Latest / Stage Talks with Bellingcat / "Building Protective Tech- Dealing with Leaked Intimate Imagery" with Madelaine Thomas
Transcript
- 0:04You're listening to a stage talk titled Building
- 0:07Protective Tech Dealing with Leaked Intimate
- 0:10Imagery. This week, we spoke with the founder
- 0:13of Image Angel, a company developing forensic
- 0:16watermarking technology to help trace the source
- 0:19of online image abuse. Madelaine explains how
- 0:22the technology works, the challenges of encouraging
- 0:24adoption by platforms and how her own experience
- 0:27led her to found the company. We also discuss
- 0:30why the technology isn't open source and explore
- 0:34alternative approaches to addressing this systemic
- 0:36issue. You can find links to all of the resources
- 0:39mentioned in the talk in the podcast description.
- 0:42This talk was hosted by me, Charlotte Maher, on
- 0:44Thursday the 2nd of July, 2026, in the Bellingcat
- 0:47Discord server. Welcome back to the Stage
- 1:02Talks podcast. Today we are diving deep into
- 1:05the frontier of digital safety, journalism integrity
- 1:08and the engineering of online accountability.
- 1:10A short trigger warning that in this discussion
- 1:12we may touch on themes of sexual exploitation
- 1:15and image -based sexual abuse. When sensitive
- 1:18media leaks, when war crime imagery is manipulated
- 1:21or when intimate images are shared without consent,
- 1:24the digital damage happens in seconds. For decades,
- 1:28our cultural and legal response has been entirely
- 1:31reactive. We try to flag, report and take down
- 1:34content after it's already spread across the
- 1:37internet, as many of you will know. But in an
- 1:39era of generative AI and automated disinformation,
- 1:43reactive safety is a failing strategy. Our guest
- 1:47today is tackling this problem by shifting the
- 1:49entire paradigm from reaction to provenance.
- 1:52Madelaine Thomas is the founder of Image Angel.
- 1:56Her company built advanced server -side forensic
- 1:58watermarking tools that embed invisible mathematical
- 2:01trackers into the structural frequency of images
- 2:04and videos the exact millisecond they are viewed.
- 2:07It means that no matter how much an asset is
- 2:09screenshotted, cropped or manipulated by deepfake
- 2:12algorithms, it carries a permanent invisible
- 2:14digital paper trail back to its source. Within
- 2:17this next hour, we'll be exploring the exact
- 2:20methodology of how this technology is built,
- 2:22the friction between open source transparency
- 2:24and defensive security, and how this math could
- 2:26reshape everything from creator safety to global
- 2:29information warfare. As usual, you can place
- 2:32your questions in the chat accessible on the
- 2:34right hand corner of your screen, but please
- 2:36remember that this is being audio recorded for
- 2:39the podcast, so if you don't want me to mention
- 2:41your username, please add that as a note in your
- 2:43question. It's also worth noting that Bellingcat
- 2:46does not endorse any individual company. And
- 2:49today we're talking about the technology rather
- 2:52than the company itself. Madelaine, your journey
- 2:54into building this technology is deeply personal.
- 2:57I want to hand the floor over to you. Take us
- 2:59back to the beginning, your story and how this
- 3:01technology works under the hood. Thank you for
- 3:05having me here. I'm really excited to go through
- 3:07this with you. I guess I need to just caveat
- 3:11what you've just said that we're changing the
- 3:13way the world's working. We're doing it slowly.
- 3:16And the person in charge of this company has
- 3:20a long and really successful career as a dominatrix.
- 3:23So tech is not my first language and business
- 3:27is not my first line of work. So everything I've
- 3:32achieved up to now has been hard fought. We've
- 3:35made mistakes and we've done things wrong. And
- 3:39I'm proud to say that we've got this far and
- 3:41I'm really excited to show it to you. So what
- 3:43I'm going to go through is from the very beginning,
- 3:46what happened to where we're at now. And then
- 3:49in the middle of that, we'll go through a bit
- 3:50of the techie side of things as well. So everyone
- 3:53can see my screen. I'm going to click through.
- 3:55So company is called Image Angel because I just
- 3:58needed a minute. I needed someone, I needed someone
- 4:01to help me. Because in 2017, my world completely
- 4:06shattered. I was doing amazing stuff in my line
- 4:10of work and winning awards and being on magazine
- 4:13covers and enjoying the successful career that
- 4:17I fought so hard to achieve. But under cover
- 4:20of that, I was battling with this secret shame
- 4:23that my content was just everywhere associated
- 4:26with my name, my place of where I live and where
- 4:31I grow up. And there was no way to stop it. There
- 4:35was no protection. There was no safety net and
- 4:38I got to the lowest point, the absolute worst
- 4:42moment of my life, really. I look back at it
- 4:44and I think it could not have got worse up to
- 4:47a point that I thought I am worth less. I don't
- 4:55need to be here anymore, almost. It was a sense
- 5:00of just exposure and helplessness and I had nowhere
- 5:05to turn. I did go to the revenge porn helpline
- 5:10and I asked for some help. But because my work
- 5:14was commercial, I was considered unable to be
- 5:18helped. And this is happening all the time to
- 5:22so many people and more and more. Unattributed,
- 5:26fraudulent, leak, fake. It causes irreversible
- 5:30damage. to an individual, but also just to truth
- 5:33and trust and your sense of self and belonging.
- 5:37And so I started looking at the world and I'm
- 5:40like, what can I possibly do as one woman who
- 5:44does not have a technical background? What can
- 5:46I do? So I wanted to focus on huge impact with
- 5:53nothing. So I started with no money, no experience,
- 5:56but I also wanted people to be protected without
- 6:01So I built a system that could potentially change
- 6:07the way harm is happening in these spaces in
- 6:10a way that could prevent. And I'm looking at
- 6:12all these things and I need to shape my company
- 6:15in a way that is going to survive the future.
- 6:19But like, what the hell is that? What does good
- 6:22cutting edge tech look like? It looks like all
- 6:24of these things. And I'm looking at the world
- 6:27and I'm thinking, Prevention is everywhere in
- 6:30that we prevent an image being displayed through
- 6:34hashing. And we tell people, don't express yourself
- 6:38online and don't go online and don't be part
- 6:40of this digital world because of the possibility
- 6:44for things that are bad to happen. So I thought
- 6:48I must be able to do something. And it took me
- 6:54a good few years to come up with the correct
- 6:57solution. Because as policy is evolving, as GDPR
- 7:02is evolving, as tech is evolving, things became
- 7:06possible and then impossible. And I actually
- 7:11found a solution, which was forensic watermarking.
