---
id: "workshop:getting-started-with-source-grounded-ai"
type: "InnovateUS workshop"
title: "Getting Started With Source-Grounded AI"
url: "https://innovate-us.org/getting-started-with-source-grounded-ai"
date: "2026-04-17"
duration: "60 minutes"
series: "The Prompting Lab: Real Prompts, Real Challenges, All Platforms"
instructors: ["Deborah Stine"]
recording: "https://www.youtube.com/watch?v=_f8ll2doV4Y"
transcript: "automatic, from the recording"
source: "InnovateUS (innovate-us.org)"
licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org"
---

# Getting Started With Source-Grounded AI

InnovateUS workshop · 17 April 2026 · 60 minutes

This session introduces source-grounded AI tools, with a focus on NotebookLM, which generate outputs using only the curated documents provided by the user. Participants will also explore practical ways to approximate retrieval-augmented generation (RAG) in tools such as ChatGPT and Microsoft Copilot by constraining inputs and structuring prompts.

The session will examine both the opportunities and limitations of source-grounded AI in public sector contexts. Participants will learn how to set up a secure workspace, upload and organize documents, and use core NotebookLM features—such as Mind Maps—to explore dense materials and generate insights that can be used in policy communication.

Participants should bring a publicly available document to use during the session exercises.

## Learning goals

By the end of this session, participants will be able to:

- Explain how source-grounded AI differs from general-purpose AI tools

- Identify key opportunities and limitations of source-grounded approaches

- Upload and manage source documents while adhering to public sector data governance standards

- Use AI-generated mind maps to navigate complex materials and support downstream communication

## Speakers

- **Deborah Stine**, Senior Fellow, Innovate US

Part of the series [The Prompting Lab: Real Prompts, Real Challenges, All Platforms](https://innovate-us.org/workshop-series/the-prompting-lab-series).

Recording: https://www.youtube.com/watch?v=_f8ll2doV4Y

## Transcript

Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/getting-started-with-source-grounded-ai.txt (plain text) and https://innovate-us.org/getting-started-with-source-grounded-ai.json (every segment with its start and end second).

**Deborah Stine**

[0:00](https://youtu.be/_f8ll2doV4Y?t=0) Hi everybody, my name is Debbie Stein and I'm pleased to be with you here today to talk about getting started with source-grounded AI. As I mentioned in the chat, in case you don't see it, I'd really like to get a feeling from you as to what it is that you like. Have you ever used Notepik LM or some other source-grounded AI, and if so, how you use it for your work activities, just so we know sort of where everybody is at. And as always, I like to know where you're from and what it is that you do where you think this might be useful. So with that, I want to start by mentioning this is the first of four. So this is the overview version that talks about really what the opportunities and challenges are with source-grounded AI. Over the next two Fridays, I'll talk about in more detail how to go to written and visual communications and also presentations, video and oral communications. There's a lot of tips and tricks and things like that that you need to use to get these tools to meet your goals. And then in the last session, we have a lab office hours where it's just a Q&A session. One of the things I'm going to ask you to do is, so you can go get started, is go ahead and bring up Notepik LM. I did check to see that you could do this stuff with the free version. Reality might be something else, but please try to bring it up. And then just identify like one document, something that's original. It might be from your organization. It might be something you have done, but any sort of original information or alternatively, maybe legislation, a bill, a law, something that is unique that you're trying to understand or communicate to others. So I'm going to start with an overview. I'm going to do

[2:07](https://youtu.be/_f8ll2doV4Y?t=127) sort of a meet Notepik LM. We'll go through this activity I just mentioned, talk about non-Notepik LM options. We can kind of approximate what it does with other things like ChatGPT and Copilot, some of the challenges, and then talk about what's next. Because Notepik LM takes time to generate this stuff. It's not like forever, but it is minutes. I'm really going to focus a lot at the beginning on the activity itself. So your information from your document can be generated in the background at the same time as we're kind of going through the content. And hopefully before this webinar is over, you'll have some actual content that you can look at for what it is that you're doing. So what is grounded intelligence and why does it help in government settings? So oftentimes you have a lot of things going on. Some big giant bill comes out, some law, and it can be a bit overwhelming to figure out exactly what's in there, what it means for you, and to try to do these things. And if you try to do like some other tools, let's just say ChatGPT as an example, there's a number of challenges. You know, it's based on the knowledge that's brought into it. Well, if that bill just say, let's just say it came out today, and somebody wants you to summarize it for tomorrow, it's not going to be a very useful tool for you because it's not part of its memory. So you want to be able to upload things that are happening like right now. The other thing is if you're like me, I'm a content creator, and I don't mean I'm on Instagram or something. What that means is I write original policy analysis for organizations. And so I'm trying to, and oftentimes where the areas I work

[4:09](https://youtu.be/_f8ll2doV4Y?t=249) in are areas where there's not much information at all. It's leading edge on cutting edge of science and tech. And so therefore I only want to look at my document. I don't want to look at everything else out there on the web, but I do want to use AI tools because they can help me communicate the results of it better. And then this knowledge gap for us is really related to the fact that, you know, as people retire, things like that, we kind of lose the knowledge that we have inherently in our governments and other systems. And so this is, you know, a way to help focus that information. One idea that someone shared with me, for example, is they got their, you know, their senior person who was the person everybody went to for questions, and they started having that person write down the answers to all those questions and then create their own knowledge database. And then by incorporating that into whatever AI tools that they are using, then they have not just what's in the web, but this internal document that can really help new people entering the system for like common questions, problems, and so forth that come up.

