# Q&A with Charlie Winter and Stephanie Rotolo, ExTrac > Source: > Published: 2026-09-11 13:00:20+00:00 **Emerging Technology Series Episode 1: Open-Source Intelligence with ExTrac** *Open-source intelligence has moved from a secondary discipline to a core one, and the advantage now goes to whoever can turn public data into a decision fastest. The catch is that it has to be the right data, because bad data plus faster systems only produces faster bad decisions.* *That is the argument Charlie Winter and Stephanie Rotolo make in the first episode of IWI's Emerging Technology Series, a three-part look at how open-source intelligence, ground autonomy, and unmanned aircraft work together on the battlefield and what that means for joint force readiness. Both work at ExTrac (extrac.ai), which gives decision-makers facing crisis, conflict, and geopolitical risk foresight they can evidence, defend, and act on.* *Join IWI's co-hosts Kristina Kempkey and Curry Wright to learn more.* Questions or ideas for future content? Contact us at*EmergingTech@IrregularWarfare.org* Note: This version has received an initial publication-focused editorial pass emphasizing readability while preserving speaker intent. Irregular Warfare Initiative **Kristina Kempkey:** Welcome to the Irregular Warfare Initiative's Emerging Technology YouTube Series. I'm Kristina Kempkey, Co-Director of IWI's Emerging Technology Focus Area. **Curry Wright:** And I'm Curry Wright, an advisor for the Emerging Technology Focus Area.  **Kristina Kempkey:** This episode is the first in a series on how the changing nature of warfare requires a layered and integrated approach across technology solutions. Over the course of three conversations, we will be joined by practitioners who build and field this technology. We go past the individual platforms to see how open-source intelligence, ground autonomy, and unmanned aircraft work together on the battlefield. And we look at what that means for joint force readiness in the future. **Curry Wright:** That’s right- no single capability is decisive on its own. Our conversation today will focus on open-source intelligence, and how a company delivers the right information to the right person at the right time to make better decisions faster.  Today we're joined by Charlie Winter, Chief Research Officer, and Stephanie Rotolo, Head of Customer Success, North America from Extrac. Extrac’s mission is to combine human expertise, high-quality data, and domain-specific AI to enable faster, more effective decision-making in the OSINT environment. **Kristina Kempkey:** Thanks, Curry. We hope you enjoy the conversation. If you have comments or questions, we'd love to hear from you. You can reach us at EmergingTech@IrregularWarfare.org. Over to you, Curry. **Curry Wright:** Charlie and Stephanie, welcome. Let’s start off the conversation with getting to know Extrac and its mission. What problem were you actually trying to solve when the company was started? What did you see in the field that convinced you a company had to be created to fix it? **Charlie Winter:** Hi, guys. Thank you so much for having us on the podcast. I think there's a bunch of stuff that led to us coming up with X-Track as an idea and then implementing it. But a lot of it was tied to work that I was doing when I was an academic about ten years ago, eight years ago. I used to spend way too much time hanging out in strange parts of the Internet that ISIS supporters were also hanging out in. And I essentially spent my time collecting data from those places on Telegram, RocketChat, TamTam, Element, weird platforms that people may not have heard of, as well as the more mainstream and obvious ones. And there was a huge amount of value to that data. So, in the academic work that I was doing at the time, I spent a lot of time working on information warfare, but also that same data, that same access was a great way into understanding how ISIS was innovating in its use of US technologies, how it was conceiving and implementing its doctrine for suicide tactics, what its defensive doctrine was. All of this stuff was in frame based on the data that I was putting in from ISIS supporters and operatives in Iraq, Syria and Africa. And for a while I was just doing that from an academic perspective, but then because of that academic work, I got looped in with some people who are working on the government and military side of the aisle in counter ISIS operations, and they essentially have been pulled into engaging with me because the data that I was working with was demonstrably useful. And one of those organizations that I was working with really likes the data. They're using it in a bunch of different programs, but they're also like questioning the time that it was taking for me to get in front of them. I mean, as an academic, I was working slowly as is an academic's want, but also, I was inhibited by the fact that I wasn't a machine. I couldn't pass 10,000 data points. I couldn't geolocate 10,000 data points in under an hour. It would take me a week or two weeks or three weeks. So the data was proven useful was getting to the decision maker and the people who are actually deploying those programs a few weeks or a few months after it needed to be so with this organization that's where the two other co-founders of extract are from we essentially built a an early version of what the software is today which was essentially an engine for collecting data from the same sources that I was putting data from and then also analyzing it so you could have that geospatial layer you can have graphing and visualization and network analysis and so on without needing to spend a week or two weeks in an excel spreadsheet and just fundamentally really worked it improved the targeting of the