- 7:13And I genuinely had the thought of, oh, I can
- 7:17do that. Like it's hard. So I'm really pleased
- 7:22to be able to say that I have developed Image Angel.
- 7:25So it is not perfect, but it's a start. It adds
- 7:30invisible robust data in a private and cheap
- 7:35way into media. So live broadcast, imagery, and
- 7:41also video. And it makes the risk of leaking
- 7:46something outweigh the reward. And throughout
- 7:51this presentation, I'm going to explain things
- 7:54to you guys in the sense of, as a consensual
- 8:00adult content creator, selling consensual content
- 8:04to a consensual buyer. That's my viewpoint. It's
- 8:08my origins, although this can be applied to so
- 8:13many different places. And I would love it if
- 8:16you could... Think of it from your perspective
- 8:19and see how it would work in your setting and
- 8:21what the barriers might be or what the changes
- 8:23may be. So Image Angel is more than a technical
- 8:29integration. It is a software as a service, but
- 8:32it's also a network where creators, regulators,
- 8:38abusers can see that this platform doesn't tolerate
- 8:41abuse. This place uses this tech, which means
- 8:44it doesn't tolerate anything. being used to abuse.
- 8:49We also have a network where people can seek
- 8:50help and then give platforms the data so that
- 8:53they can act upon that verified data. So I see
- 8:56it as not tech. It's a revolution. It's me using
- 9:00my harm for the good of the world. And I also
- 9:03see myself as a little bit like Liam Neeson.
- 9:06I will look you, I will find you, and there will
- 9:09be consequences. Because when we do something
- 9:11and there are no consequences, it feels powerless.
- 9:16So that's what I wanted to do is give the power
- 9:18back a little bit. But enough about the background.
- 9:21I'm going to go into how it works and I want
- 9:24to show you what it looks like and what our system
- 9:26does and how it's built. So we're going to use
- 9:30this beautiful little pup. This is a AI image
- 9:34of a pup. And this is the data that is embedded
- 9:38into the image. This is what the computer would
- 9:41read. You as an individual would see it. as it
- 9:45is when it turns over. You'll see it as it is
- 9:50without any data embedded in there. But it is
- 9:56essentially adding super pixels into the image,
- 10:02overlaying it on top, depending upon the system
- 10:05structure or injecting it into the image itself.
- 10:09Those pixels contain data. They rotate. They
- 10:14change. And we also have the ability to change
- 10:16the density and intensity. We also have the ability
- 10:20to adjust the amount of data that it holds. But
- 10:24at the moment, 64 bits is our limit. And what
- 10:29happens then? So you're joining a platform, you're
- 10:35loading media and somebody comes along and screenshots
- 10:38it, which is so common. And they may edit it.
- 10:43They may add something to the top of it. They
- 10:46may scribble all over it. They may print it out.
- 10:48They may invert the color or they may change
- 10:51it, prop it. You will eventually find that out.
- 10:57That's how this world works. We always find content
- 11:02out there somewhere in the world. And at that
- 11:05point, that's where I was. that, oh my God, what
- 11:10the hell am I going to do moment? And I felt
- 11:13so powerless. And when I was creating this product,
- 11:17I needed to be able to say, even if nothing was
- 11:20possible, something must be possible. Even if
- 11:24there is no legal recourse or no change that
- 11:27can be made, it's important at this point to
- 11:30be able to go somewhere and get the help that
- 11:33you need, even if that is just a couple of lines
- 11:36of advice. in one centralized place because you're
- 11:39panicking. So we created the ImageAngel report
- 11:43and support system. So as well as creating a
- 11:47system that injects data into imagery, we also
- 11:50created a system where you could get the help
- 11:52that you needed if you needed it. So at that
- 11:55point, you would find your image out there in
- 11:56the wild and you would report it to us. And there's
- 12:00lots of people that ask the question, But how
- 12:04do I find where that piece of information is?
- 12:09There are numerous ways of finding data and information
- 12:14out there in the world. What I didn't want to
- 12:17do was recreate a system that already exists.
- 12:20So I've created the two parts of a system that
- 12:23compliment the tech that already exists in finding
- 12:26and tracking content out there. We are developing
- 12:31a system. But I will go into that a bit later,
- 12:34but we'll search the internet for watermarked
- 12:38imagery. So then what happens? You submitted
- 12:41that piece of information to ImageAngel. You
- 12:44could report it to an NGO. You could report it
- 12:47to the platform itself. You could report it to
- 12:50policing or anywhere that works in conjunction
- 12:52with ImageAngel. That report will come to us.
- 12:55We then triage it. We take that report and we
- 12:59see how serious is this. Is this life -threatening?
- 13:01Have you reported to us that this is danger status,
- 13:05in which case it would go right to the top of
- 13:06the pile? Or is this just, I'm a bit annoyed,
- 13:09my commercial work has gone missing and I'm frustrated?
- 13:13So we like to find out what's going on, what's
- 13:15the situation, so we can really triage. Then
- 13:18we extract the data that's in there. So there's
- 13:21two watermarks that are in there and they both
- 13:24contain a small piece of information. One contains
- 13:28the platform data. So that was where you uploaded
- 13:31it as the user. And in this instance, let's say
- 13:35it's a place called Fan Vault. Then Fan Vault,
- 13:39we contact them and we give that other piece
- 13:41of information that is held within the watermark.
- 13:44And we again, don't know what this is. It's just
- 13:46a random hexadecimal numbers and letters. And
- 13:51we look at that and go, this belongs to this
- 13:53platform. you'll be able to decode that. So even
- 13:57if we wanted to, we wouldn't be able to tell
- 13:59anyone what that data pertains to. All we know
- 14:03is where it came from. And then we advise the
- 14:08platform of Origin what should happen next. And
- 14:11this information is gathered because the user
- 14:15has told us what they would prefer. I'm furious.
- 14:19I'm upset. I need therapy. I need help. I need
- 14:23takedown help. I need this person to be blocked
- 14:27from the system. I need the police involved.
- 14:30Or I need nothing. I don't want to do anything.
- 14:34I just want to let you know that this has happened.