[5:20](https://youtu.be/_f8ll2doV4Y?t=320) So what this granite intelligence does is it really, it's not, it only looks at the documents that you either approve for use or the documents that you provide to it. And it can synthesize those all together and create an amazingly coherent message for what it is that you're trying to do. Now, this is not to say it's perfect. You have to know what's there. This is definitely something where, you know, you need the knowledge yourself to make sure that it's accurate. But if you have that knowledge, like in other words, like when I'm writing an original analysis, I know what I wrote, right? So I'm just checking to see if it accurately is portraying what it is that I'm seeing. And if it's a document you don't know, let's just say something that's new in legislation, you still have to go back and check and say, okay, I'm going to go, but it'll provide that citation. So you can go back and look at the actual language and make sure that it's an accurate representation. So you'll see here, these kinds of usage standards, you know, some of these acceptable uses like internal brainstorming, research synthesis, non-sensitive data analysis, things that you can't do typically are personal information, confidential health data, any sort of outputs that haven't had like, you know, using what's there without any human review, what they call human in the loop. And in terms of constraints, you don't want to treat these things as final legal documents. I've never produced any of these things where I haven't wanted to change something in it. Even I see here, I'd miss this typo that it has here.

[7:06](https://youtu.be/_f8ll2doV4Y?t=426) It says traded instead of treated, right? So there are always going to be at least small arrows, arrows, typos, things like that. It's just like an intern or something like that. It does a good job, but it's not a perfect job. Okay. So this is a video, this was produced using a notebook LM, and it talks about, it gives you an overview of how this works. You know, generative AI has this huge potential to completely change how our governments work, making everything faster and way more responsive. But what happens if the AI just starts making stuff up? That's really the big question, right? Can we actually build an AI for government that we can trust? Okay. So let's get right into it because this is the absolute heart of the problem. What do you do when an AI, which you've given a really important government job, just confidently invents information completely out of thin air? And this isn't just some random glitch. It actually has a name. It's called the hallucination problem. And honestly, it's pretty much the single biggest roadblock to using this amazing technology in high stakes places, like, you know, public service. So what exactly is an AI hallucination? Well, to put it simply, it's when the AI generates something that's totally false or has nothing to do with the documents it's supposed to be working from. It's not just getting something wrong, it's literally fabricating it. And this right here, this illustrates perfectly why it's such a massive deal. This is a real policy from Minnesota, and it makes it crystal clear. The responsibility doesn't land on the AI. Nope. It lands squarely on the government employee to catch those errors.

[8:46](https://youtu.be/_f8ll2doV4Y?t=526) At the end of the day, the human is still accountable. So how do we fix this? Well, there is an answer. And it's a really powerful approach called source grounded AI. And it represents a huge shift in how we think about using these tools. The easiest way to wrap your head around it is like this. You're essentially building a digital fence for the AI. You're telling it exactly where it's allowed to get its information and basically forbidding it from wandering off into the wild west of the internet. And the process is actually surprisingly simple. First, you give it trusted, vetted documents. Step two, you tell the AI to only use those sources for its answers. And then, and this is the most important part, you make it provide citations for everything it says. That simple loop means a person can quickly and easily check every single answer. Now, this isn't just some theory. Let's take a look at how this is being used with real tools that government agencies can get their hands on right now. So the key here is picking the right tool for the job. You'd use something like Google's Notebook LM when your starting point is a specific set of documents you already trust. On the other hand, you'd use something like ChatGPT when you're starting with a question and just want to explore some broader ideas. And for government work, this is absolutely crucial. Tools like Notebook LM and Microsoft 365 Copilot are specifically designed to plug right into an organization's internal data. This makes sure the AI is working from your agency's secure, vetted information, not just scraping the entire internet. Just think about the old way of doing things. You've got legislative staff facing a literal mountain of paper, spending days, maybe weeks, just trying to understand all the details of a new bill. Now imagine feeding that same 500 page document into a grounded AI and getting an accurate summary with citations in just minutes. That's a total game changer. Or picture this, a public information officer is on a super tight deadline. Instead of scrambling, digging through a bunch of different reports to make sure every fact is perfect, they can just give the approved background documents to the AI. It can then generate a first draft where all the key information is

[10:52](https://youtu.be/_f8ll2doV4Y?t=652) correct and cited right from the start. That frees them up to focus on crafting the perfect message. And check this out. This is a really high impact example coming out of Georgia. They're using this kind of AI to analyze their own data from public assistance programs. But keeping that analysis strictly grounded within their own data sets, the AI can spot these really subtle patterns that might point to fraud, which protects millions in taxpayer money without just making baseless guesses. So what's the bottom line here for the public sector? Why is this source grounded approach so important? Well, it all really comes down to one word, verification. This is the big takeaway. The goal for government isn't to have an AI that knows everything about everything. No, the goal is to have a tool that can reliably give you the correct answer from your own trusted information every single time. And when you get that right, the payoff is just enormous. You create accountability because every fact can be traced back to its source. You guarantee accuracy. You build public trust. And you do all of that while making your whole agency run more efficiently. So by keeping AI grounded in verified sources, we're not just preventing errors. We're actually building a foundation of trust. And taking this responsible path forward really raises a big question. Is this how we build the future of public service?

[12:11](https://youtu.be/_f8ll2doV4Y?t=731) Okay, so this is a good example. Now, that video did not come through by magic, right? It wasn't like I typed in, oh, do something about, you know, NOPIC-LM and public service. I had to do a lot of work to focus that video. Hopefully, you guys liked it based on the questions and things that I thought you might have. And then it mentioned these examples from Minnesota and from Georgia, right? And so what I did is I said, okay, well, someone's going to say, like, what are those sources? And so I went back. And, you know, you'll see here on the left, these are all the sources that were identified. And the checkmark means that I looked at the sources and I thought, And the checkmark means that I looked at the sources and I thought they were okay, right? So I could not check the sources and they would not upload into it. And then when the video played and I saw that they picked these sources, I thought, well, like, which out of all these sources have those? And you'll see here that, you see the little numbers there on the side? Those are sort of the citations that led them to pick those different sources.