programs it improved the ability to understand what the adversary was doing saying how it was responding to external events how it was responding to internal events and so on and so forth and we realized that that model that structure that approach could be applied to other threat actors it wasn't just something that could be limited or should be limited to ISIS because adversaries wherever they are whoever they are whatever they are they'll use the internet in some shape or form they're all generating data in some shape or form and it was a question of taking that methodology of focusing on quality over quantity not trying to swallow up all of telegram or all of twitter or all of Facebook but just thinking about where this most relevant data could be found and then bringing that into an environment where it could be analyzed could be visualized and then translated into insight and forward-looking foresight to go to the operator and that's basically how we've evolved over the course of the last few years moving from being a bespoke kind of terrorism capability or CT capability and moving more into the state threats space into the organized crime space and also globalizing the coverage that we have and as we've done that as we've built out the product built out the solution built out the areas and problem sets that we're covering we've engaged increasingly with the us market and are getting a lot of traction there from a few different communities that are interested in leveraging Austin in a quicker smarter more kind of expertise and data-centric way and that's why we brought on Steph to work with our customers in the us and make sure that they can benefit from the software in the same way that we're our original customer base customer base over in the UK and NATO was with leveraging it sorry Steph do you want to jump in and introduce yourself **Stephanie Rotolo:** Thanks, Charlie. I'm Stephanie Rotolo, and I lead Customer Success for North America. Like Charlie said I am very intimately familiar with that environment because I am freshly out of it so just got an army most of that was in the special operations community but I was really drawn to extract and very much so aligned with the mission in that I saw what happens when the information is available somewhere but we're not getting that information to people on the ground because not too long ago I was the person on the ground trying to sort through all this information quickly so whether it was like sitting somewhere in a report it hadn't been disseminated it was buried in other reports it was trapped particularly by like classification procedures when it really maybe didn't need to be classified at that level it was rarely about the collection and more the delay between discovering something we needed and then getting that into the hands of the people who could actually do something about it and I’ve seen that extract can do that and I really wish I had had this because it would have made my life a lot easier but really what it is reducing the delay between that information and getting it to the hands of the decision maker makers the people that can action it I think especially important as we're in like the coalition space as well that minutes matter. You have to shorten the gap between the discovery and the decision, and that's what the platform aims to do. **Kristina Kempkey:** Thanks, Charlie and Steph. I want to pull a little bit on the point that you were talking about, about the evolution of your work and the demand for OSINT. But historically, the military has looked on OSINT as one of the lesser INTS. Can you explain for our audience perhaps why the thinking on this has changed and why it should be considered an equal value and importance to some of our other methods of data collection? **Stephanie Rotolo:** Yes, absolutely. I would say that 20 years ago, the battlefield wasn't also online. I think we have an opportunity now where we're seeing conflicts unfold in real time and everybody's able to access that, whether it's somebody on Telegram, TikTok, flight trackers, things like that, adversaries leaving these digital footprints. We just have more access to that stuff. Also, the commercial technology aspect of it is exploding. The government no longer has the monopoly on collecting information. And so that just expanded what is possible, how fast we can get it, how good the data might be. I also think there's an element of what we saw with OSINT in the Ukraine war really proved the value of OSINT as an int function. To be able to locate movements, verify attacks, counter disinformation. I mean, we're really seeing how the OSINT world directly expanded our understanding of battle space. And then also just that right now, speed really matters more than exclusivity. So, if an open-source report is available now, it might be more valid and relevant and necessary than waiting until that gets processed through the classified systems. And then again, the coalition's aspect of it. When I was at NATO, it was a constant choke point of trying to pass information from unclassed side to high side. But then you also have NATO classifications that you're having to learn and make sure you're following. That just slows everything down. You have this giant organization that's trying to be responsive and get things done. And then also, I think the big thing that's really pushed OSINT up to be a little bit more respected is that we have AI now, right? So, we have all of this enormous access to all of the data we could possibly have, but we're still just humans. And the ability to actually exploit that, process it and do what we need to do with it, it's just very difficult. But when you have these, this AI function that you can layer on top of it, now one