- 14:37So the extracted data is only decodable by the
- 14:40source platform when a verified case is raised.
- 14:43No exposed personnel identifying information.
- 14:46It is just a technical identifier. So it's privacy
- 14:49preserving. GDP are compliant at every step.
- 14:52That is vital. And one thing that we're still
- 14:55working out and something that may come up in
- 14:58questions is, well, how do you verify what a
- 15:01genuine request is? And that's a really good
- 15:04question. And we're still working through that.
- 15:06A genuine request we would imagine comes straight
- 15:09from a platform or policing and we can verify
- 15:12those. But sometimes you might get a request
- 15:14that comes in from an individual and that needs
- 15:16to be treated with care. But how do we make sure
- 15:19that it isn't an ex -partner trying to find out
- 15:22if someone is broadcasting on a platform or working
- 15:25on, let's say, OnlyFans? So we have to still
- 15:29work on that and build that part out. And then
- 15:32what happens? What happens then? So you've found
- 15:36this piece of leaked information out there in
- 15:37the wild. You've reported it to us and we've
- 15:40identified the person. Well, what happens then?
- 15:42And at this point, there is so many different
- 15:45options available. It could be small claims.
- 15:48It could be that it is against the Sexual Offences
- 15:51Act. It could be a copyright infringement. Is
- 15:54the user in the US? Are they in India? How do
- 15:58the two jurisdictions work together? When you
- 16:01regenerate a piece of material entirely, you
- 16:06regenerate it with the pixels that we also embedded
- 16:10in there. So it is somewhat possible to be able
- 16:15to say the presence of those pixels could denote
- 16:23that that came from a private channel. And of
- 16:28course, It's not going to be possible to use
- 16:31this as an evidence of its in and of itself,
- 16:34but it could stand as supporting evidence if
- 16:39you have the rest of the data to back it up.
- 16:42And of course, we've tested it. We want to know
- 16:45what people think. We don't want to put something
- 16:47out there that people don't want. And the response
- 16:50has been that third party neutrality, lived experience
- 16:54is really important. And this is the kind of
- 16:56perpetrator accountability that creators are
- 16:59desperate for, but no single mechanism is enough.
- 17:03Because when a harm that is this complex is happening,
- 17:08you need complex responses. You need layered,
- 17:11holistic responses. So when used together with
- 17:15things like hash databases, it can provide a
- 17:17safety infrastructure that is both preventative
- 17:20and reactive. But if someone is determined to
- 17:22do harm, then there is nothing that will stop
- 17:26them. It's a regulatory ally, not a big full
- 17:32stop. And then of course, we've asked exactly
- 17:36the market that we want to sell to the content
- 17:38creators on their platforms. What do you want?
- 17:42Does it work for you? And then there's a really
- 17:45important part that I haven't really gone over,
- 17:46which is who pays for this? The last thing an
- 17:51individual needs to do is pay for their own safety.
- 17:55when someone else is profiting from their work.
- 17:59It is vital that creating safety infrastructure
- 18:04is not accessible only to those who know about
- 18:07it or who can afford it. And that is why in creating
- 18:11this tech, we embed directly into the platforms.
- 18:14This also means that the data that is held on
- 18:19an individual doesn't need to be touched by the
- 18:21creator. But if we want to have real -world impact,
- 18:25We need real -world data. So we embed a platform
- 18:28level. They have the consumer data. They embed
- 18:32it into their system. And then we give them the
- 18:35key that unlocks that information. So I hope
- 18:38there's loads of questions. I'm sure there are.
- 18:40And that's it from me. Thank you very much. Thank
- 18:45you, everybody, for your many questions in the
- 18:49chat, actually. And thank you, Madeline, for
- 18:51taking us through that. You touched on this very
- 18:54quickly. We had one question, which I thought
- 18:56we could just kick off with actually. Somebody
- 18:59said, is this a stop NCII .org direct alternative?
- 19:04I was wondering if you could maybe explain about
- 19:07how you could work, for example, with the revenge
- 19:10porn helpline or how does that work? How do you
- 19:14differ from those kinds of services? Yeah. So
- 19:17there's kind of three things. There's forensic
- 19:21watermarking, which this is. which adds user
- 19:24data in there, or it could add origin data, or
- 19:27it could add the price that someone paid for
- 19:29something. It could be used in many different
- 19:30ways. There is hash matching, and that's both
- 19:36exact hash matching and perceptual. Like, is
- 19:39it potentially the same image? And then there's
- 19:42fingerprinting. And fingerprinting is a way that
- 19:44you can then find that exact copy of that thing
- 19:47out there in the world. We work in conjunction
- 19:51with all of these wonderful things. We would
- 19:55love to work more with Stop NCII, but we're completely
- 19:59self -funded and it costs money to do that, even
- 20:02though they're a charity. And to kind of touch
- 20:05on the kind of commercial side of this, I've
- 20:08paid for this development and this tech completely
- 20:11from my own pocket because this is so important
- 20:14to me. charge the platforms because that's when
- 20:19the harm is happening. We don't have a penny
- 20:21to waste and we are super, super, super tight
- 20:25on what we spend our money on. So to work with
- 20:29a charity in partnership with them, we need to
- 20:33be able to fund ourselves to grow big, to be
- 20:36able to prevent this harm happening. And then
- 20:39if they need funding, then we can give it to
- 20:42them because we've earned enough and we've made
- 20:44sure our ecosystem is running. The last thing
- 20:47I want to do is make a profit -making machine.
- 20:50If any profit that we make is going to go straight
- 20:51back into the system, because this is what needs
- 20:54it. And when I was thinking of setting up a company,
- 20:57I was like, how? How do I do this? I don't have
- 21:00any money. I could set up a company that's a
- 21:04charity. I could set up a community interest
- 21:06group. And to do that, I needed to have so many
- 21:13kind of grants and things like that. And then
- 21:16there was also the model, do I sort of do this
- 21:20with fundraising and stuff like that? And I thought,
- 21:22I don't want to pull from the community. Like
- 21:25I need this protection now. I don't need to fundraise
- 21:28and get money in to be able to do this. So I
- 21:32need this now. So I'm just going to pay for it
- 21:34to happen. So we do work directly with Stop NCII,
- 21:39but we are not a partner because we can't afford
- 21:42to pay that fee. That's a long winded answer,
- 21:45hey. No, it's a good one. It's a good one. As
- 21:50you've just mentioned, obviously this is a commercial
- 21:52entity and we've also spoken about how Bellingcat
- 21:57is open source and how, to be honest, a lot of
- 21:59the tech that we do spotlight in this platform
- 22:03is open source oriented. There is some protective
- 22:08tech. that is open source source code for public
- 22:11audit and transparency. For example, Teller is
- 22:14a really cool tool that helps automatically encrypt
- 22:17a reporter's data on the front lines if they
- 22:20are filming within it. Or we have Bellingcat's
- 22:23auto -archiver tool, which uses one -way cryptographic
- 22:27hashing to archive an image, video, or post,
- 22:30even when it's taken offline. But ImageAngel
- 22:32is closed source, so why have you chosen... to
- 22:35be closed source and if your algorithm relies
- 22:38on the source code staying a secret to remain
- 22:41secure, isn't that a structural vulnerability?
