Disclaimer: Asymmetrical Haircuts is produced as a podcast, meaning it is meant to be listened to and not read. Because of this, we recommend that you listen to the episode while reading, because the written word does not do justice to the emotion or tone used by our speakers. However, because we recognise there might be bandwidth issues or you might be using a hearing aid, we have provided written transcripts for all our available episodes.

Episode 161 – Seeing Isn’t Believing: Deepfakes and Evidence with Yvonne McDermott Rees and Anne Hausknecht 

[INTRO TUNE]

 Asymmetrical haircuts Justice. Update with Janet Anderson and Stephanie Vandenberg in partnership with Justice info.net. 

Janet: 00:38  Hi, Steph.

Stephanie: 00:39  Hi, Janet.

Janet: 00:40  So we are in the middle of doing a whole series about artificial intelligence, and we thought it would be interesting to go back to somebody who we’ve had on before about evidence, particularly.

Stephanie: 00:54  Yeah, we had Yvonne on to talk before about how judges were learning about the use of social media in trials, and I think we then talked about Bellingcat and we had the Judge Joanna Korner say something about that.

Janet: 01:08  Yeah, that was fun. I mean, it showed how sceptical a judge can be, but also how an expert witness can kind of help them along and explain new and unfamiliar stuff and really get them to accept user-generated evidence. 

Stephanie: 01:22  But we’re now living in a world of AI slop on social media, and we kind of see everything, and nothing is believable, and judges are also human beings. So, how will reality be assessed in courts and in international trials from now on?

Janet: 01:38  Our trusty source, as mentioned already, for user-generated content or – I don’t know what I was writing there, “user generated concrete” no – user-generated content and evidence is Yvonne McDermott Rees. Hi, Yvonne.

Yvonne McDermott Rees: 01:56  Hi, Janet. Hi, Steph. Great to see you.

Stephanie: 01:59  Yvonne is now at Queen’s University Belfast School of Law, and she’s still doing experiments with courts and lawyers, and jurors even, God forbid, and testing how people understand standards of proof and the limits of expert evidence. And we will put a link to one of these in the show notes, where there is a great team of prosecutors and defence, and former ICC judge Howard Morrison, put in the hot seat, where they’re testing people’s reactions to pattern-recognition evidence. 

Janet: 02:27  I don’t even know what that is, Steph. I mean, I wrote it down. It’s copy-paste. So maybe we’ll come back to that.

Stephanie: 02:34  Let’s put a pin in that and ask it later. But it’s part of the “TRUE Project,” where Yvonne is the principal investigator, and I’m sure she will tell us much more about it later. 

Janet: 02:45  And this time she’s joined by Anne Hausknecht. Hi, Anne.

Anne Hausknecht: 02:49  Hi, Janet. Hi, Steph. Thanks for having me on.

Janet: 02:52  And Anne is a postdoctoral researcher at the Oxford Institute of Technology and Justice. She’s also on the same ‘TRUE Project,’ and in some of the work she’s been doing with Yvonne, she’s been exploring the impact of deepfakes on trust in user-generated evidence and how that works in accountability processes for human rights violations. And her research focuses primarily on the use of user-generated evidence in domestic prosecutions of core international crimes. And Anne is being supervised in a PhD by Yvonne. So let’s see how much they agree with each other. Over to you, Steph.

Stephanie: 03:28  Well I wanted to ask one definition thing before we dove into the deep end. What is the definition of “user-generated evidence?” Because “generated” to me kind of implies manufactured. So are we talking about, you know, people who make videos of alleged violations, or are we talking about people who generate some kind of content via, for example, reconstructions of things that happen?

Yvonne McDermott Rees: 03:54  We define user-generated evidence as information, so this is images, video, text, audio, created by – so that’s the generated part – by ordinary people basically, not professional investigators, on their own personal digital devices. And what makes it evidence is its potential use as legal evidence in accountability processes. So we’re talking really about the eyewitness who records a video, someone who takes an image, not someone who decides, “I’ll create an AI model or an AI reconstruction.” That would be something different.