person is basically doing the work of 15 analysts, which again, if the publicly available information OSINT is there and there is just as much of it as you could possibly want, you need something to help you do that in the absence of 15 analysts. I never had access to 15 analysts. My last team I had four, it was four people and trying to do that in countries where internet connectivity is challenging. I can't have classified spaces in certain areas. So, I relied a lot on OSINT, but that wasn't necessarily the right answer. So, I don't think that it's for debt where like suddenly discovering OSINT. I just think that the operating environment has changed drastically so the information available to the average person is so much more available that our process is now finally catching up to being able to fully process it and turn it into products at speed. **Kristina Kempke:** You mentioned speed several times. What does "speed of relevance" actually mean for someone operating at the tactical edge? How do we balance rapidly sharing information across partners and allies—such as NATO—while ensuring that information remains useful and actionable? Why does this matter? **Stephanie Rotolo:** Because the people that have the right information at the right time control the battle space, right? Like this is where we're no longer in like the tank versus the tank fight. So much of this is also in the cognitive space. So, it's not just moving data faster, but it's moving actionable information faster. So, it is the right information, right people, right time, appropriate confidence level, those types of things. That is the objective is towards decision advantage. So, it's not going to be whoever collects the most information. It's going to be whoever is able to at speed, turn information into action and then get it to the people who have the access authorities and all the other things that you need to do something with that information so that that tactical edge being the person on the ground calling back from wherever I might have been in the world to a commander that is somewhere in a SCIF in the States. I didn't always have the luxury of being able to do that. Right. So, if I if the system is sped up to the point where I don't have to call back and do that. And I know kind of what my left and right limits are. I am so much more effective as a one person on a four-person team than I could have been before. **Kristina Kempkey:** As we think about integrating data across the battlefield, the next conversation in this series focuses on ground autonomy—specifically resupply and casualty evacuation, where the same operational clock is constantly ticking. Is this fundamentally the same speed problem you've been describing with OSINT, or is it a different challenge? **Stephanie Rotolo:** Yes, I think it's the same clock, just a different application, if that makes sense. So, the clock's running. Having autonomy doesn't eliminate the need for good information. It just makes good information even more important. So, if you have bad data plus faster systems, it's just faster, bad decisions at that point. So, nobody is really benefiting. Technology is really only creating an advantage if it's helping humans make. Better decisions. **Curry Wright:** Steph, thank you. I think that's a really compelling way to think about decision advantage. Charlie, over the past several years we've seen tremendous growth in demand for open-source intelligence and platforms like ExTrac. Governments increasingly rely on commercial companies to help fill critical information gaps. How has your relationship with government evolved? What's working well? Where are the biggest pain points? And how do you see that relationship developing in the future? **Charlie Winter:** It's a great question. All sets of questions. I think since we emerged a few years ago, I think the appetite that public sector organizations in the national security and defense space have for the kind of tooling and data that we provide has definitely increased a lot. It's increased a lot because there's a kind of learning by doing that this is an effective way to leverage data and improve systems and make better decisions. And reach a better understanding of what's happening in spaces where your understanding is undermined or difficult to layer up. But also, I think as has been broader changes in technology and like this massive AI revolution now where people talk about AI all the time, it's no longer kind of a conversation that certain people will have around, you know, machine learning from a like, well, any number of different perspectives, but where you're really drilling down into the kind of more mathematical side of the problem set. AI is just something you talk about all the time across all tasks or your day-to-day life as well. So, I think people are more aware of what the potential is for technology to be integrated into government systems, government decision making. And that has really created an environment which is really fruitful for people like us to go to people working either at the policy level, at the operational level, right at the edge and engage in language, in terms that they already understand, they're already thinking, wouldn't it be great if I could do this with X, Y or Z technology? So, I think there's that aspect of it. But I think there's also been a realization that the data that is accessible in publicly available sources, albeit sources that may be hard to reach, that you need to understand the information environment you're digging into in order to be able to identify where those sweet spots are for where you're going to be talking about making bombs or planning attacks or where IRGC sources are going to be setting out in very simple terms, the