In this episode, recorded during BICOM briefing, Tal Hagin presents his VIST method for identifying AI generated misinformation, and discusses the use of AI in psychological operations and information warfare. Tal also outlines how AI-driven manipulation is reshaping conflict reporting and what can be done to counter deception with case studies from Gaza, Sudan, and Iran.
Tal Hagin is an Information Warfare Analyst and Media Literacy Lecturer with over a decade of experience in information warfare and strategic communication. He also works as research fellow for FakeReporter, an Israeli watchdog group focusing on combating fake news, online incitement, and foreign influence campaigns.
Transcript
(This transcript has been automatically generated by AI — please excuse any potential errors.)
00:00:06:24 – 00:00:29:03
So hello, everybody, and welcome to this BICOM briefing and also to this, the recording of this podcast. I’m Richard Pater the Director of BICOM, and I’m delighted to be back here in London to be with you all. Today is December the 4th. And we are going to be discussing the impact of, of AI, both in its coverage of the region and how it also affects the conflict itself.
00:00:29:07 – 00:00:59:17
I’m delighted to welcome Tal Hagin, who will be our guest speaker today. Tal is an information and warfare analyst and media literacy lecturer specializing in open-source intelligent investigations and disinformation analysis, AI generated media verification, and to tragic intelligence research. He has worked with government entities, private companies, NGOs and educational institutes to detect, mitigate information threats, threaten digital resilience, and support evidence-based decision making.
00:00:59:19 – 00:01:22:06
So over to you tackle and thank you very much indeed. So, thank you very much for having me. And thank everyone here for, attending this the a little bit about what we already kind of went over myself so we don’t even have to go here, but, when we’re talking about AI in this current area that we’re in, it’s something that is on everybody’s mind, but nobody really knows how to handle it.
00:01:22:08 – 00:01:44:07
And we don’t also understand the risks that I can have in the information space, especially when we’re talking about the Middle East, but Middle East or Africa or other nations where the amount of information that is coming out of these places is either very, very limited or heavily controlled. And so, one way in which we access information from the field is through imagery.
00:01:44:09 – 00:02:06:18
This imagery can take the form of satellite imagery, circulated surveillance footage, eyewitness testimonies, or footage that is directly filmed from the point of view of the perpetrators or the victims. Now we see digital sources all over, of what’s happening, for example, in Sudan towards the end of October. And so many of these different images came onto our feeds.
00:02:06:18 – 00:02:30:17
Now, what’s the point of these images? Why are we seeing them? Well, one, the point is accountability that we have photos of the perpetrators and the victims so we can try to help those who are, harmed, but also try to hold those who perpetrated it accountable. Awareness as an individual’s need to know what’s happening. And so, if you don’t film anything and you just write it in a bylines report, then how many people are actually going to read it?
00:02:30:18 – 00:02:53:12
Imagery is the best way in which to have that conduit of information, political pressure. When people are aware of the situation and they understand how horrific it is, they want there to be accountability. And so, there’s political pressure on different, entities, either, government and agencies or NGOs to enact pressure on those points. And lastly, material for investigations.
00:02:53:12 – 00:03:18:13
We’ve seen various different entities like Bellingcat, or the centre for Information Resilience, being able to track down the perpetrators by, creating a timeline of events, where things have occurred through this imagery. But the question is, what happens when the majority of the content that’s coming out to us is AI? This photo, for example, that you see with the woman holding her child with the two gunmen was the front page of social media for about a day.
00:03:18:15 – 00:03:44:11
And that was the image of the massacres happening in Sudan completely AI generated. And what happens when these AI imageries become the focal points? Well, three things. First, it sends investigators on UN force, chases. Now you’re going to be trying to investigate the geolocation of video. It isn’t real. And that’s completely irrelevant to your investigation. And now you’ve wasted valuable time.
00:03:44:13 – 00:04:03:24
Secondly, it gives atrocity deniers ammunition, which is something that I see constantly in my feeds where when this type of ER imagery is showcased online and then it’s debunked, it’s like if you guys had real imagery of such horrors happening, why did you have to share AI imagery? And it allows those who want to deny horrors happening on the ground.