Janet: 04:33  OK, but I think we will also include a bit on reconstructions, because I think that’s interesting, sort of where the line lies between one thing and the other, because one can lead into the other.

Yvonne McDermott Rees: 04:45  Absolutely. In the Al Hassan and in the Al Mahdi trials – I know you’ve had people talk about this before – the digital reconstruction that was created by SITU Research, an architecture firm in New York City, and that basically used user-generated evidence. So, videos they found on YouTube, satellite imaging and things like this, and put it all together into a digital reconstruction. So, yeah, they’re closely related.

Janet: 05:12  I think there’s an opportunity for a Stephopedia here.

Stephanie: 05:15  I am diving in for the Stephopedia. Yes, the Al Mahdi and Al Hassan trials are the ICC cases that have to do with the destruction of religious sites in Timbuktu by Islamists who took over that city in Mali. And so, Al Mahdi was specifically destruction of tombs and religious sites, and Al Hassan was that and also floggings and basically terrorising the population of Timbuktu. And they had reconstructions of some of the sites where crimes happened.

Yvonne McDermott Rees: 05:47  Just one last point to add on why we use the term “user-generated evidence,” as opposed to “open-source evidence” or something like that is, as you know, previous research I carried out with my colleagues really looked at open-source research or open-source evidence, but increasingly we’re finding this kind of content isn’t posted online. It’s sent directly to investigators. It’s captured using maybe dedicated apps like the eyeWitness to Atrocities app, which was created by the International Bar Association. By today, The Office of the Prosecutor has got this platform where anyone can upload relevant evidence of international crimes. So we didn’t want to exclude that closed-source content, because actually, when it comes to evaluating it and considering its credibility, the questions are the same, regardless of whether it’s open-source or closed-source. That’s why we’ve broadened out to user-generated evidence.

Janet: 06:42  OK, I’ll start using that from now on. But, I wanted to also just ask this, I mean, crazy philosophical question, maybe, but it comes up for me again and again. Are we any good as human beings about distinguishing between factual stuff and fictional stuff? How do we know what we’re watching is reality? And obviously, I don’t want to spend the whole podcast doing philosophy, but, what can you, how can you describe that to us?

Anne Hausknecht: 07:09  Well I could start, so we’ve been doing some research. I’ve got a wonderful colleague, Ruben, who’s been looking more from a psychology perspective at how good are we actually at spotting deepfakes? And it turns out, really not that good, it’s like flipping a coin. And I think we see deepfakes are getting better and better, so it’s getting harder and harder. But one thing and one sort of argument that we always try to make is that manipulation of audiovisual content is not new and this is something that we’ve been talking to judges at the International Criminal Court, that they mention that deepfakes are just the newest form of manipulating content. But there’s always been other ways you can edit content. What we see a lot is misattribution, so taking a video out of context and saying, “This just happened in Ukraine.” So I think what we’ve seen is that this awareness that a video can only ever show so much, you don’t know why someone started to film at one point and then actually stop, whats actually not been captured on camera. That’s really important, and I think that a lot of judges are quite aware of that, and that’s where there’s so many factors that we can consider when it comes to authenticating content and telling whether something is fake or not. And I think the key one is corroboration. We see that in most of the cases, you either have another video or a photograph to corroborate the finding, but a lot of times it’s eyewitnesses that are called in to confirm that something actually happened.

Stephanie: 08:34  You say judges are kind of used to doctored images or that they’ve had this before. But one of the things we see with judges is that sometimes they’re not necessarily au fait with kind of new technology and they need new skills to look at this. Do you think, you know, everybody talks about AI being this revolutionary change, especially if you work in journalism; it really gets touted as something that is going to fundamentally change our job. Do you think it will also fundamentally change judges’ jobs, in that sense? In the content that they get or having to weed out what is real and what is not real?

Janet: 09:14  Are you saying, Steph, that they’re old, like me? Is that what you’re saying? Because they’re old, they won’t be able to understand what’s going on?