[13:25](https://youtu.be/_f8ll2doV4Y?t=805) So this is just kind of to give you an example of how you can sort of ground everything and knowledge that you pick from a broader run. And then, you know, if questions come up and you're trying to figure like, where did that come from? Then you can get that information. So it's important to understand that, you know, all these AI tools all serve a purpose. I use all of them. The only one I'm not a big fan of is Copilot, but I include Copilot here because I know a lot of you are in Copilot environments, but I keep trying to use it and it never works well for me. But, you know, with No Pickle Limits, you know, remember I talked about having an intern or whatever. It is like a research assistant. Like, you know, I've had in my career a million, and I'm not saying, definitely hundreds of interns, postdocs, things like that who've worked for me. And, you know, they're really good in terms of being thoughtful for synthesizing insights and comparing claims, all these different things. I can tell them something, they would go off and do it and see what happens. But when I'm doing something related to communication, like say a presentation or like just before this, I was working on a workshop that I'm helping an organization organize here in Pittsburgh, I go to ChatGPT. No Pickle Limit would not help me for that particular task, except for maybe creating an infographic after I've done everything in ChatGPT.

[14:45](https://youtu.be/_f8ll2doV4Y?t=885) And then with Microsoft Copilot, the big advantage there is that, you know, when you have office tasks, things that are relatively automated, or you're trying to summarize meetings, you have data that can only be within like your own enterprise environment. The Microsoft Copilot is there. The one thing that is true across all of these is that obviously you don't want to have any sensitive data and you need to have that human in the loop verifying all the outputs against the source records. So here's some, this is just like an overview of what you can do, like if you don't, you can't use Notepad LM or you have these other different options. So I'll talk about that option later, but right now I'm going to talk about, in Notepad LM, the option of using your own source. So in this case, this is what Notepad LM looks like, and this is the point, by the way, if you have your own document, go ahead and log into it and upload your document.

[15:47](https://youtu.be/_f8ll2doV4Y?t=947) So what you want to do is, you see here it has create a new notebook, and what I have done is I have uploaded a document here. You'll see here, Reimagine Appalachia Catching Heat Final. That is an analysis that I did for an organization called Reimagine Appalachia. It's about capturing waste heat from AI data centers and using it for some sort of community It's an original analysis, and I'm assuming a lot of you don't know about it, so I'm going benefit. to be using it as a demonstration as to how you can do these things. And when you upload these things, it will then produce a brief summary, and you can read over the summary in this middle area to say, okay, does this make sense to you? So that's a first step in terms of what it is that you do, and again, I'm hoping you guys will follow along with your own document as I'm talking. So on the right-hand side, you're going to see something called Studio. Studio are all the different things, outputs that you can get from Notepad LM.

[16:50](https://youtu.be/_f8ll2doV4Y?t=1010) So you'll see here audio, slide decks, videos, mind maps, reports, flashcards, quizzes, infographics, and data tables. The little arrow that you see there on the right, that is a way to refocus things. When I first started using this, I would just click infographic, and it would produce one, but it wasn't necessarily focused the way that I wanted it to be focused. And so I learned over time that if I click the arrow, you'll see that it gives you a lot more options as to what it is that you can do, but otherwise, it's just like click Also, the other thing to know is that for data tables, it does both qualitative and and play. quantitative work, and so I'm going to show you an example of qualitative analysis later. The other thing I like to mention is that it doesn't always listen to what I say despite me giving instructions. Like for example, you'll see some of these videos, I asked for a video that was one minute, and it instead produced one that was two minutes. So it doesn't always listen to what it is that you say.

[17:53](https://youtu.be/_f8ll2doV4Y?t=1073) Also, every time that you produce a product, you can't like easily edit it, oftentimes you just have to regenerate it. So I'll talk about that too as we go along. So one way to start is to generate a mind map. So you take your document. So again, this is this one Appalachian waste heat, and it took, you know, what I wrote, which is about an 80-page document, and it broke it up into parts. These parts are not, by the way, the same as the table of contents for the report. And then you click this arrow, and you get policy options, and then it broke it into these three categories of policy options, all of which is correct. Once you do that, then there's ways that you can use it to help you do your rest of your So you can, you know, like you can get a visual index for a super long document. notebook LM work. You can identify gaps, you know, knowledge gaps, and then try to figure out what's happening You can pick one area, like I picked policy for this presentation, and I'm going to focus there.

[19:03](https://youtu.be/_f8ll2doV4Y?t=1143) on that branch of what's there. And then you'll see here that you can use those branches to customize the videos, refine all your research tools, and just kind of pick the nodes that you'd like to use. So oftentimes starting out with a mind map, particularly for a super long document, which is often the case in policy, is a good starting point for you. So here's one of the other options. This is like creating a report. So this is an example of what happens when you like click that little arrow on the right. So the first part, format, these are the options that you can pick, like create your own, briefing guide, study guide, blog post, okay. But then what's nice is that you'll see this little like magic thing here where it says suggested format. So it has now suggested to me what things I might want to consider. Just kind of like oftentimes with the other tools, it'll say, oh, well, do you have this question or that question? This is kind of the same thing. So for example, this one says, oh, community impact primer.

[20:04](https://youtu.be/_f8ll2doV4Y?t=1204) And I thought, oh, that sounds pretty good. Anything different than what I've already done on this particular report. And then when you do it, it generates something like this. This is like an overview of it, the challenge and opportunity why it matters. This by the way, goes on for, I don't know, 10 pages or something. I pick this one as an example five, which is like learning from leaders, case studies. And the reason why I mentioned it is because it picked these examples from all across my document. Again, they're not just repeating what it is that I say. I put examples throughout and then it picked these examples. Now I did not like the examples it picked and I'll explain that a little bit later. And so if I was doing this for real, I would modify it, make suggestions, say, hey, this is too long. I only want it to be like 500 words or less, things like that. And you can do all that within that screen I just showed you prior to this.

[21:06](https://youtu.be/_f8ll2doV4Y?t=1266) So one of my favorite things that it does is create infographics. So this is, I'm going to show you some of the kind of iterations that we have here. So this is like when I just like hit the infographic button and I didn't do any refining at all. And you can see here, we have the opportunity, it puts in some numbers here. It shows the paths for Appalachia and it gives some numbers. And then also these examples here. The reason I don't, there's a couple of reasons I don't like this infographic. It's not perfect, at least in my mind. The first part is that you see that it's red. And the reason it's red is because it says heat, that's part of the title. But in reality, what comes out of AI data centers is steam. Steam is white. Okay, well, this is not helping me. And then sometimes these numbers, they're close, but not quite right. So I do have to sometimes go in and edit them.