00:04:04:01 – 00:04:25:02
An amazing opportunity to do so. And for journalists, it ruins the reputation. Journalists have to constantly keep their reputation in order to be able to express themselves, to, the common individual. At the end of the day, they very much live within a place of appeal to authority. People have to trust what they have to say. They can’t always divulge their sources.
00:04:25:02 – 00:04:46:02
They can always show, their work ethic and how they got to every single conclusion. And so, when they accidentally share one of these, pieces of AI content, then the viewers, both those who want to harm them and derail their journey, their journalistic careers, have no ammunition, but also those who trusted them start to question why they are listening to them.
00:04:46:04 – 00:05:09:00
Now, this is a lot of complaints, but the question is, what do we do about this? Because every one of us can discuss why AI is harmful to us, but how we solve it. And so, we’re going to go through an AI verification model that I’ve the way in which I conduct myself online called visit methodology. There’s no AI because I couldn’t, figure out a good word for it.
00:05:09:00 – 00:05:42:22
So, it’s just visit. But the staff first visualized the landscape of manipulated media. You need to be aware that AI generated content isn’t the only type of manipulated media out there, but there’s also different forms of AI generated content. And we’re going to explain, shortly why that’s important to understand. Secondly, identify AI specific indicators. Once you know what type of AI is out there, you need to know the differences between different platforms, what type of watermarks they have, what type of metadata they contain in order to be able to identify it.
00:05:42:24 – 00:06:20:05
Second source the original creator. Real media has real people behind it. When there’s a real photograph taken, somebody took that photograph. We can track that person. And if somebody created false media, then we can also track that down as well. And lastly, treat tools as signals, not proof. This is something that’s very, very important. And we saw this recently with when Elon Musk, conducted, a new strategy on Twitter where now you can see everybody’s location, except, for example, there was one Palestinian journalist who was located in the Gaza Strip, but his, location was in Poland.
00:06:20:11 – 00:06:39:10
And that was likely due to the fact that he either worked with a media company or had an individual in Poland to work with him. And so, this was an indicator to me as a researcher, that he works with somebody else who’s not just in Gaza, but for somebody who doesn’t understand how to operate with tools and sees them as full proof, oh, that’s proof that he’s in Poland when it’s not.
00:06:39:10 – 00:07:03:00
When I was able to verify the actual geolocation that he was on the Gaza Strip. And essentially every time you see a piece of media before trusting any image, visit it. Go through the list. Let’s go through the first part. Visualize the landscape of manipulated media. You need to know what is possible from the quote by tens who know your enemy and know yourself.
00:07:03:00 – 00:07:23:01
In 100 battles, you will never be defeated. You need to understand your enemy and what they’re capable of. What I think, for example, of the effects of AI generated content. I also think if I was a PsyOps unit, how would I conduct operations against the civilian population? And by understanding the way in which I would conduct those types of operations, I know how to fight against them.
00:07:23:03 – 00:07:49:21
It’s a different example that we’re going to go through. This doesn’t come for all of them. Repurposed imagery, individuals trying to claim that this shows the destruction of an Israeli F-35 during the Iran Israel war. It actually showcases a crashed F-35 in the United States from a year prior. Photoshop. This was released by various different, pro-Iranian, individuals claiming that this was a drone shot of Netanyahu, kind of like exaggerating that they are very close to the prime minister.
00:07:49:23 – 00:07:56:00
It’s actually a photo from 2009 that they photoshopped in with the drone deepfakes.
00:07:56:02 – 00:07:57:17
You can kind of listen a little bit
00:07:57:19 – 00:08:14:18
You see what’s happening. The people of Palestine are suffering beyond imagination. You have to change this. That’s I think, in a second. We sometimes also see the real video in this field, one when you actually just realize how bad the picture is. A lot of people actually fell for the, deepfake of Leonardo.
00:08:14:20 – 00:08:31:23
Let him then, for example, we also have CGI. We have this video that was, that they prepared to show, Iranian missile strikes on the state of Israel. And a lot of people said that it was AI. It wasn’t, it was CGI. This is an example of one of the type of meteorites that they use and the exact same footage.