Stephanie: 09:22  I’m saying they’re even older than you, and even I am, you know, very much Gen X. And it means that I kind of stopped at Instagram. I have an Instagram account, but I still don’t quite know how to do it and I see how quickly technology becomes obsolete and how, at a certain point, you just don’t deal with the new technology anymore. So, I think this goes for judges as well.

Anne Hausknecht: 09:43  Definitely, so I think there’s definitely a need for training, giving them at least a basic understanding of what is a deepfake, how we can spot it. Deepfake detection tools, they are not very accurate, so I think it’s important for judges to understand that relying on a tool isn’t this sort of panacea. We also have to have an expert coming in to authenticate content. But I don’t think that’s necessarily just the case for judges, I’m speaking to investigators and prosecutors, I think everyone still has this big need for more awareness about deepfakes and for training. So I think it’s sort of overarching, everyone should be aware that what they see online isn’t necessarily authentic.

Yvonne McDermott Rees: 10:24  And just to add to that, I’m not sure I agree with the premise that the role of the journalist is going to be changed completely in the age of AI, nor do I think the role of the judge is going to be completely transformed in the age of AI. I mean, as Anne said, this is maybe a new face on the old problem. But, you know, if we were ever, I don’t know, sent a document or shown an image, did we ever believe that that’s what it purports to be? Just as journalists have those skills in terms of looking, well, let’s look at the source. Where did this come from? Who sent it to me? Is this a reliable source? Is it, as Anne said, is it corroborated by other information? So, you know, I think sometimes there’s a perception that “oh back in the old days we used to take things at face value and we can’t do that anymore with AI,” but I’m not sure that was ever true you know? Photoshop has been around for quite a long time. And Anne, in her PhD, has this really good example of: in 1870, in the US, there was a guy called William Mumler who created this sort of spirit photography. So he would say to people, “Oh you’ve lost your loved one in the American Civil War. I can bring them back in my images.” And essentially what he was doing was sort of double-processing the negative. So you take a picture of someone and then have what looked like the spirit of their loved one in the background and he got prosecuted for fraud. But, you know, that was, what, 150 years ago? So, yeah, I think the idea that we were once in a golden age of, you know, we took things at face value and we can’t do that anymore, I’m not sure that really stands up to scrutiny.

Janet: 12:03  But are you finding in the work that you’re doing with judges, whether domestically or at the International Criminal Court, that they are, I don’t know, pushing back to some degree on visual evidence and saying, “OK, that is less reliable than we might have hoped it would be, and we’re really going to have to go back to basics”? Or are you finding them saying, “OK, understood. We now need to get our heads around things like authenticity, watermarks, you know, how to assess something”? Which way are judges going?

Anne Hausknecht: 12:41  I would say it’s more the latter. I think it’s more an interest in how we can actually authenticate content, geolocation, chronolocation. So determining where and when was something captured by looking at specific landmarks,looking at the weather patterns or the shadows in a video or photograph. So I think this sort of idea that we can’t rely on user-generated evidence anymore, that doesn’t really pan out. It’s more like “what role it can play?” Uh, I don’t think we will see a case at an international court, or even in the domestic cases that I looked at, where we just rely on one video or photograph because, even if it’s not a deepfake, it could have just been staged or taken out of context. So again, corroboration and just being a bit more thorough with regards to authentication and, I think also expert testimony is really important. Having an expert to come in and not just go through the steps they took in authenticating content, but also explaining how they go about that and how we can tell fake from true.

Stephanie: 13:42  One of the things I wanted to ask about is that in some kind of content generators you have this kind of digital watermark or added metadata that comes from what is known as C2PA, the Coalition for Content Provenance and Authenticity. Can you explain a bit what that does and how we are seeing that in user-generated content that is maybe used in legal cases?