[22:07](https://youtu.be/_f8ll2doV4Y?t=1327) And in not this webinar, but in future webinars, I'm going to show you how you can edit these documents and try to fix different things where you kind of like the figure, but don't like the figure. But in general, if you don't like it and you have changes, you have to regenerate the entire thing. So you really need to get it super close to what it is that you like. So what I don't like about this one is that it does not focus on examples in Appalachia. It has Seattle, Washington, Quebec, and Sweden. So I said, hey, I want you to focus on examples in Appalachia, right? Because that's my audience is people in Appalachia. I have those other examples there to show that this works, but I don't want it for my Okay, so here it generated something. communication. And okay, I did not fix the red heat problem, but it is focused now on Appalachia. It talks about projects in Ohio and West Virginia. It does talk about this thing about unlocking the grid and so forth.

[23:09](https://youtu.be/_f8ll2doV4Y?t=1389) So it's not perfect, but it's pretty good. I think probably for this, I would take out like it has this, the same thing above and below the figures. I probably would take out the step below the figures, things like that, but it's still not good enough for me. So then I go to like option three. Okay, here I said, I want you to focus more on policies, right? Not just focus on like how it works. So again, I still haven't fixed the red problem, but it does do pretty good in the middle part saying, okay, here's the two things it can be used for, community residential heating, ag and industrial. And then it did add my policy things that I added, clearing grid roadblocks, talking about gains in terms of energy efficiency. And it talks about the policy toolkit. So it's getting closer to what it is that I want. And then this is the one where I pretty much got everything into what it is that I like.

[24:11](https://youtu.be/_f8ll2doV4Y?t=1451) And by the way, you notice that this is horizontal. Oftentimes the default is vertical. Vertical doesn't work very well for slide decks. So you have to tell it that you want things horizontal and things like that. So this was pretty good. First I told it to change it, say, hey, this is steam, this is not fire, make this white. So it changed everything to white. It does do the policy categories, talk about like implementation. It does talk about things that are specific to Appalachia and covers all the policy options and things that I want. If I was going to do another iteration, I don't know why it put tax fiscal incentives over here when it should be over here with strategic policy categories. So I might say, well, this graphic is pretty good, but put all the policy categories together. But the thing is, it's going to regenerate the entire thing. So just know that that is something that can both drive you crazy, but shows that you can get it to do something different depending on what instructions that you provide.

[25:14](https://youtu.be/_f8ll2doV4Y?t=1514) Okay, so the next thing is a data table. So what I did is I said, develop a table summarizing the policy options and the pros and cons of each. And then I thought it did a pretty good job. Now these four E's for anybody who's taken any of my workshops within a bit years before are what I teach, it's what's in my book. They're effectiveness, efficiency, equity, ease, and political acceptability. So it has a policy options, the categories, and then the pros and cons for each. And you'll see the sources. The source in this case is just one, and that is just because it's only using the document that I wrote. But overall, I thought this was pretty good. I have a table somewhat like this in the very back of the appendix of the whole report, but it did a good job of doing a balanced analysis and providing it in a chart that When you get this, by the way, you can download this into like Excel or whatever your favorite can be used.

[26:14](https://youtu.be/_f8ll2doV4Y?t=1574) tool is, and you can manipulate it there. Okay, so the next thing, another thing you can do is slide decks. So with slide decks, I found you really have to be quite careful. So you can have a detailed slide deck or just presenter slides. You can have like short, that's what I picked for this one. And then you can just like describe the deck you want to include is really important. Like I did this one twice. I'm not going to show you the first one, but it was very gray. And so I gave it instructions to say, make this very colorful for users. And you'll see here, this is it. It made it colorful. And again, on the top right, you see, you can download this as a PDF and download as Then you can bring it over and edit it there. a PPT. Now there are some challenges with editing. I'll describe that in future webinars, but just so you kind of know what it looks like. But you'll see here, it did a pretty good job. It includes numbers, shows how things work. And I recognize these from some of the diagrams and pictures that I have in the report.

[27:20](https://youtu.be/_f8ll2doV4Y?t=1640) Talk about these two pathways, community thermal, industrial, and ag. Has these perfect concepts. Again, it still has these things all over the place. So I would fix this and tell it to just maybe have one from Appalachia, depending on the audience that I was talking to. And the thing I just mentioned, this thing is from something that's done at the National Renewable Energy Lab. And it was a box in the report, and it did a pretty good thing for the box that I had. Talked about, again, having some of the numbers in here, the roadblocks to adoption. But again, it has icons twice, so I probably would edit out one of those. And this is the policy playbook for Appalachia. This is it took my four E stuff and did a nice little graphic here based on the pros And then it has a nice sort of concluding kind of slide. and cons.

[28:24](https://youtu.be/_f8ll2doV4Y?t=1704) So overall, pretty good job. Again, not perfect, but close enough that I could edit and definitely far better than I could do on my own using a slide deck, particularly as I might be good at analysis, but I would fail any sort of class related to graphics. Okay. So the next thing is audio. So this is so you can do like your own activity. So here you can do deep dives, briefs, critiques, and debates. So I put explain the policy options in the report in 30 seconds. So I gave it an idea. So the first one, which is what I call solo, and I'm going to play it so you can just kind of hear what the explanation is. But what I think is a little bit more fun is the debate, depending on your context. Like if this is just for the public and you just want to say, oh, this is what this analysis report rule or whatever is about, you probably want the solo option. But if you're like me and you love debates and you want to show that you do both sides, then you're going to do the debate option.

[29:33](https://youtu.be/_f8ll2doV4Y?t=1773) So I'm going to play a bit of this. This is the brief on AI data center waste heat policy options. AI data centers generate massive amounts of leftover heat. And rather than just venting it up into the sky, governments can turn this waste into a heating resource for nearby communities. But it takes some really intentional policy action. To actually turn that heat into local community heating, policymakers basically have three progressive levers to pull. Clearing the rules, assisting the planning, and offsetting the costs. First, regulatory streamlining. Think about building an addition on your house, right? Before you can even dream of pouring the foundation, you've got to cut through a maze of red tape and zoning laws. Data centers face that exact same headache. So, governments can step in and reduce that timeline uncertainty with one-stop permitting and pre-certifying sites with standard access conditions, which totally removes the guesswork for developers. Second, project development support. Now you might be wondering, even with permits, how do you figure out the logistics of connecting a data center to a town without going completely bankrupt on engineering studies?

