* Musik * Herzlich willkommen zum Podcast "Fokus strafrecht". Wie immer mit der Charmanten Nadine zur Kinden. Assistenzprofessorin für Straf- und Strafprozessrecht an der Universität Zürich Hoinerding. Hallo, Glow. Und wie immer mit dem ebenso Charmanten, Glowed Eric Bertschinger, rechtsanwalt unter Zend an der ZHRW. Nadine, wen hast du uns heute mitgebracht? Heute habe ich Sarah Somers eingeladen, die uns etwas erzählen wird zum Thema "Predictive policing". Also, wenn ich das richtig verstehe, vorher sagen, der Polizeiarbeit könnte man das übersetzen, was muss man sich darunter vorstellen? Ja, ganz genau. Da geht es um vorher sagen, der Polizeiarbeit kann man durchaus so übersetzen und die Polizei braucht "Predictive policing" jetzt auch nicht erst seit aller Neustum, sondern schon seit längerer Zeit. Früher ging es vor allem um Tools, die auf Statistiken basierten. Und man kann da unterscheiden zwischen Tools, die auf Orte abziele. Geografisch vorher sagen, aufgrund von Statistischen Daten, wo Straftaten mit hoher Wahrscheinlichkeit begangen werden. Zum Beispiel Einbruchdiebstelle. Dann gibt es Tools, von denen ich nicht weiß, ob die in der Schweiz zum Einsatz kommen, die voraussagen machen, in Bezug auf Individuen, die möglicherweise straffällig werden. Jetzt hat sich die Technologie natürlich weiterentwickelt in den letzten Jahren. Wir sprechen immer mehr über künstliche Intelligenz und so hat sich auch das "Predictive policing" weiterentwickelt. Und ist heutzutage weniger Statistik passiert. Sondern es gibt mittlerweile auch Tools, die mittels Maschinenlearning-Muster erkennen. Und dann aufgrund dieser Mustererkennung, Patron Recognition, vorher sagen machen, wo Straftaten passieren könnten und wer da allenfalls involviert sein könnte. Und darüber wird Sarah uns heute berichten. Ich bin schon sehr gespannt. Ja, das ist ein spannendes Thema von dem, wo es sich auch noch nicht sehr viel bescheid. In der klassischen juristischen Ausbildung, wenigstens an der Uni, hab ich eigentlich nie von diesen Tools gehört. Und dann auch in der Praxis, auch in meiner Erfahrung als Straferfolger und Weiterbildung. Ich weiß auch nicht, ob viele Straferteidiger*innen, also die Fachexpert wahrscheinlich schon, aber ob alle Anwältinnen und Anwältinnen immer wissen, dass ihre Klientschaft vielleicht gefasst wurde auf Grundlage von solchen Tools oder Ähnlichem. Das weiß ich nicht. Also da muss ich spekulieren, aber ich fließe von mir auf andere. Von daher ein sehr spannendes Thema, dass uns Sarah's Wammers hier erörtert. Ja, und du sprichst hier, welchen noch kurz darf, ja bitte. Du sprichst hier indirekt einen anderen interessanten oder wichtigen Punkt an, im Zusammenhang mit Protective-Polising und zwar, dass es ja eigentlich primär mal um die preventive Polizeiarbeit geht und nicht um die Straferfolgung, um die repressive Polizeiarbeit, wobei die Grenzen hier eben zunehmend verwischen, was dann ein weiteres Problem natürlich gibt, so aus Ruhl-of-Law-Sicht, wie uns Sarah sicherlich erklären wird. Fokus strafrecht. Der strafrechtspot-Kaas der Universität Syrich. Für Profis und solche, die es werden wollen. Mit ihren Hosts Prof. Dr. Nadine zu erkennten und Dr. Klod Ehrig-Bärtschinger. Today we have a very special guest, Professor Sarah Summers. Sarah Summers is Professor of Criminal Law and Criminal Procedure Law and Criminalogy at the Faculty of Law of the University of Syrich. She studied Law at the University of Glasgow, in the UK, and received both her Doctorate and Habilitation from the University of Syrich. Her research interests lie in the field of Criminal Law and Human Rights. Her research concerns considerations of the empirical realities of criminal justice and of the importance of normative principles underpinning systems of criminal justice in the rule of law. She is co-director of the Digital Society Initiative, the S.I., Commission Member of the Criminal Institute of the Canton of Zurich, a Member of the Prosecutorial Regulatory Oversight Commission of the Canton of Basel City, and a Member of the Board of Trustees of the K.I.H.Z.Kiz Daycare Center of the University of Zurich and ETHZ. She is also on the Editorial Committee of Questiofakti International Journal on Evidencial Reasoning, and is on the Editorial Board of the International Journal of Evidence and Proof. This much I have learned from your biography page of the University of Zurich, are you happy with that introduction? Hi, I'm indeed. Thank you very much. It's a pleasure to be here. It's a pleasure to have you. Thank you. Welcome, Sarah. I'm very happy you are here today. A predictive policing. Maybe we start right with what is predictive policing. So predictive policing is the idea in some sense that you can stop a crime or prevent crimes from being committed before they happen. So the idea is that what if you could stop a crime before it happens? And it's at once the promise, but also the threat of predictive policing techniques. So these techniques, they promise efficient, proactive, not just reactive policing, low crime rates, perhaps safer communities. But they are also accompanied, as we'll see, I think, in the course of this podcast, by a number