Yvonne McDermott Rees: 14:05  Yeah sure. So C2PA is like an industry standard. So, it’s inherent in some hardware, I think, for example, Sony have signed up, other, you know, particular camera makers, Google, if you’ve got a Google Pixel phone that will be embedding the C2PA message data, but also in software. So if you amend an image using Photoshop, Google Gemini, ChatGPT, in theory at least that should have that stamp on it and then you can go to contentcredentials.org, upload the image and then you can see that C2PA stamp. So, for example, if I edit an image using Gemini and say, “Here’s a picture of me and Steph; can you add Janet in?” it will just say “this was amended using Gemini.” So we can see that there was some AI manipulation of it. So that’s something useful. I think, like Anne said earlier, you know, we can’t rely on it as a panacea. I think sometimes people say, “oh, well, this is the answer, we just need to get C2PA in everything and then we’ll be able to tell that it’s reliable or not.” Some of the dangers with that line of thinking is, first of all, not every device will have it. So, famously, Apple hasn’t signed up, because it kind of goes against their privacy, the importance they put on privacy. There are all sorts of other device manufacturers that won’t be part of C2PA. Secondly, I think, you know, we have to be concerned about potentially there, there are witness-protection kind-of issues here. So, C2PA doesn’t say “Yvonne McDermott took that image,” but it will say “it was taken by this kind of device.” You know, depending on the device or whatever, it might say in Belfast on this date. And you know, if you think about, if you’ve got a human rights defender in a region where there’s a crackdown on human rights defenders, there might be ways in which you could use that information and triangulate it with other information to identify the identity of someone. So I think we need to, kind of, push back on the idea that C2PA is the answer. Like everything, it’s just one of these factors that can be taken into account when holistically assessing a piece of evidence.

Janet: 16:20  OK, so my question was about judges. What about prosecution and defence? I mean, are there differences there in the way they are regarding the use of images? I’m trying to think particularly defence work is to try and poke holes in things and to say “this isn’t right” and so on. So, are we getting things like a, I don’t know, a battle of expert witnesses, where one says it’s this and the other says it’s that in terms of images and, you know, whether it’s been AI’d? How is that working in court?

Anne Hausknecht: 16:55  I think with regards to the defence, one issue is really equality of arms and whether they have access to as much information as the prosecution would have. A lot of the content that we see is guarded by civil society actors. They don’t necessarily collect the kind of information that the defence would find useful. So that is a big issue there. I don’t know, Yvonne, if you wanted to say something about the battle of experts.

Yvonne McDermott Rees: 17:19  Yeah. So in the cases that Anne has looked at, there have been usually expert witnesses from like national forensic institutes. Sometimes the defence will call their own expert witness, but it really hasn’t been that battle of experts thing, it hasn’t come out in very many cases.

Yvonne McDermott Rees: 17:36  There was one case I remember from Sweden and it was a case where, it was one of these cases of like returning ISIS fighters, and the defendants were shown in this image of a like a beheading video, but they had their faces covered, so they had like a mask and also something so all you could see is the eyes. But one of them had a tattoo on his arm or on his hand, and they also spoke a dialect of Swedish but like a dialect that’s quite unique to western Sweden. So, someone from the Swedish Forensic Institute was like, “I’ve done an analysis, and the likelihood of it being someone other than this defendant, whose got the same tattoo and is from western Sweden, is so small.” And I remember the defence called their expert, a professor from Lund University called Christian Dahlman, who said, “You’ve chosen the wrong reference class. You’ve compared it with everyone in Sweden, whereas actually you should look at what’s the likelihood of a returning ISIS fighter from western Sweden also having this tattoo and then actually it could be someone else.” But that didn’t fly in the case, and the defendant got convicted. But I think, yeah, we’ll see new – we saw it in our mock trial – the defence chose not to call their own expert witness, but rather to point to the limitations of the expert called by the prosecution and to try and, again, poke holes in “okay, well, what do you not know?” You know, “did you approach this with a sort of a biased mindset?” Yeah, different ways of sort of undermining the expert. And interestingly, the majority of our juries acquitted. So, nine out of our eleven juries went not guilty, and the other two were hung juries. So, they just couldn’t reach a decision, even when they were told “we’ll accept a majority verdict.” So thats what one of the ten out of the twelve of them can agree on.