[30:36](https://youtu.be/_f8ll2doV4Y?t=1836) Well, policymakers can de-risk those early stages by providing grants for front-end engineering design or by funding county-level maps to identify thermal-ready infrastructure corridors. Finally, fiscal incentives. Look, even with blueprints in hand, building this infrastructure is incredibly expensive. To tackle those massive upfront capital costs, governments can offer refundable investment tax credits for thermal networks, exempt sales tax on the equipment, or even give local property tax abatements tied directly to verified heat delivery performance. By streamlining regulations, supporting development, and offering fiscal incentives, policymakers can transform AI data centers from massive energy drains into anchors of resilient regional energy systems. This is the brief on AI data center waste heat policy options. Welcome to the debate. When we picture, you know, the cloud, we usually imagine something totally invisible. Oops, hold on a second. There we go. Oh, back, back. Okay, so that was the solo version.

[31:38](https://youtu.be/_f8ll2doV4Y?t=1898) You can see it was like all pro. It didn't have any analysis, but it did a good job of connecting the analysis, you know, the challenge to the policy option. So here is the debate option. Here we go. Welcome to the debate. When we picture, you know, the cloud, we usually imagine something totally invisible. Right, it's not. No, exactly. If you step inside a new AI data center, say out in Appalachia, you hit this physical wall of heat. I mean, it's millions of microchips acting like tiny, relentless engines. Yeah, and to keep those engines from just, you know, we use massive cooling systems that essentially, completely melting down, well, they just take all that heat and throw it away into the atmosphere. Exactly. We're just venting it out. So today we're looking at that colossal amount of untapped waste heat. And the big question is, should states step in and mandate that these new builds capture and reuse this heat, or should they rely purely on incentives?

[32:40](https://youtu.be/_f8ll2doV4Y?t=1960) And critically, you know, who writes the check for the plumbing to actually move that heat? Is it the tech developers or local governments? Right. So my stance here is that states need strict regulations for new builds. And I also think local governments should chip in on the infrastructure to secure those community benefits. And I take the opposite view. I think rigid mandates will just chase these highly capitalized tech companies out of town. States should use voluntary incentives. And honestly, the developers themselves need to foot the bill for the entire infrastructure. Okay. Well, let's start with the mandates. I mean, we really can't afford to just wait and hope companies do this voluntarily because the physics of retrofitting are a total nightmare. They're terrible. Right? Most existing data centers use air-cooled systems, meaning they just blow warm, diffused air around. Trying to capture and transport that later is, well, it's like trying to catch wind in a net. It's incredibly inefficient. That is ready-made to be piped directly into a local heating system.

[33:44](https://youtu.be/_f8ll2doV4Y?t=2024) Right. That makes sense. That's why baseline rules are necessary. I mean, look at Germany. They're requiring new data centers to reuse 10% of their energy by 2026. And New Jersey's proposed bill, S4043, is looking at similar energy use plans. We need that plumbing built in from the start. But look at what's already happening without those heavy-handed rules. I mean, this is a fiercely competitive market, right? If you drop a strict mandate on a developer, they'll simply pack up and build in a neighboring state that doesn't have those rules. You really think they'd just leave? Absolutely. You don't need a mandate when you have flexible, outcome-based incentives. Okay, so that kind of gives you an idea of what the debate form is like. It's very, I don't know, NPR-ish, I would say, in terms of doing it. But I would say, again, as the author of this thing, I thought it did a pretty good job, you know, debating the issue. And in some places, that would be very useful to have.

[34:44](https://youtu.be/_f8ll2doV4Y?t=2084) Okay, so video. I've already showed you, like, one video. So I've asked for a 30-second brief overview of the policy options, you know, from the report. Of course, it did not do 30 seconds. It did longer. But I'm going to show this to you again so you can just sort of see how it's presented in a video. You know, have you ever stopped to think about where the internet actually, actually lives? I mean, it's not just in the cloud, right? It's in these massive buildings called data centers. And wow, do they generate a ton of heat. But what if, what if we could actually turn all that waste heat into something useful for our communities? It really leads to this pretty wild question. As our digital lives just keep expanding, could the data we're creating literally start heating our homes and businesses? Okay, so to really get how this could work, we first need to understand where all this heat is even coming from. Think about your phone after you've been scrolling or gaming for a while. It gets pretty warm, right? Now, just imagine that on an absolutely massive industrial scale.

[35:48](https://youtu.be/_f8ll2doV4Y?t=2148) That's a data center. They're the engines of our digital world and they get so hot that basically all the electricity they pull from the grid gets turned into heat. So here's the big idea. What if we just stopped calling it waste heat? What if we started treating all that energy as a brand new resource? The process itself is actually pretty clever and straightforward. First, you've got all that intense heat coming off the computer servers and it's used to warm up a liquid. But that water isn't quite hot enough. So then a heat pump acts like a booster, cranking up the temperature to make it really useful. From there, that hot water is just piped over to nearby buildings, homes, offices, even greenhouses. It's basically a giant recycling system for energy. It sounds like a perfect solution, doesn't it? A total win-win. Well, it's not quite that simple. There are some huge opportunities here, for sure. But there are also some pretty big challenges we'd have to figure out. So on the one hand, the potential is massive. We're talking about cutting carbon emissions and creating new ways for data centers to make money.