- 22:44Like how do you address why it's closed source?
- 22:48It's closed source at the moment because we're
- 22:51using a specific type of tech that we have to
- 22:53pay for because we want to be the best. So we
- 22:55outsource, we buy that and we use it. What we
- 22:58need to do to remain robust against manipulations
- 23:02going forward is get rid of that dependency poly
- 23:06third party and build our own. But we also then
- 23:11have that dependency on us. And what if someone
- 23:14works out how to manipulate that? So what we
- 23:16are now doing is taking a mixture of open source,
- 23:21closed source, and combining each of those in
- 23:25various platforms so that it is harder to then
- 23:28decode in the long run. We're always fighting
- 23:31against the next thing to kind of break and and
- 23:35close the stuff, right? So it's always a constant
- 23:38fight. So what we want to do is just always change
- 23:40that model so it's harder to decode. But also
- 23:44what we do is depending on a platform's risk
- 23:49appetite, vulnerability to this kind of thing,
- 23:52we work with them directly to figure out the
- 23:54best model and method for adjusting the algorithm
- 23:58of the implementation of the watermark. Yeah,
- 24:01it is interesting, the technicality of everything.
- 24:04And I really want to dig into that, but we've
- 24:06had a few, we've still got quite a few questions
- 24:08to work through, so I want to quickly get them
- 24:11in. Someone's just asked how much of this is
- 24:13about risk awareness and accepting that content
- 24:16placed online may be misused versus preventing
- 24:19misuse or responding to it. For example, through
- 24:21techniques like that hex technique. In some cases
- 24:24is understanding and accepting the risks of sharing
- 24:28content, such as having an OnlyFans account,
- 24:31more important than focusing on response. Yeah,
- 24:35I mean, there is a risk to everything. And honestly,
- 24:41everyone says to me, oh, well, what do you expect?
- 24:45You know, having an OnlyFans account, what do
- 24:46you expect? Well, I expect a little bit of decency
- 24:48and privacy. I would never share the images that
- 24:51are sent to me from a user. There is that implicit
- 24:53expectation of being in a private chat that this
- 24:57communication doesn't go elsewhere. And our watermark
- 25:00will only ever apply to content that is private.
- 25:03There's no point watermarking stuff that's out
- 25:06there in the world and free to view. That's not
- 25:09what we do. We're not necessarily here to enforce
- 25:12copyright. If someone wishes to do that, we can
- 25:15support them. The main aim of the company is
- 25:20to prevent harm, and that is with a deterrence
- 25:23method. Adding in data of consumers into the
- 25:27content they watch could be seen as a surveillance
- 25:30sort of, well, that means every piece of content
- 25:33contains my information and what if this and
- 25:36what if that? And the simple answer is if you
- 25:39do no wrong, nothing happens. It is not illegal
- 25:43to share your own content. You can do what you
- 25:46want with that image once you've taken a screenshot
- 25:49of it or a screen recording. You can do what
- 25:51you want, but you shouldn't be doing that. And
- 25:53if you do do a negative thing, you could potentially
- 25:56be caught and understand what the repercussions
- 26:00of that could be anything from just being banned
- 26:03or kicked off that platform, for example. So
- 26:05for example, if you were a content creator who
- 26:09had a private paid for space, like for example,
- 26:13people had to pay to gain content and that content
- 26:16was leaked, would your company help in that sense?
- 26:20Yeah, definitely. If a consumer was talking with
- 26:23a content creator and a content creator threatened
- 26:26to leak or something, their private material
- 26:28in the same regard, that's a private space. Interesting.
- 26:33We've had quite a few questions focusing on whether
- 26:37you've had any success with the big tech platforms
- 26:40adopting your tech or showing any interest in
- 26:44it. I wonder if you can speak about that. So
- 26:49we've had, because of my experience in the adult
- 26:53space, I used my connections and was like, look,
- 26:56I've been working on your platform for XYZ time.
- 26:58Do you think you could use this tech? And mostly
- 27:02they said no. But there was one platform called
- 27:08Friends To Follow who were like, we love this,
- 27:11we want this, make it happen. They've been really,
- 27:15really great in paying. They paid us, they gave
- 27:20us money to put this system in their platform
- 27:23and it's live and it's active and it's protecting
- 27:25half a million users. And that's the kind of
- 27:29good tech I see. I've approached, you could pretty
- 27:35much name every tech platform. We have in some
- 27:38way approached them. I can't say who has and
- 27:42who hasn't decided to go ahead with it because
- 27:43there are some people who are under NDA. The
- 27:50big, big ones. But I guess I can say, can I say
- 27:53I can say if someone has decided not to go ahead
- 27:55with something. Yeah. Yeah. Meta said that they
- 27:57didn't have a problem with stop NCII. So I was
- 27:59like, okay. If you think that, cool. I think
- 28:05there is a certain amount of unwillingness to
- 28:09accept that this is a good solution, partly because
- 28:13of my background. In fact, I had somebody who
- 28:17pitched the deck to, and they said, yeah, you've
- 28:19not done bad for a whore. And I thought, yeah,
- 28:22I've not done bad for a whore. Fuck you. But
- 28:26there are some people who don't like what I do
- 28:28because of where I've come from, I suppose. I'm
- 28:32not a native in the tech land, I don't speak
- 28:34in the same way. So there has been some people
- 28:38who don't want to be anything to do with this.
- 28:41And I've actually experienced quite a lot of
- 28:43negativity from people who are really, really
- 28:47big voices in violence against women and girls
- 28:52spaces. Even though I felt that harm, you know?