Yvonne McDermott Rees: 19:31  So for me this was a real eye-opener, because I’ve always thought well it doesn’t really matter if you don’t know the source of a piece of user-generated evidence, or the person who captured it isn’t able to come and testify, because you can always get an expert to independently verify it. And we saw that our juries just didn’t think that was possible really. They said, you know, “we think this expert seems qualified, reliable and so forth” yet they had doubts about whether the piece of user-generated evidence could be independently verified.

Stephanie: 20:01  We kind of see that in journalism as well with images that now, because the world is kind of less reliable and those deepfakes are getting better and images get doctored and end up on Twitter, or X, or wherever. There is a lot of people doubting genuine narratives or genuine pictures. Or if you want to cast doubt on something, you can just say, you know, “this is just like all the other stuff,” and you can never now know that something’s true. Is that something you think, is that an effect that you see with juries having to decide? And is there a big difference between say jurors and professional judges in that sense, if you’ve looked at that?

Yvonne McDermott Rees: 20:41  Yeah, I, I, I’ll let Anne come in here first, because she has some really interesting findings on the extent to which that sort of deepfake defence has come up in court.

Anne Hausknecht: 20:51  Yeah, so we haven’t seen a sort of liar’s dividend or deepfake defence come up in any of the international crimes trials, I think mostly because the cases at the ICC and in the domestic courts, they all predate crimes that were committed before deepfakes. But we’ve seen this deepfake defence being raised in a couple of cases in the US. There are two cases that came out of the Capitol Hill storming, where the defendants themselves had been capturing a lot of photographs and videos, proudly sharing them with family and friends, and that content was then used as evidence against them and Counsel for them stood up and Court said, “Well, this could be fake. We can’t trust this kind of evidence.” And luckily, in those cases there was CCTV footage and police-camera footage to corroborate the findings, so there was enough corroboration there.

Anne Hausknecht: 21:40  But I do think this is definitely increasingly a problem where experts are really important. One case in which the deepfake defence was used successfully was a murder trial in the US, where prosecution wanted to show a video on an iPad and used the zoom-in function to sort of expand the video and the defence said that if you use the zoom function, it sort-of manipulates the video. So, the court said, “ok well, the prosecution has to bring in an expert to explain that using the zoom function doesn’t actually manipulate or use AI to change the video.” And they couldn’t find an expert, so they just could not use that as evidence in that case.

Yvonne McDermott Rees: 22:20  So, on our juries, we did see a little bit of this. Some of the, I remember one of the jury members in one of our groups said, “Oh, you know, I saw a photo competition recently, and the very famous photo competition and the guy who won it, it turns out that actually the whole photo was generated using AI.” So that sort of doubt, I guess, is creeping in. But I think, you know, people are logical and rational in terms of how they approach the evaluation of content. So we do a lot of like online studies and the TRUE Project, and people say, you know, “Well obviously I’d have to see is there’s anything else on this incident,” or, you know, “I’d look at the source,” or I thought, you know, “I might ask friends and family.” I might, you know, make this evaluation holistically. So I think, yeah, we’re not quite there yet ,where we hear it a lot, right? Like no one believes anything they see anymore, but I think it probably, this is such a lawyer’s answer, it depends.

Yvonne McDermott Rees: 23:18  We did see the idea of like, “oh well, anything might be a deepfake” come up a bit in the European Court of Human Rights. There was a case that Ukraine and the Netherlands took against Russia around the attacks that started in eastern Ukraine back in 2013 was it? or 2014. But In that case, there was one particular image that had been circulating online that was used to geolocate this Russian soldier, and the military unit to which he was a member to a particular site of strikes from eastern Ukraine. And there were actually two versions of the same photograph in circulation. So one had him with a tank in the background and the tank number is on the tank and there’s like a column of soldiers behind him. And the other was without the tank number. And Bellingcat had used one in one of their reports, and the Atlantic Council had used the other version. So before the European Court of Human RIghts, Russia really seized upon this discrepancy and said, and you know, even the column of people behind, that could be cloned from a single individual. So it’s that sort of deepfake defence coming in, in practice. And what the European Court did was they actually got a witness statement from Eliot Higgins from Bellingcat, who explained, uh, what happened was there actually, the soldier himself had used this image as his profile picture on VK.com, the social media site. And he had used the first version, and then later he took it down and replaced it with the second version. So a very simple explanation. And the court said, well, that sort-of explains that. And as for the other things that Russia has referred to, they are only what the state has called “indications of potential fakery.” And the European Court’s Grand Chamber said that falls quite significantly short of what we would need to see to be convinced that this is, in fact, a doctored image. 

