languages and all the regions of their culture. So, I think AI's going to have a great, great, positive impact on culture.
Then, a huge one, education. We are at the place ‑‑ it's amazing that we are basically, in terms of technologies, we are at the place where everyone can have their own teacher. And what I really recommend you to try out is, try to learn a new topic, for example, with Gemini. So, use Gemini as your personal teacher, and you will be quite amazed about how efficiently you can learn with AI. So, this is something I'm really hoping that everyone in this world would have their own AI teacher.
Okay, in health care, this is the other huge one. I have a few of my friends work as a doctor, and they use generative AI models at their daily work, and they are really impressed about the capabilities, and they've told me numerous times how great it is that you have this kind of sparring partner when it comes to diagnosis, for example. So, yeah, hopefully, really amazing things happening in health care, and I think AI can enable every one of us to have a doctor available at all times.
AI is also going to bring us a lot of new jobs. So, there's a lot of new kinds of roles emerging, so we are going to need people who are capable of working with AI and prompting AI, and there's a lot of positive estimates about how AI will affect the job creation. And I wanted to bring this, because in the media, we see a lot of kind of negative news related to job creation, and it's unfortunate, because I think there's also this positive side, that we know that there are jobs that AI will probably replace, but then, there are a lot of jobs that AI will also create. So, therefore, I really wanted to end with this positive note, that AI will bring us a lot of good things.
All right. Let's discuss. So, what I would like to do next is let's share your thoughts, what is happening in your country? How might cultural or regulatory factors influence AI adoption in your country? Or what ethical or practical concerns should governments consider when implementing AI tools? Feel free to share your comments, or if you have any questions, I'm more than happy to answer any questions.
Actually, yeah, we have a mic coming. Yeah. That's excellent. Yes.
>> AUDIENCE: Thank you so much. Congratulations. This has been a great learning experience for me. As a politician, it's difficult to come up with some sort of policy or regulations on this area without having a solid base on this technology, so this session is very, very important for people like myself.
But I have sort of some concerns. You mentioned about one of the main pillars of Google is responsible AI development and deployment. Would this include an effort to make your machines more transparent? For instance, how they achieve their conclusions, especially when it comes to policy suggestions, so that we know that their suggestions, their recommendations are solid, based on some sort of logic? That's one.
The other more practical question would be, from the ranges of AI applications is other Google resources, would this be available in one simple gadget? For instance, if I buy ‑‑ would it be possible for me to have access to all of these applications and other facilities as well, such as Google for small businesses, enterprises and others? So, that kind of discussion is really important for people from my background or from developing countries. So, we would love to access all of these facilities, but then it's kind of scattered around. So, centralized access would also help people to enjoy these technologies.
To end, I would like to congratulate one more time. This has been an amazing session. Thank you.
>> ALEKSI PAAVOLA: You want to take that?
>> Olga REIS: Hello, everyone. Maybe a brief introduction because not all of you were here in the beginning. I understand it's an early‑morning session. My name is Olga Reis and I cover public policy for emerging markets at Google.
So, let me cover your question related to how people can access Google technologies and what we as a company do, especially given that there are, you know, users in different countries, have different, you know, purchase power, right?
First of all, many of our products, such as Gemini, they are available for free, so you just need to have access to the Internet. And really, with the latest updates, we are constantly, practically on a quarterly basis, updating our underlying models. Gemini's capacity and capability of the free version is really great. I personally, for my personal capacity, don't use actually paid Gemini. Of course, I do have access to it as a Google employee. But for example, when I do some of the stuff for me personally, I don't use my work account. And for me, free version is just enough.
For example, I learn a new language, which is Turkish, and I use Gemini to help me check my homework, understand why I might miss text, and it's just enough. The free version is already big, and we as a company have always been focusing on giving access to information, and Gemini is one of the tools of how we give access to information.
And we also do ‑‑ we have special programmes and offerings for some ‑‑ for example, for NGOs. We do understand that very often civil society and NGOs are low‑resourced, let's say actors in the field. And so, with Google for nonprofit programme, we give access to most powerful AI technologies for users who are officially registered for free. And we actually just announced two weeks ago that we're expanding this programme to 100 more countries, so it was available in 65 countries, and now, 100 more countries will come and be supported in next quarter of this year. So, that's on, let's say on the commercial side of things.
Maybe I will quickly comment and then give it back to Aleksi in terms of transparency on our models. This is something, we understand that there is a huge, you know, expectations from companies like Google, but also from our peers to make models more transparent. And one of the ways of how we do that is we publish the so‑called models scorecards, where we practically, you know, in brief terms, discuss what the models are capable of, how they were trained. And this is also part of our commitment and what we are supposed to do as a company that signed up for G7 Hiroshima process, and it is a company that is a member of Frontline Models Forum, kind of international self‑regulatory organizations, but not organizations as really forums, where we as a company are a member of and where we are disclosing as much as we can while protecting our commercial interests, because of course, model is a commercial interest. What we are doing, what these models are capable of, how we tested them, including redteaming. So, that's something that we are doing, and it's a process. It's not ‑‑ it's a process because we and our industry peers are constantly updating our approach to making ourselves more transparent. And definitely, you know, if we have this conversation in one year, probably the picture will be slightly different.