- 28:57And I created this to be able to prevent... content
- 29:00from getting out there and leaking and that kind
- 29:03of harm that comes through. So the response has
- 29:07been really mixed. Some places are really keen,
- 29:11others don't even want to talk about it. And
- 29:13that's probably because trust and safety is very
- 29:17different to compliance. Compliance is, oh, I
- 29:21made my system this safe. Look, we made it. Whereas
- 29:25trust and safety is anticipating harm, adding
- 29:27in prevention instead of reaction. I hope that
- 29:31that will change. There's some quite sarcastic
- 29:35shock in the chat about Metta saying no. I am
- 29:40so sorry if somebody said that to you as well.
- 29:42That's disgusting. You mentioned the tech platforms.
- 29:47What about governments? For example, why is forensic
- 29:49watermarking not part of the UK government strategy?
- 29:53They recently declared intimate image abuse a
- 29:55national emergency, yet regulations like the
- 29:57Online Safety Act focus almost entirely on reactive
- 30:00measures, giving platforms 48 hours take down
- 30:02content after it's reported, for example. Why
- 30:04do you think the government and Ofcom in the
- 30:06UK and more globally as well are hesitant to
- 30:10mandate preventative server -side forensic tracking?
- 30:14I wish I knew. I wish I knew. I have been, I'm
- 30:21not going to lie, bugging Ofcom particularly,
- 30:25but also the EU commission. about this tech.
- 30:30I think they don't understand it. If I read,
- 30:32and if you search through Ofcom's guidance and
- 30:36you search the word watermarking, in one of them,
- 30:40probably the most, not the most recent, I haven't
- 30:42checked that one before that, you search the
- 30:45word watermarking, it comes up with one thing
- 30:47referencing, you know, like the old Getty Images
- 30:50watermark, like Getty Images, proof of origin.
- 30:54I don't think they get it. But I have been thankfully
- 30:58invited to teach Ofcom about the tech itself.
- 31:03Now, of course they have to be really impartial
- 31:05and I respect that. I'm not interested in kind
- 31:09of dominating the market and making loads of
- 31:13kind of moves and money. I want people to be
- 31:16safe. I don't care if there's another company
- 31:19out there that can do this. Go ahead and do it.
- 31:23I need Ofcom to back me, but I don't feel like
- 31:28I have that. And I don't know why. There you
- 31:31go. There's your call to action guys. If you
- 31:33want to make an open source product that takes
- 31:36the image, go wild. There is someone else who
- 31:45does this. In fact, there's a couple of companies
- 31:46that do this and I've raised with them that I
- 31:50would love to collaborate. I see a world in which
- 31:56we decode someone else's watermark and we can
- 31:59say, hey, your content came to us. It actually
- 32:01belongs to you. You know how to decode that.
- 32:03And we work as a collaborative approach. It wasn't
- 32:08welcomed. I tell you what, I'm surprised. I really
- 32:14am. I grew up in a world where I thought when
- 32:18I grow up, it's never going to be like that.
- 32:20And I've grown up and I've realized It's capitalist
- 32:24hell out here. I hate it. I hope I can weave
- 32:29a little path out here and do things a bit differently.
- 32:32Well, we're all about collaboration here, hence
- 32:36this space. And we do have, by the way, as I'm
- 32:39on this subject, we do have a space in the server.
- 32:42If you want to contribute to Bellingcat tools,
- 32:44you can. It's called Tech Contributors. I'm sure
- 32:47one of the mods can link it. And then we also
- 32:50have a space called hashtag tool development,
- 32:52where you can work together building tools, open
- 32:55source ones, hopefully. But you can have a little
- 32:58chat in that space too. Somebody's asked, do
- 33:01you think that there is some merit in taking
- 33:02this product to T &S terms and services conferences
- 33:06that happen around the globe, just as a way to
- 33:09show the teams firsthand on how this can add
- 33:11value for users on their platforms? Have you
- 33:14ever thought about doing that? Yeah, we are actually
- 33:17presenting at Trustcon in the next few weeks.
- 33:23It's tough though, because it costs money, doesn't
- 33:25it, to go to these things. And we're only earning
- 33:27a little bit of money from the platforms that
- 33:28we're installed on. Look, it all kind of comes
- 33:31down to money, doesn't it? It sucks. I hate it.
- 33:33But yeah, it's something we're doing. It's something
- 33:36we're doing. The thing is I'm only as powerful
- 33:38as my team and as the presence of people who
- 33:44are kind of in the movement with us. We're only
- 33:46as powerful as those voices. And we're a minority
- 33:50voice. Like sex workers want this. Sex workers
- 33:53are a minority voice. No one listens, right?
- 33:57So to get where we've got to, I'm damn proud.
- 34:02And to be at TrustCon is pretty cool. But it
- 34:05all takes time or takes money, you know, effort
- 34:09and... I swear I have not seen my husband for
- 34:13a couple of bloody months. Poor guy. He's such
- 34:17a wonderful guy. He's really supported. This
- 34:20wouldn't be where it is if it weren't for him.
- 34:23In regards to money, you were talking at the
- 34:27beginning as to why you keep the service free
- 34:33for those that have been subjected to online
- 34:38image abuse. But platforms like Telegram, for
- 34:42example, are primary battlegrounds for information
- 34:45warfare. I mean, a lot of the Bellingcat investigations
- 34:49start in places like Telegram, but they fiercely
- 34:53guard their architectural privacy and refuse
- 34:56server -side moderation. If a platform refuses
- 34:59to integrate your API, is there a way that journalists
- 35:04or client -side solutions could work? For example,
- 35:11allowing journalists or investigators to use
- 35:17the tech. Is there a way that it could work on
- 35:20the client side where it doesn't necessarily
- 35:22have to have the platform's approval? I don't
- 35:27know. I'm going to say I don't know because I
- 35:30don't come from a tech background. We have an
- 35:34amazing tech team working with us. they are so
- 35:37much better place to answer those questions.
- 35:39But if a platform does not want a system like
- 35:44this, that's okay because we do have to have
- 35:50spaces and places that can be anonymous and you
- 35:53can go via VPN. You need to be able to do these
- 35:55things to be critical and to share things that
- 35:59need to be shared in kind of anonymous places.
- 36:03This tech is not about changing the whole infrastructure
- 36:08of the internet. It's about changing the mindset
- 36:12of people who view content that is privileged
- 36:15and private. How that happens in a space that
- 36:19won't do that is a tech question I cannot answer.