of quite serious concerns, concerns about bias in already highly or overly police communities, concerns about their potentials to generate increased citizen and police tension. So these are some of the issues that can arise. And I think I could maybe begin by illustrating the idea of predictive policing with an example. And it's a very old example, but nevertheless one which I think might be quite useful for sort of thinking about these ideas. And it's the case of Robert McDaniel, who was a citizen in Chicago. And then in 2013, he was visited by some police officers from the Chicago Police Department. And they told him that they were keeping an eye on him and he should just be careful about what he was doing when he was going about his day-to-day business. And he was surprised because he hadn't committed a crime and the police officers were not accusing him of having committed a crime. And what happened is he'd been put on the Chicago Police Department's strategic subject list, SSL, which was designed to determine whether people might be involved in a shooting. And this was based on their proximity to gun violence. And the data couldn't predict whether the person was going to be shot or whether the person was going to be the shooter. But the idea was that there was this proximity and he was on this list and there was someone that was a potential target or a potential criminal. And the police hoped in this case that by warning McDaniel that they would be able to avoid obviously him being the target or the perpetrator of gun violence. But in fact it didn't work out quite like that. So this is quite interesting thing about this case. What happened was people in his neighbourhood noticed him speaking to the police. They were worried that he was then an informer and he ended up getting shot because people had seen the police visiting him. So we see here very much the worries, the dangers. But this is a sort of idea. This is the idea that by kind of working out who is likely to be a target or a perpetrator of criminality or violence, then the police might be able to get involved and try and prevent it. And how did they think they could prevent him from being a target by telling him very unspecifically that he should be careful? I think that's part of the problem is that the police get this data and it's still at the end of the day, the police are very, there's a lot of expectations about police experience and what they're going to do with this data and how they can use it to effectively target or prevent violence or criminality. So there has been sort of a prophecy effect, basically, the fact that they have alerted him being cautious in his day to day has actually made him a victim of a gun violence in the end, right? Yes, and I think this is something that we often, I think if we read the literature on the subject of predictive policing or think about the subject more broadly, we see these are the concerns that we see also in the context of location-based predictive policing. So the worries that by having more police involvement in certain areas, that this is likely to be a self-fulfilling prophecy if the police are there more often, they're likely to uncover more crime. The fact of police intervention might result in not just a more criminality although that might be part of the reason like in this case, but also just that more crime is uncovered and it turns into self-fulfilling prophecy where overly-policed areas, there seem to be more crime committed there than in other areas. But I think it's quite important, I think, at the outset, as well to think about the fact that predictive policing is very much a normal everyday fact of policing. So the idea that one tries to prevent crime is a very important part of the police mandate, as important, at least as the repressive part of policing. So we have repressive and repressive.
preventative policing and in that context, predictive policing, even AI, predictive policing might be seen as something which is just a continuation of a broader phenomenon, of broader attempts to try and prevent crimes taking place. And I think we can see these attempts at prediction that trying to work out who might be dangerous for instance in society. We see these are issues which have, there's nothing new, this is like something that we've seen in the criminal justice system throughout, certainly the 20th century, in various points in the criminal justice system. So we see attempts at prediction, attempts at assessing or predicting dangerousness in the context of parole decisions. We see it in the context of sentencing decisions. So we might see that predictive policing is based on quite a distinct understanding of crime and delinquency, which is connected not so much to the structural society, but more to the individual or social pathologies. And I mean that's quite a dominant strand of positivism in the history of criminal justice system in more recent times. So I think it's important to see that. At the same time that I think, and that's perhaps what we want to discuss a bit more today, that this specific problems associated with artificial intelligence and the increasing use of algorithmic