[36:52](https://youtu.be/_f8ll2doV4Y?t=2212) But on the other hand, the hurdles are very real. The upfront cost to build this stuff is high, you need a whole new network of pipes, and it only really works if the people who need the heat are located pretty close by. And that brings us to the most important point. How do we actually tip the scales and make this a reality more often? So, how do we make it happen? Well, it really comes down to having smart policies. Governments can help smooth the way by streamlining the rules, offering support to get these complex projects off the ground, and providing financial incentives to make it all worthwhile. Which leaves us with this final, really interesting thought. Could the very buildings that power our digital lives also become the power plants that heat our actual neighborhoods? It's definitely something to think about. Okay. Okay, so that kind of gives you an overview of all the different ways that you can use Notebook LM. Here's how you can do it, but not when it's not from a single source, but is from internet sources.

[37:57](https://youtu.be/_f8ll2doV4Y?t=2277) So, as I sort of mentioned, this is, so this one I asked, what are the challenges of using source-based AI like Notebook LM? And you'll see here on the left, here's all the sources that identified. Again, I could click these on or off, and then you click where it says import at the bottom, you just hit import. I've noticed that it only does 10 sources most of the time, so I think you'd have to prompt it for more. And I believe there's also a maximum of 40 or 50 sources, something like that, that you can use. So that's just, you know, sort of something to think about. But this is a topic where I did want to go to the web, but I did want to know like where everything was coming from. And then I asked it to create an audio and video version for basically what are the challenges and what are the common mistakes that people make. It then does a summary. I then go to the video overview, and I say develop a one-minute video that discusses the challenges in using Notebook LM for state and local government policy work.

[39:00](https://youtu.be/_f8ll2doV4Y?t=2340) Oh, I think I missed the video itself. Well, you know, I know what happened. No, I think I put it in the end. So you'll see that video at the end, but that kind of gives you an idea. Okay, option three is mimicking the source-generated AI and other Gen I tools. Okay, so this is a video, you know, that talks about, you know, ChatGPT and also Copilot. Notebook LM guarantees accuracy because its architecture physically restricts the AI to only the documents you provide. But everyday tools like ChatGPT and Copilot pull from their vast training memory, meaning they naturally guess and blend facts when they shouldn't. To fix this, we are going to build a six-step workflow, a process that overrides those default settings and forces your AI to pull and cite exact evidence.

[40:01](https://youtu.be/_f8ll2doV4Y?t=2401) To follow along, grab two or three sample PDFs, ideally ones with some overlapping information, and open a fresh chat window. We are explicitly trading the ease of a quick conversational prompt for absolute control over factual accuracy. Step one is uploading. When you drop your documents into the chat, you need to explicitly name each one in your prompt, like source A and source B. Giving the AI strict naming conventions provides it with a tracking framework. It establishes a clear tag the model must reference later when we demand to know exactly where an idea came from. This diagram illustrates a built-in shortcut for Microsoft 365 Copilot users. As you can see by the laptop connecting directly to specific folders, you can bypass manual uploads by prompting Copilot to strictly use documents located within a designated SharePoint or OneDrive folder. For everyone else, we move to step two. We use our first prompt. Extract all relevant evidence from the provided sources.

[41:04](https://youtu.be/_f8ll2doV4Y?t=2464) Do not use outside knowledge. If the answer is not in the sources, say, not in provided materials. We want the output as raw bullet points. We are deliberately stopping the AI from writing flowing paragraphs at this stage. By forcing the model to pull raw data first and interpret it later, we cut off its ability to invent the logical bridges that usually cause hallucinations. Now that we have the raw facts, step three is finding the friction. Prompt number two is identify where the sources disagree or present uncertainty. Do not reconcile the differences, just surface them. This illustration highlights a major risk called coherence bias. Looking at panel four, we start by trying to force a puzzle piece into place. Left to its own devices, AI naturally smooths over contradictions, which hides vital discrepancies in your data. With our prompt, the AI stops trying to solve the puzzle and instead explicitly lists the exact points where source A and source B disagree.

[42:05](https://youtu.be/_f8ll2doV4Y?t=2525) Step four is synthesis. We run our third prompt. For each conclusion, provide a direct quote from the source that supports it, then summarize. This step, before claim formatting, forces the AI to show its work, creating a highly traceable audit trail for every single point it makes. You no longer have to blindly trust the output. Forcing explicit citations turns the AI into a transparent research assistant, providing the exact receipts for its conclusions. Step five exposes the limits of your documents. list what is not known Prompt four asks, or not addressed in these sources that would be critical for decision making. If you skip this step, the model will often invent plausible-sounding statistics to fill any empty spaces in the narrative. Finally, as this graphic of a person cross-referencing paper documents shows, human verification remains mandatory. We lack hard architectural constraints. Looking at our technical chart's failure points, the AI might generate weak citations, hallucinate page numbers, or drop context entirely if the session resets.

[43:07](https://youtu.be/_f8ll2doV4Y?t=2587) Prompts can get you 90% of the way there, but your manual review is the only absolute guarantee against source contamination. Look at the difference. On the left is the vague, unverified output of a standard query. On the right, the result of our six-step chain, a structured, explicitly footnoted report tied directly to your data. You sacrifice a bit of conversational flexibility and speed, but the return is a massive increase Take a moment to copy this exact four-prompt chain in enterprise-grade reliability. into a local text editor or Notes app right now, so it's ready for your next session. Mastering this structured approach removes the guesswork. It gives you the confidence to rely on AI for rigorous, fact-critical workflows, knowing the evidence is right there on the page. Okay. So I'm not going to go through this, but for people who are like, okay, the video's not going to work for me. I want something different. So I asked it to generate a very simple for users slide deck.

[44:11](https://youtu.be/_f8ll2doV4Y?t=2651) And you'll see here it goes through sort of the same thing, talks about what the prompts are. In this case, it's probably easier as a source to look at the slide deck, but watching the video is also a good way to do it. And like I said, it uses good analogies like this, expose the blind spots and use the receipts and things of that nature. So sometimes I would use both of them. So let me, I'm running out of time here, so I'm going to go to the challenges. So here's a data graphic. Now, this gives you a good example of, like, how you use multiple tools. When I first did this, I tried to use, to generate an analysis of the challenges of using Notepik LM. I did not like the result. So instead I went over to chat GPT.