escalation ladder, what's going to get targeted, when it's going to get targeted and with what munitions. That kind of stuff, as it becomes more available to the public sector organizations that we're supporting and as they begin to understand that, yes, this kind of stuff is coming through classified channels from other sources, from other kinds of platforms. It's also available and accessible through the open source. And that means that it's transferable. It means that it can get onto systems that aren't in SCIF. It means that it can be passed between allies. And I think those efficiencies are really substantial in terms of why we're looking at private sector. And I'm just going on to the next part of your question here, why the private sector is coming to play the role. It is in national security and defense strategy and at the operational level as well. I think that it's just easier for private sector to innovate and kind of try things out. They may not work, but in the private sector, you're a much smaller organization. There's less bureaucracy, less red tape to jump over, jump through. And I think what we found really useful is working with people who come from the intelligence community or from defense agencies, people who kind of have been, like Steph, really recently working operationally on this stuff and then taking the pain points, the frictions, the challenges that they're experiencing in the field or wherever they are in their SCIF, and translating what we're doing and kind of developing in the direction of those pain points. And I guess as a private sector, as a private sector organization, we just have that ability to kind of push quicker and adapt quicker to new technologies, new frontier models, new forms and ways and means of AI, but also new forms and ways and means of data acquisition and do the hard bit. So, then what we're serving to our customers is something which is the tip of the iceberg in terms of where the effort is. It's the kind of signal rich tip of the iceberg, if that makes sense. **Kristina Kempkey:** Let me follow up on that. You've described several areas where the private sector has clear comparative advantages. From your perspective, is there something government is inherently better positioned to do? How do you see the division of labor between government and companies like ExTrac? **Charlie Winter:** Yes, I mean, I think it'd be crazy and really arrogant to be like us private sector organizations can do everything better than government organizations. I mean, we in the private sector side of the aisle are only as effective as we can be if we understand the problem set that we're trying to help our government customers solve. So, the way that we approach that is by spending as much time as we can talking to our customers, understanding the problems they face, and they have a much better understanding of that than we do. We can kind of make assumptions. We can expect that X, Y or Z organization will be interested in this or that actor in Latin America or MENA or whatever it may be. We can go out, find the data for it. We can go out, figure out the functionality that makes that data sing. But ultimately, it’s the folk who are in national security and defense who are kind of working. Operationally on this stuff, who know what they need to be doing, know what their objectives are, know what strategy they're trying to implement and what change they're trying to make. So, we're a complementary good to something which is going to do it regardless of whether we exist. We're there to try and make the customers we have better able to achieve the kind of effects and measure the kind of effects that they're seeking to achieve. **Kristina Kempke:** I think that's an important distinction. The private sector brings tremendous agility, specialized expertise, and, in many cases, more advanced technology for processing information. But those capabilities still need to be grounded in government strategy, operational objectives, and military experience. **Charlie Winter:** I think one kind of example of where private sector can move fast. So, when we are looking at a new problem set or something, it's a data coverage thing. So, say, a customer comes to us and say says they want to know anything and everything they can about fentanyl precursors and how they're making their way into the North American market, we're able to take that intelligence requirement, respond to it within hours, within days, so that having done some work on our side, we can make data that the person who issued that intelligence requirement is interested in the data that gives them kind of stuff that they don't otherwise see or the data in an environment where they can analyze it and exploit it really, really easily. And that's one of the big things that extract is all about, like making stuff as interpretable as possible and not kind of hiding behind, you know, complex methodological language. The complex methodology is there, but we try to bring insight and foresight to the surface. There's a ton of different examples of where we can respond really quickly to a new requirement that's come in, whether it's functionality or coverage or back-end requirement that a customer has. But I think fundamentally the capability only makes sense in light of the use case that we're supporting, in light of the kind of objectives that our customers are going after. And again, as the public private sector engagement and relationship has kind of blossomed over recent years, I think being able to have folk on our side engage at the operational level, talk to the end user, understand really what they're trying to do and have a frank conversation rather than one which is kind of skirting around certain issues. It makes us much