models for predicting violence or crime. May I go back just once then? Yeah. Because the case you mentioned was in 2013. Yeah. So did the police learn anything of that? So that self fulfilling prophecies do not happen that often anymore? Do they have strategies? Do you know anything about that? What do they continue to use? Well, I think these types of systems are still quite prevalent in different places. The metropolitan police in London used like it's called a gang violence matrix, which is very similar in some sense that you attribute a score to individuals who have been identified through their behavior as being potentially likely to be involved in criminality. And by involved, you mean as a victim or as a perpetrator? Exactly. So maybe for our listeners who are maybe not aware of all these tools that are already in use and how they work, if we had to simplify it a little bit. I mean, I recall having watched a movie back in the days. It was called Minority Report and it was a movie from the year 2002. And they had these ladies in the pool floating in a pool and they were like alien like the head visions and they could kind of extract those visions from where what would happen in the future and they would use that to prevent crime or try and prevent crime is predictive policing something like that. I don't think we're quite at the level of minority report just yet. You have to say sorry to disappoint you poor Hannah. No, not sure. But it's important, I think, to sort of like say, what is it we're talking about when we're talking about predictive policing? So on the one hand, we have a predictive policing that's been defined as something like the use of analytical techniques, particularly quantitative techniques to identify likely targets for police intervention and then to make like kind of crime prevention by making statistical predictions. And I think it's quite important to distinguish this kind of general idea of predictive policing from AI or algorithmic predictive policing. It's not the minority report level, but it goes rather in this direction that we it's the use of algorithmically mediated data analysis for the purposes of finding a patterns in data sets on which risk estimates are produced either for individuals or for locations and are operationalized if you like in the form of preventive measures. So targeted preventive measures. So we see here a distinction between statistical mapping, if you like, and the use of algorithmic tools to enable pattern recognition. This is a distinction that we have to keep in mind. And we have to also keep in mind that these two distinctions apply across these two major categories of predictive policing possibilities. So on one hand, the person-based, which represents an attempt to identify as we've discussed already individuals who are likely to commit or be the victim of crime by analyzing factors like past arrests, victimization patterns, social networks. And examples include the Chicago Strategic Subject List, and we've discussed this gang matrix used by the Met Police. But there's other forms of person-based algorithmic policing measures, which include also automatic facial recognition systems, which have been used in London as well. For instance, a few years ago, the London Police had admitted that it had been supplying information. So it entered into something like an information sharing agreement with the private operators of Kings Cross estate, which was a huge estate in London, close to Kings Cross railway station. And the police supplied then ostensibly for crime prevention reasons with the images of people who'd been involved in crime or who'd recently been arrested. So this was a very, with the view to sort of like enabling the private operators of this estate essentially to probably remove people from their ground, with the view to try and prevent them committing crime. So very, very problematic, obviously. So example is used in Zurich, this derias in team partner system, which is used in the context of domestic violence. But again, the idea is that people have drawn awareness to themselves in some sense. They become like persons of interest. Persons of, of instances without actually having committed a crime. So they've just drawn some sort of awareness from the police to themselves and then their subject to kind of like a risk assessment tool whereby there's like a scoring system to try and work out the extent of the risk that they propose. Dr. B explains. Derias stands for dynamic risk assessment systems. Derias intimate partners is a tool that assesses the current risk of a male individual committing a serious act of domestic violence. This tool has been developed at the University of Durham Stutt Forensic Psychology section. So this is like examples of person-based, predictive policing. More widespread of course is the use of place-based predictive policing and that's what I think we are also more familiar with here in Switzerland. And whereby data is essentially analysed with a view to predicting when and where crime is likely to take place. Where the view perhaps to enabling the police to send out targeting patrols, but not necessarily that