Yvonne McDermott Rees: 25:13  So I think it goes to show that the sort-of blithe assertion that like anything might be AI these days is probably not, you know, the best legal strategy. And, there are probably other ways that one could challenge a piece of evidence: that’s not the defendant in the image, or there was someone out of shot holding a gun or, you know, we’ve seen those kinds of challenges a lot more in the case law database that Anne has put together. Sorry for the plug, but on our website, trueproject.co.uk, Anne has compiled this amazing database of cases, and it’s freely available. So that might be something people want to check out after the show.

Stephanie: 25:52  I’m going to slightly Stephopedia this as well. I think the European Court case you’re talking about is also the one involving the downing of flight MH17, which is why the Netherlands was involved. And we’ve talked about this on several podcasts, notably, I think, with Marieke de Hoon, the Dutch international law, and we’ll link to those in the show notes if you want to know more about that specific case.

Janet: 26:14  I want to introduce an element that’s also quite kind-of popular at the moment: sloppaganda. But just before we do that, I wanted to say that one of the reasons we were talking about doubting true narratives was because long-term supporter of the podcast Ronald Sly said that he’d been thinking about that as part of the AI series. So thanks, Ronald, for helping us with the question. You too can support the podcast and send us in questions if you ever want. Just a plug to counter the TRUE Project plug as well, or to add to it, let’s say.

Janet: 26:48  So, sloppaganda, and, you know, the sense that we have that there’s lots and lots of slop around – we’ve seen it particularly in the conflicts Iran, US, Israel – it has a purpose. Maybe we could describe that purpose and then maybe also describe what we think might happen with that kind of stuff in court. Is this going to play any part in, I mean, I’m not saying they’re going to be any trials out of this, but is that the kind of thing that might end up being discussed in court sometime? Start with what, what’s it for this sloppaganda?

Yvonne McDermott Rees: 27:30  Well, you know, a lot of research has been done on this, like Neve McIntyre, who is a journalist based here in the UK, did a big deep dive into who who’s behind all this AI slop, particularly in the context of the war in Iran, and it was found that actually a big incentive is monetary. You know, theres a lot of these accounts that are based in Thailand, Cambodia, the Philippines – in other words, in countries that aren’t actively involved in the conflict – generating this kind of AI slop. And that content drives engagement, and engagement drives clicks, and clicks bring money. So that does seem to be a big incentive. Then the other side, of course, is the sort of propaganda side. So we’ve seen, for example, Donald Trump reposting AI-generated stuff. We’ve seen the Iranian government repost AI-generated content. So that aspect of it can’t be set to one side, but neither can we ignore the very real monetary reasons behind some of this content.

Yvonne McDermott Rees: 28:25  In terms of what can be done about it, the Meta Oversight Board had a really interesting decision recently on this exact decision. Like what should Meta be doing on AI generated content from the Israeli-Iran war? And they found that, you know, first of all, meta wasn’t really actively implementing its own C2PA standards. So it wasn’t actively doing that and flagging content as AI generated. I think that would go a long way in just putting that sort of watermark or warning to viewers that this is AI generated. And also, you know, there’s also a role in terms of demonetising that kind of content and downgrading it. So there’s a lot that could be done that falls short maybe of taking it down altogether, that platforms could be doing. 

Stephanie: 29:14  We’ve talked a lot about the problems with possibly AI in this evidence. But is there also a role for AI-assisted evidence reconstruction of what houses looked like before and after, you know, plans of a prison, those kinds of things?