00:08:31:23 – 00:08:57:15
And this was actually taken; some place in Asia. I forgot the exact country. Another example of things is AI generated images. Now air generated content. There’re various different types. It’s not just images or videos. We’re going to go through a few examples. This image was, produced and showcased by the Fars News Agency in Iran. This, post is by the actual foreign minister of Iran, and they tried to claim that one of their missiles did not strike the Soroka Hospital in Beersheba, Israel.
00:08:57:21 – 00:09:23:02
In fact, it had hit two military sites adjacent to the hospital, and it was due to the blast wave that the hospital, was harmed. Now this map is completely AI generated. It’s not real and doesn’t coincide at all with where all the different things are located. When I conducted an actual investigation into this using open-source methods, I was able to pinpoint, through videos and photos exactly where the missile struck by about one meter off from this real location.
00:09:23:04 – 00:09:49:01
And these types of images are abundant. And we see that it’s not just random activists online. These are government officials posting this type of stuff. Then we have air generated videos such as this, individuals alleging that these are Israelis protesting for the end of the war with Iran, trying to demoralize the Israeli population. But of course, when we understand the signs of AI, the video, watermark from that generation, we look at the signs that don’t make any sense in Hebrew or English.
00:09:49:03 – 00:10:10:08
And the weird finger, extrapolation from these images. Now, there’s another type of AI that isn’t really talked about a lot, and it usually is able to trick a lot of people, including mainstream media giants, who usually do some type of verification process. AI splicing is what I call it a splicing. Essentially means that you take real photos and real videos and you extend them.
00:10:10:10 – 00:10:30:18
And the reason this is very, very important and very useful for those who want to spread propaganda is because you already have all the details in the photo. The AI doesn’t have to generate anything new from that. One of the most famous examples was the Israeli missile strike on the Evin prison, where this, video of the strike went all over social media.
00:10:30:18 – 00:10:54:11
It was shared by major news organizations as authentic. I was on the first people to raise the alarm that this actually was based off an image of the prison from a few years prior, and someone had AI generated and made it grainy, so it looked like it was surveillance footage. And another example that was, shared is allegedly these were, this is a video of Iranian diplomats and officials fleeing the country in their vehicles.
00:10:54:13 – 00:11:12:17
This was posted here, but the video is actually from a video from 2000 and, 23 where it cuts off. Right. Here we can see it’s the exact same photo, and they don’t have to really manipulate too many details. It looks real when you’re looking at it, because Israel, a big part of it is. And so, you don’t, question it.
00:11:12:19 – 00:11:30:20
Now, the second thing is identifying these AI indicators. Now that we’ve gone through the different types of manipulations, we need to discuss AI, itself. One of the biggest issues that I always hear is how old to identify AI, we just have to look at the weird fingers or, you know, they’re they don’t make sense when they’re talking.
00:11:30:20 – 00:11:55:13
AI is this today, by the way, that was 2014 with the cows. This was 2022. This is 2025. You’re not identifying AI with the fingers anymore. AI is far more advanced. And we need to be aware of how advanced it is and no other methods to actually identify it. Now, many of you guys who work in the field, you know how to look for the obvious details, which is the fingers that are a little bit off, or maybe the numbers in the background don’t make sense.
00:11:55:13 – 00:12:20:17
But I’m going to go through two other examples of ways to identify AI imagery. That is usually able to work, even if you can’t tell that the video is AI. First are watermarks, watermarks. They’re patterns within the image itself. These are three different examples. The image of the lions with the guns, which is obviously AI. By the way, this is from Rock, and you can see that they have a little watermark in the bottom right side on the bottom.
00:12:20:17 – 00:12:42:06
It was posted actually, by Tehran Times and other Iranian newsletters from video showcasing an Iranian missile strike on, there was actually a strike that night and, but, so they made a fake AI video of the strike, and this is from an alleged Palestinian protest, but it actually was completely generated by, of now the main thing that we need to remember here is that I have patterns.