better able to respond. And in a manner which is effective and actually gives them what they need. And I think that's another kind of structural change that we've seen in recent years where because it's kind of necessary for the public sector to engage with the private sector, there is a realization and an openness to talking through those pain points, those frictions, those kind of problem areas in greater detail, perhaps than there was before. **Curry Wright:** I'd like to turn to the importance of data itself. As AI continues to evolve, it seems increasingly clear that data may be just as important as compute. Could you talk about what makes a good AI model, and what happens when even an excellent model is fed poor-quality data? **Charlie Winter:** So, I mean, good models are everywhere now and they are everywhere. And also, it's able it's possible to leverage and exploit those models in a secure manner without kind of giving up user data. Or passing intelligence requirements out of a secure environment into an insecure one. I think over the next few years, there are going to be even better models. We just saw the other day Claude release a new version of Opus that comes on the back of Fable. I mean, these models are just progressively going to get better at passing data, interpreting data, performing quant analysis, whatever. It is identifying patterns and anomalies. The issue, I think, is going to be one of displacement. So, the problem is going to be less. Do you have a good model that can perform a task and is going to be more? Do you have data that actually helps you do that? Because fundamentally, if you put bad data in, you get bad results no matter how good the model is. And I think over the coming five years, 10 years, 15 years, that need for data, which is genuinely representative. Of the stuff that you wanted to be representative of from sources that are actually associated with the people, the actors, the entities that matter to whatever problems that you're looking at. That is absolutely integral, because if you're working with just any old data or huge volumes of data that aren't enriched before they're entering and passing through the model, said model is just going to treat it all the same. So, there's. Multiple layers of enrichment that need to happen. There are layers of assurance that need to happen as well to make sure that that data is actually relevant to the problem at hand, not just secondhand circular reporting, disinformation stuff that's being cranked out by bots. That being said. Having a feed of disinformation and stuff being cranked up by bots is equally really, really interesting. But you want that feed, that data, to be understood by a model as coming from sources which are known for coordinated and authentic behavior or bot-like activity. It's only useful if it has that awareness. So, the problem of good data is, I think, integral now, but it's only going to become more integral further down the line. **Curry Wright:** Before we started recording, you gave us a demonstration of ExTrac, and one thing that stood out was the role of your human analysts in tying everything together. Could you talk about where human judgment fits into this process? **Charlie Winter:** We think about human-machine teaming as something that's essential at both the beginning and the end of the analytical process. On my side of the organization, as Chief Research Officer, I lead a team of analysts drawn from academia, intelligence, and defense. They're genuine subject-matter experts—people who've spent years immersed in their respective problem sets. Some specialize in the Russian military, others in Iranian networks, Chinese activities, or private military organizations. In my own case, that experience came from years studying ISIS. When you've spent enough time studying a particular adversary, you begin to develop an intuitive understanding of how they communicate, how they attempt to conceal their activities, which platforms they migrate to, and how they adapt when circumstances change. That expertise is critical because it determines what data enters the platform in the first place. When I say "sources," I don't simply mean social media platforms. I'm talking about specific Telegram channels, military bloggers, discussion forums, image boards, local chat groups, and the many niche communities where relevant conversations actually occur. We're able to identify those sources without relying on deception, persona operations, or other intrusive techniques. Instead, it's about developing a deep understanding of the information environment and knowing where to look when a customer needs insight into a particular actor operating in a particular location. All of that work happens before any data ever enters ExTrac. Once the data is inside the platform, it becomes available both historically and in real time. Users can perform sentiment analysis, network analysis, investigations into individual entities, and a wide range of other analytical tasks. Artificial intelligence plays an important role throughout that process. But ultimately, the analyst remains responsible for interpreting the results. Technology can identify patterns, surface anomalies, and organize enormous quantities of information. What it cannot do is answer the most important question: So what? What does this change mean? What are the implications? What decisions should a commander or policymaker make over the next seventy-two hours? Those are fundamentally human judgments. Our goal is to give analysts their time back. Instead of spending hours collecting data, cleaning datasets, building visualizations, or preparing information for analysis, they can focus on interpretation, implications, and decision-making. Those data preparation