Yvonne McDermott Rees: 29:30  I think there’s a role for this kind of thing in advocacy, in human rights reporting. I would be very wary about using it in criminal trials because, as I think Sarah Zarmsky, my colleague from here in Queen’s, I think she’s been on the show talking about this before and she’s also written on this. You know, there is that risk of the sort of cognitive bias that comes in seeing something with your own eyes and the way that we process that visual content. It can be highly, highly prejudicial, I would say. And probably that prejudice outweighs its probative value. Sorry to use some legal terms, but in terms of the actual weight that it should be given, its strength as evidence is outweighed by the prejudice that would be caused by it. So I’d be quite wary about using it in the courtroom. I don’t know if Anne would agree or disagree.

Anne Hausknecht: 30:23  No, I agree with that. Yeah, I find it very tricky. 

Yvonne McDermott Rees: 30:24  I’d say on this, someone asked me recently, “Well, what about like witness protection? Couldn’t you have like deepfakes, you create an avatar of someone and then you feed that AI model their words and then it looks like someone is presented, but its their identity is completely protected?” I mean, again, I think that could something that would be useful. You know, maybe, we’ve seen some documentaries, for example, using this, or, that kind of thing. But in court, you know, really, evidence needs to be tested rigorously, and there are well-established ways of protecting witnesses that don’t involve turning to AI and creating AI avatars and things like that.

Stephanie: 31:08  I was going to say, I saw a documentary on, I think, a Mormon polygamist cult, where they had the underage wives of this particular kind of prophet. And they said, “well we doctored them because, you know, it looks better if, so, they’re not recognisable, but they do look like real people.” And they’d say that upfront in the documentary but when you look at it, its really, your brain doesn’t really compute that, they don’t look AI, they don’t. It wasn’t uncanny valley kind of where you’re like, something is off. So I was really looking at that thinking: I can see why you would do this in a documentary, and I can see that it looks more engaging right? Especially if you have evidence in courtrooms here where, you know, when we have a protected witness, it’s like blocks and pixelated. And of course that creates a layer of distance. But then I was also thinking, well, the judges don’t see them pixelated. They see the real person and they have to judge most of the time, unless it’s video evidence. So, you know, I, when you talked about documentaries, that sprang to mind because I was looking at that thinking, this is so odd, but also very effective visually and my brain is being tricked into thinking that this is completely normal.

Janet: 32:18  What was in my mind when I wrote the question was work on North Korean prisons that I’ve done, where we used, quote-unquote, blue bodies to represent people because, you know, and didn’t give them any characteristics at all. They were just vaguely robotic, but to, you know, one had a whip and one was cowering and one was, you know, ten were packed into a cell and so on. But, I would have felt distinctly uncomfortable with showing that kind of detail with people that look like people. AI’s, i mean, it felt right to make it as, I don’t know, as unreal, but effective, as possible.

Janet: 32:59  Yes, we’ve gone off on our own tangent, of course, as, as

Stephanie: 33:01  As we are one to do.

Janet: 33:04  Yeah, we had another 15 questions, but we’re not going to get around to doing them. It’s going to wrap up here and just ask you to start with, is there something that you wanted us to ask you that you haven’t had a chance to say? Anne, to start with?

Anne Hausknecht: 33:19  I think the main point that we’re trying to make is that it’s not a new phenomenon. Manipulation has always been around. So, I think the big fear with our work as well, when we’re constantly talking about deepfakes, is that we are doing more fearmongering than good. So, I think it’s important to be aware of deepfakes and to know that they are quite tricky and risky and difficult to detect. But yeah, we should always approach anything we see online with a healthy dose of scepticism, but not cynicism that just dismiss everything outright.

Yvonne McDermott Rees:  33:50  Yeah, I couldn’t have said it any better. 

Stephanie: 33:52  And then we’re going to move on to our kind of traditional Asymmetrical Haircuts question. If you follow the podcast, you’ve probably heard it before, but we like to ask our guests if there is anything that they’ve been watching, reading, listening to, or even that they’re doing, that they want to share with us. It can either be related to the field you’re working on, or it can be whatever you do to get away and keep your mental health intact, which is also important in this kind of work. So let’s start with Anne. Do you want to share something that you’re enjoying listening, reading or watching?