[45:12](https://youtu.be/_f8ll2doV4Y?t=2712) I generated the content over there so that it met what I thought was the right level of knowledge, you know, like what you guys would like to have. Then I brought it back to Notepik LM and then I asked Notepik LM to create the infographics. Cause I do, as I said, I do love the infographics option in Notepik LM. And you'll see here, so let me go through these. So some of the data silos hinder insights. I have come across this as a challenge in that if let's, you know, you can't like go into a dialogue with it, like chat GPT. And I've done a previous webinar for Innovate US on how to use chat GPT for dialogues and really discuss issues. The problem is, is that like, just say you have question one and you uploaded all those sources, but now you think, Oh, well I have question two. Well you can't just ask question two. You have to go back and add in more sources to answer question two.

[46:19](https://youtu.be/_f8ll2doV4Y?t=2779) Well in chat GPT it'll, it'll do that for you automatically. Right. The second thing is that you still can have hallucinations despite what it says. This is 17 to 33% error rate. This is just like one case. It's like legal stuff. Okay. But again, it emphasizes you need that human in the loop thing because you know, you gotta make sure that those numbers are correct. You can have this interpretive drift where it's overconfident, it's unsupported. It does kind of do universal facts. It doesn't do necessarily deep reading. Like it, you know, like if you have a 900 page document, you know, again, it's just like a student. Just always think about like this intern or something like that. I don't have time to read 900 pages. I'm only going to read the beginning and the, and like a little bit in the middle and the end. Well, it will do that for sometimes.

[47:19](https://youtu.be/_f8ll2doV4Y?t=2839) And then there is this other thing, which is, which is called rag poisoning because our rag stands for retrieval, augmentation and generation overload. And if you have somebody malicious, then they can go and put something opposite to, to what it is that you from your source and get it, the AI to believe it. Okay. So that's something out there that there's, there's evil people out there who, who will try to poison your data and you need to watch out for that. And again, that's why you had this human loop. This doesn't make any sense to me. So this is again, things that you can do. So it's going to answers note again, it can do 50 documents at once, about 500,000 words, trust the citations, not just the text, pretend says information and optimize the files for better accuracy. This is, I think getting towards my final thing here, but I wanted to welcome back. So today we are diving into something, well, something pretty revolutionary in the world of AI.

[48:32](https://youtu.be/_f8ll2doV4Y?t=2912) We're not talking about your everyday chat bot. We're talking about a new kind of AI that's designed to be more truthful, more reliable because every single answer it gives is grounded in sources you provide. Sounds like the perfect fix, right? But as you're about to see, this solution comes with a brand new kind of hidden danger. So let's just start with the big one. Is there an AI that can't lie? I mean, we've all seen the headlines, right? AI, just making stuff up, stating things as fact that are completely wrong. This whole issue has been one of the biggest roadblocks to really trusting these incredibly powerful tools. Yeah. The official term for this is hallucination and it happens because of how these models are built at their core. They're expert word predictors. They're just trying to figure out the next most likely word in a sentence, not trying to state a verified fact and that design, well, it can lead them to generate stuff that's totally made up, which is, you know, a massive problem. But what if we could actually fix that? Okay. Oops. I think I also did the wrong video there. Now that I think about it, hold on. I think I, I think I put the, the, there we go. Oh no. I just want to do this. How RAG works. I think I did.

[49:40](https://youtu.be/_f8ll2doV4Y?t=2980) I haven't done that. Okay. Well name RAG. So let's take a second and break down how this thing actually works. So RAG it stands for retrieval augmented generation. And the best way to think about it is like this. Instead of asking AI to answer a question from its own giant jumbled memory, you're giving it an open book exam. It has to first go and retrieve the relevant facts from a set of books. You've already approved. Those are your sources. And then it generates the answer based only on what it found in those books. The process itself is actually really simple. This is where our story takes a bit of a dark turn. This is where we learn how to poison the well, because that core strength, the fact that it relies on sources, that is also its single greatest vulnerability. What happens if the sources themselves can't be trusted? This brings us to a whole new kind of threat called knowledge base poisoning. Now an attacker doesn't even try to hack the AI model. No, instead they attack the knowledge. They slip malicious, totally fake documents into the collection of sources the AI relies on.

[50:47](https://youtu.be/_f8ll2doV4Y?t=3047) Basically they poison the textbook for the open book exam. So how would an attack like this actually work in the real world? Let's walk through the anatomy of one because it's a lot simpler and frankly, a lot more terrifying than you might think. This image from a research paper shows it perfectly. Okay. Look at the left. A user asks where the movie Manchester by the sea was filmed. The rag system correctly pulls up facts that Manchester is in Massachusetts, which is in the U S and it gives the right answer. Simple enough. But now look at the right. An attacker adds one tiny piece of false info, a perturbed edge as they call it, connecting Manchester to England. Now when the user asks the exact same question, the system pulls that new poisoned fact and the AI confidently and completely incorrectly says the movie was filmed in the United Kingdom. And here's the crucial part. This doesn't require some sophisticated hacking genius. The attacker doesn't need to know anything about machine learning. They just need to be able to upload a document. That's it. They just add a few fake files that look real enough and then they just wait. So here's a much more concrete and scary example from security researcher, Amin Raji.

[51:53](https://youtu.be/_f8ll2doV4Y?t=3113) Imagine your company's AI has a legitimate financial report showing great revenue and attacker uploads a new document. They frame it as a CFO approved correction, but inside are completely fabricated numbers showing a massive revenue crash. And this kind of attack is just devastatingly effective for three reasons. Okay. I want to stop there cause I'm running out of time and I do want to have a few questions. Uh, I just want to remind you that there's still time. Uh, this is just an overview. There are two more sessions after this, you know, the next couple of Fridays. Um, and then there's also a lab at the end. Um, and then I'll just close here. This is all the information, um, about me. These are newsletters I do on LinkedIn. Uh, if you want to connect with me on LinkedIn, uh, you can go there. Um, and then otherwise I'm going to stop sharing here. Uh, you will get, um, the video, you know, like the, the video of this webinar, you'll get the slide deck that I just showed you guys. Um, and I hope you will join me for these future sessions, but hopefully I have time for just a few more questions before in the five minutes we got left.