tasks remain essential—but they don't necessarily require human effort. Machines are exceptionally good at that work. By allowing technology to perform the repetitive heavy lifting, analysts can devote more attention to understanding the operational environment and producing better intelligence assessments. **Kristina Kempkey:** We've spent a lot of time discussing OSINT in the abstract. Let's make it more concrete. Charlie and Steph, ExTrac spent months forecasting how Iran might respond to potential U.S. military action. How do you forecast an adversary's behavior without simply making educated guesses? And why is it so important to study an adversary's information environment rather than focusing primarily on our own? **Charlie Winter:** I think from the perspective of Iran like looking back at the last six months or seven months however long it's been and also the previous rounds of conflict that we saw between the Iranians in the USA the Iranians in Israel and the kind of various proxy exchanges that we've seen since 2023 what is very clear is that there are indicators everywhere across the world not just sitting in chat rooms that are used by people who are very close to the besiege or IRGC but also the proxy networks the Houthis Lebanese Hezbollah and so on and so forth the information environment when it comes to this particular conflict but it's exactly the same when it comes to Latin America or Russia Ukraine or the gray zone stuff which is happening on the fringes of these conflicts there is a huge richness of the information that's being shared between these in the information environment where you have people who are communicating with each other from different perspectives whether they're kind of at the grassroots level or whether they're deployed or whether it's kind of fusing their perspectives with what's being pushed out by officials but I mean language barrier is a really significant thing it shouldn't be any more but it still continues to be a really significant thing when it comes to identifying those indicators but more than that it's the volume of data and the fact that it's happening on like dozens of different platforms thousands tens of thousands of different sources at the same time so when we have been working on trying to understand what escalation looks like in the event of U.S. strikes on Iran or what right now would happen in the event that there was a resumption of hostilities. There's been a 48-hour pause, perhaps with Israel jumping in as well. We're able to tap into the data that we have and kind of treat it as like a, without sounding too sci-fi about it, kind of like a hive mind for the actors involved in these conflicts, where you have thousands of people constantly and continuously assessing and analyzing and issuing commentary on what they think the next course of action will be. You also have official sources. So, I mean, the Iranian state and the IRGC have dozens of media assets that are overtly and covertly operated by them as entities. Same for the military, the Artesh, and similarly, the Iranian state. So, I mean, the Iranian state and the IRGC have dozens of media assets that are overtly and covertly operated by them as entities. Same for the military, the IRGC have dozens of media assets that are overtly and covertly operated by them as entities. Same for the But essentially, if you have a full living picture of what is happening in that or in those particular slices of the information environment, and if you're able to look back historically as well at what happened last year before the strikes on Al Udeid in Qatar, for example, or how did the Iranians respond to the embassy strike in Beirut in August last year, with a view to understanding what kind of threshold is for engaging, what their escalation ladder looks internally. And I think it's really important to do that analysis and base it not just on what is circulating in the Western press or what's circulating in the think tank community. That stuff is really important, really useful, and it should be layered onto the kind of operating picture. But getting into the Iranian military, for example, are seeing on the ground what they're hearing, the rumors, the things which are being overheard, but also what the official side of things is looking like and how that's developing. You can arrive at a really detailed picture when you have enough data and when you have enough context to understand how that data is translated into real world signal in the past. So whether it's understanding specific sites or assets that are highly likely going to be the focus or the priority of upcoming military operations or whether it's understanding what kind of weapon systems are going to be used and when they're going to be used or when a particular proxy is going to be activated, that kind of decision calculus is available in spades if you know where to get it and know how to interpret it when you're at the kind of user end. And I think when we orchestrate AI against this data, to enable us to get that kind of information, understanding on the AI's part, so in the orchestration layer, it understanding how to use probabilistic language in a very, very selective and constrained way, but also understanding source bias so it can come out with and provide like a graded confidence statement for the kinds of insights that it may surface. That's absolutely critical. That's exactly the kind of stuff that we've hardwired into the orchestration of ExTrac and co-analyst our agentic system, which is the kind of interpretive layer that you can kind of chat to pull out that kind of insight. Thinking like an intelligence analyst, it is just as cynical, just as skeptical of biased data as you would want a good intelligence analyst to