in the early days there might also be a different type of response. So there's a lot of crime in a particular area you might want to increase the lighting in the area. I'll put up more street lighting to make it brighter and it doesn't necessarily have to be like a police operational police patrol response. But I mean that's typically what you think about that they can send out police patrols. So you might just distinction between person-based and place-based predictive policing and you said in Switzerland we use the larger more, the place-based predictive policing. Does that have to do with the legal framework? I don't think so necessarily to be honest. Actually so we go back if we think about this. I mean we have this diaries in team partner that we use for domestic violence also in Zurich. It's a very much a person-based tool. It might be to do with legal basis in some sense because of this other distinction between a statistical modeling and algorithmic tools in the context of predictive policing. It might be useful then just to compare the different types of tools. Different models. So Sarah could you maybe explain a little further which are the different tools or how do they function? There seems to be certain differences between the tools you've been describing. Yeah so in the early tools, early methods of analytic methods of statistical modeling such as pre-cops which we also used here in the city of Zurich, they're very much theory-driven. So maybe it's useful to just discuss a little bit of chronological theory to explain what that might look like. They begin with a hypothesis which is informed by the subject matter. So this idea of crime and criminological causes of crime and so on. So according to rational action theory, the decision making of the potential offender depends on certain promising or unpromising offending opportunities of a given situation. Can you make an example of what is? Yeah so it seeks, for instance, in the context of burglaries. It seeks to predict burglaries by using what's known as a near-repeat statistical observation that burglaries often occur at places close to places which have already been burglarized. So this is just the statistical observation. It's quite common that when a house has been targeted, soon after, other houses in this location will also and this area will also be targeted. This is a theory, this is like some of the criminological hypothesis which is born out by empirical data and is well known in the chronological literature. And then there's another hypothesis that's called in this context, the boost hypothesis which tries to explain why that is the case. So another criminological hypothesis which tries to explain why houses close are more likely to be burglarized. [BLANK_AUDIO]
a temporal and near situation. This is the type of statistical modeling which these tools like pre-Cob use. So they have a chronological theory, there's a hypothesis behind it and on the basis of this they map the use of the data to try and work out where there's like hotspots and then they try to use that for operational police reasons that resources are sort of put in this area with a view to crime prevention. So I think this is very important for our purposes here is that these tools, the deliver results which are explicable from a chronological perspective. So you can look at the results and you can explain to people, well this makes sense to divert resources to these regions which are being and we can explain why this is happening and it feels transparent, there's certain levels of accountability, it seems sensible from a efficiency perspective and so on. And what about algorithmic tools, do they work completely differently? I would say the complex algorithmic models, they're different. So pred poll is similar to pre-Cob's to the extent that it aims to do the same thing. So it aims to tell the police where there are like hotspots for particular crimes, burglaries, car thefts, robberies to play this type of crime in particular locations. But the important thing is not theory based, at least not chronological theory based. These models are based on an algorithm which was inspired by seismology and the fact that certain crime patterns follow similar predictable aftershocks like earthquakes. So there's no, there's three vertical models, but the theory is not at all based on chronological theory. Nothing that we could explain in the sort of sense I mentioned before with these theories about near repeat statistical observations and boost hypotheses and so on. Nothing like that. I'm not sure I can follow. Maybe you can. So you would need a geologist to interpret today. Exactly. So you have like theories like to do with like how likely is that aftershocks will follow an earthquake? There's a big difference because the empirical basis of the statistical modeling tool that I mentioned earlier is experience of policing and crime and crime prevention. But in the context of, of let's say, pred poll or geological or whatever it's called. Now I'm thinking there's no empirical basis in criminology. The empirical basis is in earthquake theory. So it's basically math on the lying math or