Anne Hausknecht: 34:28  Well, I’m not reading anything AI-related because I feel like I’m doing enough work on that. But I’m currently rereading Nadia Murad’s book “The Last Girl,” about being held captive by ISIS.

Stephanie: 34:39  That doesn’t seem very calming, honestly.

Anne Hausknecht: 34:43  It’s not. It’s not fun. It’s an extremely honest and inspiring account of what she went through.

Stephanie: 34:49  Yvonne?

Yvonne McDermott Rees: 34:49  Usually I like to read fiction online, but there’s two, there are two really good nonfiction books that I’ve really enjoyed recently, and one that is kind of tangentially related to this. “Careless People” by Sarah Wynn-Williams. I don’t know if you know about it. It’s the inside story. She was the head of global policy at Meta and it’s sort of the inside story. It’s been quite controversial. Meta have sued her, and the result of that is that she’s been prevented from publicising this book in any way. So she recently was at the Hay Festival in Wales, where she had to sit in silence on the stage and not nod or, you know, blink, so, while two other people discussed her book. But it’s, you know, it’s really compellingly written and she, you know, tells a good story. It’s bizarrely, it’s quite funny in parts, so I’d definitely recommend that. I also really enjoyed Patrick Radden Keefe’s latest one, “London Falling,” which is really good. And then, podcast I’ve really been enjoying recently is called “The Expert Witness.” It’s by CBC. Have you seen that? Yeah

Stephanie: 35:58  I am fascinated by that.

Yvonne McDermott Rees: 36:01  Yeah, yeah, yeah, and related to what we’ve just been talking about, because this new thing called CyberCheck was suddenly being introduced in domestic criminal trials. And, you know, it sounded all very high-tech, it was like we use machine learning, we use, you know, hashing and cryptographic this and that. And it turned out basically, it was nonsense, and it was being used against people and very difficult to sort of argue against because it was purportedly an AI system. So yeah, I think thats a cautionary tale, that podcast.

Stephanie: 36:30  It really is. It’s especially indeed about they would just say it’s AI, and the AI would essentially look at all your social media and whatever you did online and then kind of make this footprint and place people digitally at the scene of crimes, and it was used to convict. And when it got to the courtroom, none of the judges or prosecutors could really parse it or challenge it, and there were no expert witnesses to explain what it did because whenever they got to kind of the secret sauce of the AI, the person would be, who was doing it would be like, “well this is protected, this is mine and we’re not sharing this,” which is, you know, one of the big problems with AI: there’s no visibility about the algorithm, which we’ve talked about in other podcasts. But it’s really fascinating how it then gets kind of used and people accept it for what it is and you really have to have these Don Quixote-type prosecutors who go against it. But it’s a fascinating podcast, it’s “The Expert Witness” by CBC podcast. We’ll link to it.

Janet: 37:39  Great. Now, Anne and Yvonne, thank you so much for making time to have a chat to us and help us along with our AI series and i’m not sure where we’re going to go next, but I’m sure we’ll come up with different ones. Don’t hesitate to let us know if you think we should be doing some specific topic, not necessarily related to your one. Thank you so much. Thanks for making the time, Yvonne, really appreciate, and stay safe.

Yvonne McDermott Rees: 38:03  Thank you.

Janet: 38:05  Thanks, Anne. Bye-Bye.

Anne Hausknecht: 38:06  Thank you. 

Stephanie: 38:08  Thank you.

[OUTRO MUSIC]

This was asymmetrical haircuts, your internationa​​l justice podcast, created and presented by Janet Anderson and Stephanie van den Berg. This episode was created in partnership with Justiceinfo.net, an independent site covering justice efforts for mass violence, and with the Hague Humanity Hub. You can find show notes and everything about the podcast on asymmetricalhaircuts.com. This show is available on every major podcast service, so please subscribe, give us a rating and spread the word.

Stay safe and enjoy your day.