00:12:42:08 – 00:13:00:22
They always are the same from every single platform. Go through an example of this. We have this, video that was posted about a week ago alleging, IDF soldiers who are burning the American flag now without looking too closely. And if you know what the IDF uniforms look like, or you see my new little details, let me dismiss them.
00:13:00:22 – 00:13:21:07
Fine. They were a special unit that they didn’t want to make themselves exactly like the IDF. And the small little issues that you see in the video are due to the fact that, there was some weird issue with the buffering. That’s why you see the weird details. But there’s two things here that make me, which made me very, very suspicious that this was I, the skull and the framing.
00:13:21:09 – 00:13:49:11
If I was filming my fellow soldiers burning, a flag. When did I want all of them in frame? Why would I be so cropped in the image? What am I trying to hide outside of it? And also, why does the skull suddenly appear there but then disappear afterwards? Now I know, for example, with saw AI that Sora appears in this pattern on the on its images that it has a as well as I know what the, saw.
00:13:49:13 – 00:14:09:08
I the little gif. Now you guys might not have noticed it, but if you then know where to look, you might have noticed in the top left corner they messed up. Deleting the, Sorry, I right there. If you see right above his head, just there for like a split second, you can see the, saw AI appear.
00:14:09:10 – 00:14:42:00
So, they failed at deleting it completely. But the only way that I or anyone was able to notice that is if you know what to look for, you know that these, different types of AI videos have patterns in their watermarks, and then you look for those patterns. The next is metadata. Now, not every single AI platform that creates AI, generated imagery or videos has this, but there’s currently a movement, called this, Pas, which is to have a permanent watermark within the actual metadata.
00:14:42:02 – 00:14:59:16
And the way in which this can be very helpful for you is that if you suspect that a video or photo is AI, you can go to that platform, upload the image and ask it if this was generated by them, and then the system will be able to calculate, rough percentage whether or not it was. And it’s quite accurate at the moment, but not everyone is a part of this grok.
00:14:59:17 – 00:15:28:24
AI for Twitter is not a part of this. For example, but quite a few of them are. And so, I would recommend being familiar with the site, see which companies and groups are part of it, and then that can be another tool set that you use. Next source the original creator. As we said earlier, one of the most important things that I always, have when I’m looking at imagery or videos or photos is that every real image has a real person behind them, and I can find that person.
00:15:29:01 – 00:15:53:11
We’re going to look at two different examples that were uploaded and both of them, this one was claimed to be a real image from Sudan, and this photo was claimed to be AI generated because of the weird thing with the hand that you see on her, right hand, which doesn’t really make sense. It looks very strange. So, the first image we see that in almost every single upload of that video or photos, online, that this little watermark appears, in the corner.
00:15:53:13 – 00:16:13:07
So, what do we do? We look, for where that page comes from. Turns out that it is a creative AI specialist. Who, it comes from his page. And if we look at his content, we see that he makes a lot of very, very similar content. Now, when I, corrected him and, thankfully, actually, the post went viral and a lot of people started, basically harassing him to change it.
00:16:13:09 – 00:16:37:11
He changed it a few days later to include that it was generated, in the video. So sometimes we are able to make an effort to force these types of individual individuals to change, what they say. And I can understand that people want to create AI generated imagery because they don’t want to show real victims, you know, a lot of us don’t really want to exploit, innocent individuals who are harmed in conflict zones without their consent.
00:16:37:11 – 00:16:56:23
We want to make sure that they’re okay with that. So, what we do is we uploads, AI generated imagery to illustrate, but we need to be aware of how quickly that illustration can be lost. And suddenly you are spreading fake news and your, leading to atrocity denial for those people. And so, you need to be aware of this when you’re using AI, if you ever do in your work.
00:16:57:00 – 00:17:18:23
The other image that we have is this woman from Gaza. What did I do in order to verify it? I don’t, claim it, I or not look for the photographer. So, I looked, not every person that uploads the re uploads this image is going to credit the original photographer. But some people do. And so eventually, by looking at different places on Google Lens, I was able to find this post by Eye on Palestine.