be. And it will say that right to your face in the event that you're pulling together an analysis on, I don't know, BDA or trying to understand what's next in the escalation cycle. So yes, there's a few different angles to it. But again, it comes back to this need for good data that's actually genuinely relevant to the problem set, that context window, which is holistic and historic as well. So, you can do that time series analysis and then hardwiring tradecraft into the way the agents think. That's another critical part of it. But those things come together and you can do this stuff at machine speed. You can do this stuff in basically real time and feed that directly into the decisions you're making. **Curry Wright:** As we wrap up, let's look five to ten years into the future. Imagine a world where open-source intelligence and commercial data have become fully integrated into military operations—not simply as another intelligence discipline, but as a continuous layer of situational understanding. What needs to happen over the next several years to make that future a reality? And what responsibilities belong to government versus the private sector? **Charlie Winter:** But I think for me, I don't think there's ever going to be like a replacement or a substitute institution thing. I think that the kind of work that is being done in the OSINT sector and the kind of technologies which are just getting better and better at doing what they do, that's going to become more fused with the kind of traditional collect. And also, I think a lot of OSINTS is often anchored to communication data, stuff we can pull in from the information environment. But increasingly, as technology allows for, the capabilities we have, and we do this at ExTrac, the capabilities that we have in the private sector side will be also doing their own data fusion. So, you're not just looking at stuff in the information environment. You're not just doing narrative intelligence or whatever. You're pairing that up with a live readout on what's happening in the physical domain. So, you can use this for understanding if you're planning an evacuation route from a city for embassy staff for example, you know which roads to go down and which not to go down based on what people on those roads are saying and reporting and making videos of and so on and so forth. And that kind of fusion of the physical domain with the information environment, but then also layering on top of that ad tech or maritime data or flight data means that the OSINT offering becomes much more than what's happening on, well, it's been much more than what's happening in the information environment. It's been much more than what's happening on Twitter for a long time. But for a long time, it was what's happening on Twitter. And I think that trend of increased data fusion and technology, which enables and speeds up interpretation and enables investigation and kind of cross comparison and red teaming and visualization, all of that means that that passage of OSINT from being a nice to have to being a just a mandatory layer of the decision cycle, I think is, is going to be something that takes shape, takes shape, like more and more over the coming couple of years. I think moving further out, we're likely to see a place where, like we spoke earlier about how AI models and kind of the frontier labs are brilliant, they're going to get more brilliant. And the issue isn't going to be the quality or performance of the model; it's going to be the quality or performance of the data and making sure that that data is actually relevant to the problem set. If it's not relevant to the problem set, like all of this stuff actually ends up being problematic, because it will give you really credible, well evidenced bad reports. And you don't want that if you're trying to make decisions in a really dynamic environment off the back of it. So those are just two thoughts off the top of my head. **Stephanie Rotolo:** Yes, I think there's, I think there's also an element here of accessibility and just something even as simple as the user interface with a lot of these, these platforms that are out there. Being in that space and being handed a lot of different systems to use and apply and see how they would work. The first test was always; can I hand this to the youngest, newest, freshest member on this team? And can they use it responsibly, appropriately? And can something come out of it? I think there that there's going to be, there's also an element of like, when you start talking Intel, which gets into the authority space. And I think that the benefit of OSINT, whether you're calling it OSINT or publicly available information, is there does exist a bit of a gray line there. So, I don't necessarily need a, you know, thumbs up your certified Intel analyst to create products that are for all intents and purposes. They could be Intel products, but we don't need to call them because of where the information was derived from and where the data sits, if that makes sense. So, I think that is going to become an element of it of like we are and I say we I mean, by in terms of the U.S. military is constantly being asked to do more with less and then especially in the special operations community. So, I think anything that can hand a commander a tool that, again, they can hand to the newest person on the team, and they can readily figure it out. And we can, with some certainty, fairly trust what's coming out of it. I think that's going to be critical, and I think that's going to continue to propel this broader acceptance of, yes, OSINT needs to be folded in. It adds the credibility of all the other INTS as well to getting a coherent picture of what's going on in battle space. **Kristina Kempkey:** I'd like to finish