**Agueda Quiroga**

[53:06](https://youtu.be/_f8ll2doV4Y?t=3186) Okay. David, do you want me to read the questions from the Q&A?

**Deborah Stine**

[53:08](https://youtu.be/_f8ll2doV4Y?t=3188) Yes, go ahead. I, I've been trying to follow a bit as I go along, but it's kind of hard.

**Agueda Quiroga**

[53:12](https://youtu.be/_f8ll2doV4Y?t=3192) Okay. I mean, there are six questions. So I, I'm trying to see what are really, really simple ones. Can the AI voice be changed?

**Deborah Stine**

[53:22](https://youtu.be/_f8ll2doV4Y?t=3202) Yes. Yeah. So, um, the reason the voice sounded the same is I did not bother to change the voice, but there are, you know, those options, like when you click the little arrow on the side, there are ways to change your voices. And when I do the future ones, where I only talk about video and audio and so forth, I'll show you those different options. And also, by the way, the same thing for everything else that you saw, I just stuck with the same style of, um, you know, everything. Uh, but you know, you can, you can, you definitely have options that you can pick.

**Agueda Quiroga**

[53:52](https://youtu.be/_f8ll2doV4Y?t=3232) Okay. Another question from Melissa, Melika, sorry. To combat hallucinations, it sounds like the recommended approach is to use separate prompts for the AI tool, for example, rather than one prompt with AI instruction. Is that accurate?

**Deborah Stine**

[54:10](https://youtu.be/_f8ll2doV4Y?t=3250) Um, I'm not quite sure what you mean, but I will tell you, let me tell you how I do it. Okay. I do a couple of things. So the first is I picked the right tool in the first place, right? If I'm doing analysis like that, the one I showed as a demonstration, I do not use chat GPT or anything else. I use perplexity. And the reason I use perplexity is because perplexity goes to it directly links like each sentence to the source. I click on the source. And then I verify information before I then take that information and put it in a new document that I am writing. Okay. So that's like fundamentally the very first thing, even then though, like when I'm done writing, whatever it is, cause I will use a variety of sources. I will often have tools check each other. I was doing a report for like another client, which is for a federal agency. And what I did is I said, I copied the texts that I wrote. I went and I put it, uh, in this case, I put it into Gemini cause I found that I liked Gemini for fact checking because it has the most recent data, uh, because it's great.

[55:20](https://youtu.be/_f8ll2doV4Y?t=3320) Google does everything in the entire web. And I will say, is this correct? Um, tell me, you know, any nuances and it will tell me and, and I'll say it's correct. And I said, well, can you give me some sources to show me it's correct? And then I check those sources. So there's lots of ways to do fact checking. Uh, but oftentimes I will put the tools against each other.

**Agueda Quiroga**

[55:42](https://youtu.be/_f8ll2doV4Y?t=3342) Okay. I think we have time for one more question. Not sure about this, but I think this is an important one. Uh, it's like Jen say how does not because I am with accessibility for people with visual disabilities, does it meet any standard guidelines? Not sure if you know this, but yeah.

**Deborah Stine**

[56:00](https://youtu.be/_f8ll2doV4Y?t=3360) So the answer is I do not know this, but what I would do is I would go to Gemini and I would ask that question. And, um, it's not something that I have looked at, but I do know that there are like, there are lots of hidden tools that you can use that makes sense. And part of the reason I like notebook LM is because as you know, people learn in different ways. Um, and so like audio might be better for some people, uh, while visuals might be better for somebody else. Right. Um, and so I think it all kind of works out. Um, I, let me add, let me just quickly add some of these other questions that I know that I'm on here. Um, Paul asks how precise are the reference locations, a case we want to double check. They're very precise. You just click on the link. You can go to the source. If the link does not work, which I've had happen and say, chat GPT, then that means it faked it. And, you know, you gotta find some other way to see that information is, um, is, uh, correct. Um, this other question, how can a government agency set up notebook LM for its own internal use?

[57:04](https://youtu.be/_f8ll2doV4Y?t=3424) I have no idea, but I do think, I do know that I see ads for like Google for government or something like that, uh, G suite tools for government. So I think just like with Microsoft, you can create your own sweet thing, but that's, there's other people in the innovate us world that are better answers to that than me. Can you use notebook LM to create a quiz? Yes. Uh, I did use do this on, again, I didn't, just didn't have time to show it today. Uh, but it did generate a quiz. Uh, and there's also like tools that will let you download the quiz so that you can use it. That, that are like, um, um, add ons. Um, Erwin asks, well, because AI memory is more capable than human will they have the skills to do meaningful analysis, uh, complex better. Okay. The answer is not at this time. And I do a lot of analysis. And of course I'm always worried about getting a job. And by the way, I also think graphics people, they still have a future too, because like when I work with these tools, a lot of the time it's just for me.

[58:05](https://youtu.be/_f8ll2doV4Y?t=3485) And then I will generate like five versions of the graphic. I did that actually for this, for this data center. One, I sent to the client, they picked out the one they liked, they sent to their own graphic artists and put it in their style. But the hard part is I need somebody to generate off of the graphic artist needs something to work with. Um, and then I think it makes sense. Um, if you're Melissa asked if you use more than 50, what if you do research and more than 50 documents, can it do that? If you upgrade to a higher planet, my answer at this point, I think is no, but those things are changing all the time. We're at three o'clock. And so I don't have to stop.

**Agueda Quiroga**

[58:41](https://youtu.be/_f8ll2doV4Y?t=3521) Thank you so much. That really has been amazing. So I hope to see everybody soon next Friday for part two. Okay. Thank you so much.

**Deborah Stine**

[58:50](https://youtu.be/_f8ll2doV4Y?t=3530) Thanks for attending everybody. Hope to see you next week.

**Agueda Quiroga**

[58:53](https://youtu.be/_f8ll2doV4Y?t=3533) Bye. I'm closing the session now. Bye-bye.

* * *

Original page: https://innovate-us.org/getting-started-with-source-grounded-ai · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org