00:17:19:00 – 00:17:38:01
And they include the agency that the person works and their name. What do we do with this information? We go to Google and put in his name and the company. And once we have, Getty Images, we run through Getty Images by looking at specific keywords like the Palestinian woman, crying, stuff like that that is related to the image.
00:17:38:03 – 00:17:54:14
And then we find the image on his post. Well, okay, maybe, but this is only one image on. It’s going to be hard to prove to people that it was, true. Well, once we have her name, we can find more images of the woman from very just different angles. And once we have her full name and various different angles, we also have video of the woman.
00:17:54:16 – 00:18:16:20
And so, it’s incredibly important to realize that finding the source of these images can be a very, very useful tool in figuring out if it’s real or not without even being able to tell if the videos are not. Now, for the last part, treat tools as signals, not proof. We discussed this a little bit earlier in the presentation, but it’s very important to be to understand that tools are signals.
00:18:16:20 – 00:18:38:07
For example, Google Lens is something that I use every day in my work is incredibly powerful tool. I’m assuming that everyone here knows with Google lenses, but to reiterate, you essentially can put any type of photo or screenshot of a video, and it’ll show you where else it appears on the internet, but it’s not foolproof. Here’s an example of a photo that was taken in 2021 of the Iron Dome, against, Hamas rockets.
00:18:38:07 – 00:18:59:22
It’s a very famous photo, but the Google Lens shows it appearing in 2017. And how can that be the case? Well, Google Lens has an index error issue where sometimes it will incorrectly put, images that are unrelated to the link that it’s actually giving you. If you go to that post, it and the title is correct as well as the date, but the photo doesn’t appear there.
00:18:59:24 – 00:19:16:20
And so, you need to be aware that the catalogue results aren’t the end game. Now, what about, you know, a common question that I get in all my presentations is what about AI detector tools? I don’t trust any of them. The reason is because of this. Here are two different AI detector tools that are used many, many times.
00:19:17:01 – 00:19:36:24
And here is them identifying these two photos that are real as AI. And here’s another time of them identifying an AI generated image as 100% real. And so again these are indicators. They’re not foolproof. If and if you put an image into multiple AI detectors, and all of them are telling you that it’s AI, that just tells you that there’s a strong signal that it might be AI.
00:19:36:24 – 00:19:42:01
It’s not a final verdict, and it’s not something you can publish.
00:19:42:03 – 00:20:05:23
Now, a more concrete example of this is with chat bots. Now, I recently wrote an entire piece talking about AI in fact checking, and why LMS essentially cannot do the work of fact checkers. There’s a direct a QR code, but afterwards I can also give you the link. And for example, we have this, photo of a, a Palestinian girl at a distribution site in the Gaza Strip from the 26th of July 2025.
00:20:06:00 – 00:20:29:04
And grok responded to somebody asking if the photo was where it was taken, saying it was the ZT girl from August 10th, 2014. It didn’t provide a source. It didn’t provide any context as to why it was saying this, but I was actually able to find the photograph that it was referencing. It was referencing a different photograph of a young girl who had very similar blue eyes, that looked completely different at the end of the day, but similar blue eyes.
00:20:29:04 – 00:20:48:05
And because the lens at the end of the day, they don’t know anything. There’s statistical probability. It’s all these two images are so similar that they have to be the exact same event and therefore told the user. So, while LMS can be very useful for our work, and especially with fact checking, for example, you see, you’re looking at a video, you want to verify what country it’s from.
00:20:48:11 – 00:21:18:05
So, you see that it has a license plate. You put it into the yellow and ask what countries have blue license plates. Now you can you have to then go through those results. You can’t just trust it. It helps you along the way in your process of research. It is not a foolproof tool. Now going back to what we said with visit, visualized the landscape of manipulated media, identify specific indicators, source the original creator and treat tools as signals, not proof.
00:21:18:07 – 00:21:42:07
And if you come at every single image with before trusting any image, visit it. Then we don’t have to get to a situation where atrocity denial is committed, where our researchers lose precious time in identifying, where these, events took place. And at the end of the day, it ensures factual reporting and credibility for both researchers and journalists.