>> ALEKSI PAAVOLA: Yeah. I think they were next. Before the question, just a quick reminder. I would really, really appreciate if you could scan this QR code and leave some feedback. That would be awesome, if you could do that. But yeah, let's continue with questions.
>> AUDIENCE: Thank you so much. Does the Google AI technologies murder historical patterns of racial segregation? Thank you.
>> ALEKSI PAAVOLA: Do you want to comment on that?
>> No, I have a question.
>> ALEKSI PAAVOLA: Let me comment on that first. So, I think that's a great question. Unfortunately, it's something I don't consider myself an expert enough on that area to respond to that.
>> OLGA REIS: If I may comment quickly on this point. The way ‑‑ and as discussed early on in our session ‑‑ the quality of the model and the output that you can, you know, get from the model really depends on the dataset. And what we as a company do in this, to ensure that our datasets, they are comprehensive and they represent different views, including this historical content. So, we really invest a lot of resources, both in terms of, you know, human resources, but also financial resources, to acquire the best data, to make sure that such cultural and political and historical context are taken into account while we train into the datasets on which we base our model. So, that's fundamental, I would say, in terms of how we should approach this very, very important point. Thanks for raising it.
>> AUDIENCE: Yeah, thank you very much for the very insightful presentation. I'm a parliamentarian who introduced to Egypt, the Parliament, the first draft bill on AI governance, and it is very much around whatever is ethical when it comes to the usage of the big data.
When it comes to Google, actually, and politically, I have very much of a concern on how the big data over time could be used at war times. And the situation at Gaza lately was very clear when it comes to you terminating and please were against the (?) project when it comes to the servers which serves the IDF. So, a big question here is the big text, which are actually mostly USA‑based big techs, how can this be politicized? How can the big data and the AI not used in wars? And how can we guarantee the bias?
One of the main things which I think the governments of many African countries think twice about it is to get the servers, to host the servers in their home countries, given that what guarantees the privacy. Same goes, there's always two‑sides thing. What you have presented is excellent. On every level, the AI's taking us, as humans, to a total different level ‑‑ agriculture, health care, all the good things. I'm talking about the other side of the coin, which is still for any parliamentarian or for any politician is quite considerable. So, if at Google you have ‑‑ and definitely you have discussed thoroughly ‑‑ how can you as big tech companies guarantee that the bias is minimal, at least, or that it's not used for war, it's not used for political purposes, it's not used for fake propaganda, it's not altering elections at the country or another? This is the other side of the coin of AI usage which I'm very much concerned with. Thank you.
>> ALEKSI PAAVOLA: Let me ask you ‑‑
>> OLGA REIS: Let me take this question as a Google representative. The points that you raised, they are very important. And one of the ‑‑ again, one of the ways how we address such concerns is, again, ensuring that our datasets are representative, that they take different points of view. I, myself, I am based out of Dubai, so I do live for four years now in the GCC region, right, in the Gulf region, and I can share, you know, in terms of the discussion and what we did also to mitigate and navigate this very, very challenging situation that we are all facing in the Middle East.
We have many employees ‑‑ hundreds of employees, of course, representing both the Arab culture of Palestinian descent working side by side and making sure we support our users in Palestine, in all the areas that are affected by what is going on in the Middle East. So, this is something that we are taking very seriously. We do extend our support to local NGOs supporting communities, and this goes on both sides. But this is definitely very challenging and stressful. I saw it myself firsthand, when someone who is based in our headquarters in the Middle East, and the company takes this very seriously and will continue doing so. Hopefully, peace will come to this region, not only the Middle East.
>> AUDIENCE: (Off microphone)
>> OLGA REIS: Yes. One of the things I wanted to mention, because this is where AI comes, you know, interplays with content, right? And this is definitely something that we address very seriously, two ways how we ensure that AI systems are not misused and are not used to produce fake content.
First of all, we were one of the first companies to introduce what we call SynthID, which is basically virtual labelling of content that was produced with AI, so you can, actually, as a user, check and see whether this was something that was produced with the use of AI in the output. That's one thing. So, really, technically ‑‑ okay, we need to finish.
We need to, and we'll continue working to ensure that we address such concerns with the technology. And secondly, really using AI at scale to detect and delete/remove content that is harmful content. And I can share with you some of the data, how we do that, but this is something that we ramp up internally, and we take this very seriously.
And I think with this, we need to finish this session, right? The organizers just reminded me. But I will stay here, colleagues, and will be happy to answer your questions. Do we have any more time or not? Please. No.
>> ALEKSI PAAVOLA: Yeah, unfortunately, we are out of time. Thank you so much for everyone attending. Go try out Google NotebookLM and Gemini. And I will be here. If you have any questions, you can come talk to me. Thank you so much.