- 36:23And there's probably people in the chat that
- 36:24can answer that other than I can, for sure. Yeah,
- 36:29I think there's a cripple of questions about
- 36:32the technicalities of all of this. I'd love to
- 36:36dig into that a little bit more. Like for example,
- 36:41how do you ensure that if somebody's computer
- 36:47is hacked, for example, and these are leaked,
- 36:50how do you prevent those kind of false positives
- 36:52from happening? because as you mentioned, the
- 36:55kind of watermark gives the user ID. So has there
- 36:58been cases where that's happened or is there
- 37:01a way to deal with that kind of situation? So
- 37:05there's been no cases yet that have happened
- 37:07in that way. We've worked with a charity called
- 37:11Refuge to try and understand the way this could
- 37:16be used in a negative. What could happen to try
- 37:20and attack this? And we know that it is perfectly
- 37:24possible that someone will use this method to
- 37:27try and catch someone out to find data on someone
- 37:29and create cases that are malicious or get content
- 37:33taken down. And that's something that we still
- 37:35need to figure out ourselves. How do we verify
- 37:38the person who is giving us the information is
- 37:42the person they say they are? Is that their content?
- 37:44What information are they trying to find out?
- 37:45So in the first instance, what we do is if someone
- 37:48comes directly to us, as a company, we give them
- 37:53as much information as we feel is right at that
- 37:55time, which is, thank you very much for submitting
- 37:58this. We're looking into it. And that's all we
- 38:01do at that point. We then have a little look.
- 38:04We see what's in there. If there is no data,
- 38:07we don't say there is no data. We just say, we
- 38:11can't do anything with this at this time. We
- 38:13don't say whether or not there is, or there isn't,
- 38:15or we found anything. We just sort of say, you
- 38:17know, sorry, we can't help you today kind of
- 38:19thing. If we do find data in there, we don't
- 38:22tell them that either. We then go and find the
- 38:26information on the platform that it has been
- 38:28flagged so that we then know, is this a legitimate
- 38:31request? Because we know that there will be people
- 38:34who are trying to find content that is not watermarked
- 38:37so that they can then use that in a space for
- 38:40abuse or even find content that is watermarked.
- 38:44There's also that privacy around I'm a user.
- 38:47I've had my content leaked. I don't want to be
- 38:49seen as troublesome for the platform that I'm
- 38:51earning my money on. I'm already feel like I'm
- 38:53a target or I'm shadow banned. I don't want to
- 38:56be causing this company more stress and them
- 38:59think it's a me problem. So we also don't tell
- 39:01the platform who that end user is. It's a very
- 39:05complex dance between data and finding the right
- 39:09information without giving the right information
- 39:11over and it's complex. But we have a person entirely
- 39:15dedicated to both GDPR and then we have that
- 39:19support service who works on making sure that
- 39:22there is no harm happening in these kind of edge,
- 39:25weird, odd cases. In regards to, you mentioned
- 39:29at the beginning that, you know, if it's screen
- 39:31-shotted or posted across platforms, this forensic
- 39:36watermarking still tracks. Yeah. However, we
- 39:41see a lot of manipulation. If somebody mentioned
- 39:45earlier that one manipulation that's done is
- 39:48video inversion, for example, horizontal flipping.
- 39:52We also obviously see... deep fakes now as well
- 39:55of people taking original video of, for example,
- 40:01a professional sexual performer and putting a
- 40:03different face on that. In those cases, is the
- 40:09watermark strong enough to follow that kind of
- 40:12paper trail? Is there circumstances, for example,
- 40:16with AI manipulation where that trail could meet
- 40:21a dead end? Absolutely. If you're determined
- 40:26enough, you can break this. There is nothing
- 40:30that this will survive through if you are determined
- 40:33enough. We know people who are determined to
- 40:36do harm will do everything they can. This is
- 40:40designed for casual, I'm going to call it just
- 40:45opportunistic and silly leaks. People being stupid,
- 40:51opportunistic and thinking they might get away
- 40:54with it. This is not for creating a system that
- 41:00is unbreakable. It's impossible to create a system
- 41:03that's unbreakable. I would say that there's
- 41:06a spectrum of it. The more you do to an image,
- 41:10the less likely we are to be able to extract
- 41:13data. But also the more information we put in
- 41:18there, the less likely it is that we can extract
- 41:22it. because there's a payoff between how invisible
- 41:26it looks and how much data we can embed in there.
- 41:30So it's a dance again of how much information
- 41:35can we put so we can still identify you but you
- 41:37still can view the image of disruption versus
- 41:40how much manipulation can happen to it and every
- 41:43case. when it comes down to anything outside
- 41:47of that very generic crop flip minor edit will
- 41:53then be an individual case that we have a human
- 41:56review. And that's so important because we also
- 42:00don't want false negatives. Absolutely. There's
- 42:04always a hole in every method, in every tech.
- 42:09If you are looking to the future, In regards
- 42:12to this space, in regards to protective tech,
- 42:16what is the biggest technical or cultural hurdle
- 42:19you predict you might face in making this technology
- 42:24an uncompromised standard? What are the biggest
- 42:31obstacles facing protective tech in this space?
- 42:35You're talking a bit of quantum and we're all
- 42:37up a creek without a paddle, right? So if we
- 42:40were to go that far, we're done for, we're cooked.
- 42:43But in short term, I think the biggest hurdle
- 42:47is adoption. If, for example, every, let's say,
- 42:53easily findable space on the internet that our
- 42:55friends and family use, uses a system like this,
- 42:58then it will start to become colloquial knowledge
- 43:00that you don't do that. And then the more people
- 43:03who get kind of caught for doing negative things
- 43:06in spaces like this, it will kind of become known,
- 43:09oh yeah, that was like, did you know what happened
- 43:12to him? And it becomes kind of that colloquial
- 43:14understanding that you just don't do that. And
- 43:16that starts to create a social change of, we
- 43:19respect people in this kind of space. So adoption
- 43:23is probably the biggest thing. And I need people
- 43:27to want it. You know, is this something that
- 43:30people even want? Do we need people having data
- 43:35injected into imagery? Our research shows that
- 43:37it does, but if the wider world doesn't want
- 43:40to adopt this, then it doesn't go anywhere. Yeah,
- 43:44absolutely. I mean, you need to make sure that
- 43:47you're building it for the right people, right?