00:21:42:09 – 00:21:59:16
I hope this was helpful with understanding AI as a lot of different, and a lot of information, but I’m happy to, talk with you later or through email or a conversation to better explain these concepts. Thank you so much for listening. Thank, thank you very much. I’m still I’m still recording. I’m going to ask one question directly.
00:21:59:16 – 00:22:18:06
Then I’ll turn it off and open it up to the floor. I wondered if you could if you could just drill down in terms of the ramifications, implications within the, within the Middle East and who are likely to be kind of bad actors in this space. Are we seeing non-government actors? We’re seeing any governments kind of taking advantage of this.
00:22:18:09 – 00:22:36:16
So, it’s kind of deliberately, use that for disinformation. How are these tools being used in the real world of Middle East politics? So, I would say that there’s, two ways in which, our general content is being used in the Middle East at the moment. On one hand, it’s being used for monetary gain.
00:22:36:18 – 00:23:04:02
At the end of the day, as I said earlier, photos, are much more appealing to individuals. Makes them click, it makes them interact. It makes them engage with the content a lot more. And platforms like X, for example, that supports not content but views, is a platform that is filled with individuals who are just posting and generate content because they get a lot of money from all the people that are following them and research and giving them views.
00:23:04:04 – 00:23:28:15
The other side of it is psychological warfare operations, by different government entities and activists who are trying to either demoralize the enemy or exacerbate their own, achievements in the field. We saw this, with the Iranian strikes, and the Israeli strikes in Iran, where the, the video that I showcased of the Evin prison being blown up was posted by pro-Israel, activists and government aligned, figures.
00:23:28:15 – 00:23:48:24
I think even one of the governments, one of the ministers even posted the video. And the other hand, we had Iranian officials posting AI generated imagery, such as the map that we saw in order to justify strikes, within the state of Israel. And so, we’re seeing that agent AI generation, is not just, another aspect of psychological warfare.
00:23:48:24 – 00:24:10:08
I would say it’s its own field now, where the war of reality, so to speak, of what actually happened on the ground, where we are identifying how successful the attacks against our enemies are based on what our citizens think is true. Because at the end of the day, if you think about it, it would be much better that none of the information is public that the Me.
00:24:10:08 – 00:24:33:18
As a military leader, I know that we struck all of the enemy’s, critical components and that’s it. And then I can go home. But if my civilian population, all they see is my cities on fire and burning down to the ground on AI generated content, then for the citizens, we’ve lost the war because they don’t. They’re not aware of the covert intelligence operations that have been conducted or the successful success of them.
00:24:33:20 – 00:24:50:22
And so, when it comes to the population, how they view this, matter on the ground and I generate content, is the new battlefield of that. Just a follow up. It was interesting. We saw that the Iranian regime, over the last few days have been releasing what appears to be real footage of some of the Israeli strikes.
00:24:50:24 – 00:25:15:03
First of all, are they have you if you check that they are real. And just if you would speculate kind of what the motivation of the Iranians to release that material. So, on your first point, there are four videos that the, that were showcased to the world. Only one of them is real. The others are a generated content by most likely individuals trying to exacerbate, the destruction of the Iranians capabilities.
00:25:15:03 – 00:25:32:07
I don’t know if it’s the second or the third one. I can show you it afterwards. Which one is real? The others are completely, generated when it comes to why I ran released them. Because at the end of the day, what we see in the video is destruction. We’re seeing a pinpointed Israeli airstrike destroying their command centre.
00:25:32:09 – 00:25:53:03
But what people don’t understand is that that doesn’t mean that your perception of these videos is not their perception of the video. If you actually looked at Iranian pro, Iranian government, channels and their internal communications, they see it as their people standing strong. That until the end they thought that even though they were, they were struck and that they were killed.
00:25:53:07 – 00:26:10:01
They stood strong in the face of the, in the Israeli, the Israeli air airstrikes. And that’s something that’s also important to understand that just because the footage showcases one thing to you or it can be interpreted in many, many different ways. And that’s exactly what we saw with the Iranian footage that they released. Correct.
00:26:10:01 – 00:26:11:04
So, thank you very much.