with one final bonus question that ties this discussion to the next episode in our Emerging Technology Series. Imagine an autonomous ground vehicle evacuating casualty. As it's moving, an OSINT platform detects that a protest is forming along the planned route and automatically recommends—or even executes—a different route. How close are we to that reality? And what has to happen before commanders trust a machine to make those kinds of decisions without requiring human confirmation? **Charlie Winter:** That's a good question. I'll have a go at it and then Steph can give a better answer. I don't know. I was reflecting on this kind of future of OSINT point earlier today and thinking about the idea that it becomes just a standard part of the understanding in any sort of operating environment. It's just there and it's continuously updating, continuously updated, and people, whether they're at commander level or deployed, are able to dip in and out of it and understand or get responses for specific RFIs that they have on a rolling basis. So that kind of thing is like simply happening now. I mean, not simple, but it is happening now. And the space that I'd say where we're already in, where essentially the stuff that you can do through open source is essentially an environmental data layer that underlines or can underpin decisions that are made on a minute-by-minute basis. I think when it comes to pairing that up with autonomous technologies, essentially now you can have AI interpret the data. And if you want it to make a decision based on the data or recommend a course of action on the data, you can have a data set up that's based on the data that you put in front of it. So, I think from a technical perspective, technological perspective, that side of it isn't so challenging. The real challenge will be figuring out where that human should be in the loop, because especially if it's something like an evacuation for a casualty, you don't want a model making a decision that a human wouldn’t want to make. And it's the same with kinetic UAS. Like there's a reason that people are adamant that there needs to be a human in the loop somewhere in the way those technologies work. So, I think from a technical perspective, technological perspective, that side of it isn't so challenging. So, I think that for me, it's less a technological constraint and more kind of ethical or moral or, you know, philosophical challenge that we need to get our heads around, because actually getting this information in front of a machine. That machine makes sense of it and then figures out what it thinks it should do in response to that, that data, that information. That’s kind of something that can happen now. **Stephanie Rotolo:** Yes, I think I think you make a good point with that, Charlie. And I would say it's less I don't think it's going to be like the autonomous systems being able to do it. It's going to be trust in the systems. And I think to achieve that trust or for a commander to feel comfortable saying, I'm going to let this machine act in this scenario, the AI first has to show it works. It can't just say reroute. You have to be able to audit why it's coming to that conclusion. Right. Secondly, the data has to be reliable, like we've talked about. It has to show that it can distinguish between whether this is some kind of misinformation or things like that. And then third, as a commander, you'd have to have built the confidence in the system through repetition in advance of meeting it. So that's not going to happen. It can't happen in combat when the casualty is moving. It has to have happened over perhaps hundreds of exercises and other training missions where the AI is consistently making good recommendations and earning trust. I think of AI as a partner force in this scenario. I think that's how we kind of need to approach it, at least initially, is it cannot be expected to make the decision for us. But we have to have an element of trust in it before we work with it as a partner, which is what it is essentially doing if it's moving somebody. So, I think in the next few years, from my perspective, we will see machines making those decisions and probably lower risk logistics scenarios, maybe. But I think for the higher consequence decisions, definitely the humans can never disappear from the loop. And then they just can't replace human judgment. But we need to develop that trust. Commanders need to know they can trust it. The organization needs to know they can trust it. And essentially, once that machine is able to give commanders time back to think about the next steps, because if you're making that decision, you should already be like five steps ahead of that decision. So, the commander should have. And understanding and comfortability with like, I now know and I've validated that this thing can make this decision because I have to do the five steps after this casually gets back. **Kristina Kempkey:** Charlie and Steph, thank you both for joining us. This has been a fascinating discussion about one of the fastest-growing areas of national security technology. We could easily spend another hour discussing the ethical questions surrounding artificial intelligence, but we'll save that conversation for another episode. Thank you for joining us for the Irregular Warfare Initiative's Emerging Technology Series. In our next episode, we'll speak with the team from Forterra about how open-source intelligence integrates with ground autonomy and what that means for future military operations. If you enjoyed this conversation, please subscribe to the Irregular Warfare Initiative newsletter and visit us at www.irregularwarfare.org. Thank you for watching. **Charlie Winter:** Thank you very much. **Stephanie Rotolo:** Thanks so much.