- 43:48But you did mention that you were saying that
- 43:50you, at the beginning, did ask that exact question
- 43:54to those that have been subjected to this abuse
- 43:57previously. Sarah, one of our amazing mods in
- 44:01the chats just asked, regarding humans reviewing
- 44:03complicated cases, are the people doing so compiling
- 44:07anything they're learning as they review these
- 44:08cases? EG, if they are able to see a pattern
- 44:11in the way Image Angel is outsmarted, is that
- 44:14being documented for future improvements? Yes,
- 44:18it is. Yeah. Everything that we do, we need to
- 44:20keep a record of it because then it improves
- 44:21our system. But also everything we do to our
- 44:24imagery also makes the kind of the anti, the
- 44:28people who were kind of trying to attack it stronger.
- 44:30So we're always, it's just cat and mouse. We're
- 44:32always trying to adjust and change. And that
- 44:34will always be the case. And I think that's great
- 44:37for innovation. We have ideas. for changing the
- 44:42way this works, for adding in extra data, for
- 44:45additional products that could support systems
- 44:47and take downs and searches and added in signals.
- 44:51It's all a part of development and they all will
- 44:53come in time and it all has to be for the right
- 44:55reasons. But yeah, everything. to be like, how
- 44:58are we learning? But also importantly, it's really
- 45:02vital that the people in that human in the loop
- 45:05don't view this kind of information and content.
- 45:09So we have to go through a really stringent process
- 45:12of keeping as much of that content blurred and
- 45:15blocked out as possible. So even if we do need
- 45:18to take that piece of information, how do we
- 45:20make sure that that user is protected and safe?
- 45:23And also that your nudes aren't being seen by
- 45:25some tech guy. We've had a couple of questions
- 45:29about court cases. I don't know if this tech
- 45:32has ever been used in a legal case, but they've
- 45:36asked if this is used in court, does having a
- 45:38watermark have positive effect on the outcome?
- 45:41Similar to having a lock on a property, how it's
- 45:43approving that the thief saw it and knew he is
- 45:45breaking in. I feel like regardless of the technical
- 45:47part being a hundred percent, this could be a
- 45:49big benefit in the legal part of the issue. Have
- 45:52you interacted with that? legal procedure world
- 45:56yet? Because obviously that's something that
- 45:58a lot of we do with war crimes, for example,
- 46:01war crime imagery and manipulated imagery. We
- 46:04have previously provided information for legal
- 46:10processes. So is that something that you kind
- 46:13of work with on your side? Yeah, and it's something
- 46:17that we want to do more of. The problem is, again,
- 46:20resources. We're still a startup. We're still
- 46:23scaling. What we do is test, ask, nudge everyone
- 46:30all the time. What we know is that data embedded
- 46:35in imagery is not unusual in court cases. There's
- 46:40metadata, there's even information like a username,
- 46:44a handle that could be kind of linked back to
- 46:47the person who created the deepfake. And there
- 46:49have been cases that have been successfully prosecuted
- 46:50in those cases. With the Image Angel, not yet.
- 46:54But hopefully that'll be the case. And then when
- 46:56it comes to AI manipulation and how much data
- 46:58is still left in there, it is possible to still
- 47:01see information held within an image. Only we
- 47:05can't read the exact data, but the presence of
- 47:08the regeneration of those pixels proves that
- 47:11there was an original image that came from that
- 47:13non -conceptual space. and just the presence
- 47:16of it alone. Even though we cannot record or
- 47:19find enough robust data to pinpoint exactly who
- 47:23it was, the presence of it there, even visually,
- 47:27denotes that there was something there and that
- 47:30can go along to supporting. I know a lot of people
- 47:33are going to say, oh, well, I had my device stolen.
- 47:35I was packed. The biggest defense when somebody
- 47:39is being accused of intimate image abuse or deep
- 47:43fake creation, for example, is I was hacked,
- 47:46I had my phone stolen. That may very well be
- 47:48the case and life is like shit. Things happen,
- 47:53bad things happen. What we could also do in this
- 47:57evidence gathering is prove a kind of a pattern
- 48:03of abuse and we may be able to track somebody's
- 48:07abusing across a matter of months or weeks. So
- 48:10it is quite possible that someone had the device
- 48:11taken and stolen and accessed then, but would
- 48:14that pattern continue? Who knows? It's all adding
- 48:17evidence to support people. If you are interested
- 48:22in this particular subject, I definitely recommend
- 48:25listening back to our last stage talk with the
- 48:28CNN team who spoke about some of the work that
- 48:31they've done, uncovering communities who were
- 48:33sharing, online communities who were sharing
- 48:36intimate imagery. that was either filmed or was
- 48:43leaked. They were into detail about the accountability
- 48:47there and how they conducted those investigations.
- 48:50It's definitely worth listening back to and those
- 48:52links are in the chat and also will be in the
- 48:54description of the podcast if you're listening
- 48:55back. And thank you, Sarah, for pinning that
- 48:59again. We've only got a couple of minutes left,
- 49:03so I'm going to have to squeeze in the last few
- 49:06questions now. We had a question earlier. saying,
- 49:09has there been any resistance in adoption from
- 49:12the OnlyFans sex worker space? I figured that
- 49:14would be an audience open to watermarking. You
- 49:16mentioned that you consulted a few people from
- 49:19that space, but has there been any resistance
- 49:23in adoption from those spaces? No, all of my
- 49:28friends I would say now are sex workers. I had
- 49:31such a wonderful community around me, sex workers,
- 49:34creators, you name it. artists of all sorts of
- 49:38walks of life. There has been no resistance to
- 49:41it. The only sort of thought was, oh, wait, wait,
- 49:44wait, are you putting my data in it as a creator?
- 49:46You know, can that person track who I am? But
- 49:48no, that's, you don't need to prove who you are.
- 49:50It's your face or it's, you know, you've got
- 49:52the original. No, and even in the consumer space
- 49:56as well. A lot of people I've been speaking to
- 50:00are consumers of pornography and adult content
- 50:02and social media and stuff. Nobody has said,
- 50:07I don't want that. And I think it's because we're
- 50:10used to data doing things. We know that if we
- 50:14are responsible for something online, it can
- 50:16be traced back to us. Our search history is,
- 50:18you know, it's possible to link these things
- 50:20back to us. We're never entirely anonymous. I
- 50:23think it's not beyond our understanding that
- 50:25this is a thing. And the only people who would
- 50:27potentially be opposed to it, and I welcome opposition,
- 50:31are people who would want to do something negative
- 50:33with that. The one thing that we need to make
- 50:36sure is that this cannot be used in a negative
- 50:40way. And that's something that will always be
- 50:43evolving. Yeah, absolutely. I would potentially
- 50:46challenge your notion that everybody knows how
- 50:49much information is behind each thing that they
- 50:52upload. The stuff that we find in metadata is
- 50:56sometimes hilarious or source code of websites.
- 51:00If you want a reference, go check out our work
- 51:02on the Kinahan Cartel and see kind of what professional
- 51:07criminal organizations are sharing online. You'll
- 51:11see the people's understanding of what is being
- 51:15shared. Right, we are nearly end of time. So
- 51:18I wanted to ask, as this talk is really about
- 51:22building protective tech. What would be your
- 51:24advice to people who perhaps are listening and
- 51:27want to go out and build something similar themselves?
- 51:30What would your advice be to those who are looking
- 51:33to develop protective tech? I think it's important
- 51:38that we stop and pause and think, who am I building
- 51:40this for and what are the potential negatives?
- 51:42Well, how could this be manipulated and used
- 51:44in a negative way? If we're going to build good
- 51:46stuff, we have to make sure that it's truly good.
- 51:48It really, really is good and not just good in
- 51:50our minds or in our projection of our world.
- 51:53Is this good for the global South? Is this good
- 51:56for old people? Is this good for more people
- 52:00than I can conceive? And that's what we try to
- 52:04do is stress test in as many different environments
- 52:07as possible. But also I would say, don't do what
- 52:11I did and wait. I took this idea to hundreds
- 52:15of people, I would think, over the time, you
- 52:17know, hundreds of platforms like, oh, I've got
- 52:19an idea, would you like to develop this? Everyone
- 52:21said, no, it's not possible. And I thought, no,
- 52:25it really is possible and it's really important.
- 52:28And being ready to do that was a choice. I had
- 52:31to make that choice. Like I'm never, you're never
- 52:32going to be financially ready, emotionally ready.
- 52:34Your home life is never going to be settled enough.
- 52:36If you want to go out and make a difference,
- 52:38you got to go and grab it. It's a choice. You
- 52:41need to make the choice. And if you're not feeling
- 52:45inspired enough already, as I said, do go look
- 52:48at hashtag tech development in here because there's
- 52:51lots of people trying to do the same. We had
- 52:54a question earlier and I promised them that I
- 52:56would ask this towards the end. Are you hiring
- 52:58and or supporting contributors to this project
- 53:01as you're building it out? Because I think you
- 53:05have a lot of sans in this space now. I need
- 53:10fans. Every day I struggle with like, am I doing
- 53:13the right thing? I've got imposter syndrome to
- 53:17the max. Are we hiring? We're not directly hiring
- 53:21because we are growing slow and growing well,
- 53:24but we want to build a database of people who
- 53:27want to work for us so that when we have those
- 53:30money and that time, we give back. I want to
- 53:33hire the best brains. If you want to work for
- 53:35us, please come and say hi. If you have a thought,
- 53:39I'd even love to know, what do you hate about
- 53:41it? What do you challenge? What do you want to
- 53:44do? What do you want to hear and say? I'm putting
- 53:47my heart out to the Bellingcat community. Just
- 53:51tell me what you think. I'd love to hear your
- 53:53feedback. And how can they do that would be my
- 53:59follow up question. Probably Instagram is probably
- 54:02the best way. We also have TikTok. We have LinkedIn.
- 54:07We have an X, but I don't like to use it. And
- 54:12all of it will be linked on our website. But
- 54:14if you search Image Angel with an underscore
- 54:16at the end, you'll likely find us. I'm just dropping
- 54:20that Instagram link into the chat now, and I'll
- 54:23include it in the description of the podcast.
- 54:26Finally, finally, finally, as I mentioned at
- 54:27the start, Bellingcat doesn't endorse individual
- 54:30companies. So I wanted to ask you if there's
- 54:33anybody else in this space that you would want
- 54:36to shout out? For example, is there any other
- 54:39resources that have been vital to your building
- 54:42of this project or any other companies that are
- 54:45doing something cool that you think should be
- 54:48on people's radars? Oh, my God. Oh, my God. Oh,
- 54:53my God. That's completely blank. That probably
- 54:57is. I would say Not Your Porn are a really good
- 55:03group of people. Not Your Porn is a group that
- 55:07victims, allies, and they just jump in and try
- 55:11and lobby and get change to happen. It's a not
- 55:16for profit. They are volunteer run. They are
- 55:20awesome. Go and support Not Your Porn. I just
- 55:24dropped their LinkedIn in the chat there. I didn't
- 55:27have, I couldn't find the website that quickly,
- 55:29but it is there. And hopefully you can explore
- 55:33that and I'll make sure to find the proper link.
- 55:36Oh, thank you people in the chat for the description
- 55:39of the podcast. All right. We are about time.
- 55:42Thank you so much Madelaine, for your time today
- 55:45and for going through this amazing tech. It's
- 55:49really fascinating to hear a little bit about.
- 55:51your solution for this massive problem. And hopefully
- 55:55it's inspired others in the audience to think
- 55:58about how they can shape their own solutions
- 56:01for this issue. A lot of love in the comments.
- 56:05In two weeks time, we will be back with another
- 56:08stage talk. It is not going to be on this subject.
- 56:11We have covered this subject quite a bit over
- 56:13the last few weeks. We are going to be moving
- 56:15to environmental. investigations in two weeks
- 56:19time. We have a fantastic environmental journalist
- 56:22coming in who recently won environmental journalist
- 56:26of the year awards, and they're going to be coming
- 56:28in to talk about how you can do environmental
- 56:33investigations digitally. So that's what the
- 56:36next stage talk will be about. But for now, thank
- 56:38you again, Madelaine. Thank you for having me.
- 56:42Thank you for listening to the stage talk. If
- 56:45you'd like to catch a stage tour live where you
- 56:48can ask the guest questions, join the Bellingcat
- 56:51Discord server by visiting www .discord .gg slash
- 56:57Bellingcat. The music you've heard is titled
- 57:00Dawn by Newer Self and is courtesy of Artlist.