similar mathematical reasoning like after an earthquake there's this probability of this and this happening in the aftermath of an earthquake. Exactly. And then they have just realized that the math on the lying principle may also apply or simply to apply also to human behavior with regards to criminal activity. Exactly. That's exactly the situation. So they have this it's called an epidemic type aftershock effect. And they say this sequence model is based on population genetics and epidemiology and notions of contagion. The idea that victimization can maybe be passed from victim to victim. And yeah, but the option of this is that any sort of understanding about how this system works, this type of algorithmic model, how this works is limited. So we can't really explain the connection to crime or criminality or victimization patterns or we can't explain anything about it. So we have some opacity problem there. Yeah. So it's theory driven but it's not driven by some chronological theory. And then we can take this one step further with newer models which are less well developed in Europe but certainly it being utilized actively in the United States. So models like machine learning tools, I mean which essentially allow the algorithm to learn from data and continually improve its performance by drawing probabilistic inferences from the training data. And we have things like Hunschlab which is describe itself as a patrol management system. It's used by the Philadelphia police. I think the company is based in Arizona but the Philadelphia police are the ones that kind of pioneer the use. And it estimates the likelihood of crime being committed in a particular location at a certain period, so at certain time or something like that. So it uses crime data and algorithms to sort the data and combines that with a whole host of other information. So it might be the weather. Do we actually know what kind of data goes into the system? No, I think the point is that it's not just that we don't know which data goes into the system. These systems actually are not that bothered by the type of data that goes into the system because it's just essentially if it helps train the algorithm then that's a good thing. And I think this is quite interesting to say that the complex machine learning algorithms are said to generate accurate predictions but there's no attempt to explain why that might be the case. And if we look at the people that are working on this, for instance, one criminalologist who's very like a very big proponent, if you like, of these algorithmic models and machine learning models, Richard Beck, he said famously that if all other things are equal, shoe size is a useful predictor of riskedivism, then we should use it as a predictor. So he said, if there seems to be a link on the basis of the data between your shoe size and the likelihood that you're going to be a criminal, we should use that. So this is kind of the problem with these types of tools. We have generalizations, stereotypes which are devoid of any sort of empirical foundation. So I think we call them like spew in the law of evidence. We call them spurious generalizations. Did I understand correctly that one just puts all kind of data into the system and then let's anything which might be useful? How do we know which might be useful? The machine with time, that's to do with the way that it learns. But if you do not give the system weather data or shoe size data, you will never know whether it will be useful. Yes, but remember, with these types of systems, particularly with machine learning ones, they're teaching themselves, it's like we're not even giving them the data anymore. So they are like collecting the data by themselves. So it's not that. But I think the problem is, and I think we can see here that these types of spurious ideas, like some sort of connection between shoe size and recidivism, they are not absent from the history of chronological thought, just to come back to patterns of thinking about crime and criminology. So if we think about chronology, we had this idea that there might be some connection between head size or shape and likelihood to commit crime. So the history of chronological thought is not devoid of this type of spurious claims or thoughts and claims implicit, for instance in Bex's idea of this connection between shoe size and recidivism or propensity to commit crime, they're almost certainly untrue. So there's real issues about the epistemic value, I think, of this type of pattern recognition in this specific context. But obviously, of course, this is bad enough if we're discussing, well, obviously, there's no connection between shoe size and recidivism, but then there's, of course, deeper political and moral claims that we're all familiar with. These generalizations don't just relate to shoe size, they relate to sex or gender or socioeconomic background race. And so race, of course, important questions about commitment to equality and non-discrimination, the rule of law. So maybe we have to summarise this up in a bit of a dumb-to-down version for people like my humble self. So if I on the studio, correctly. So they were the old systems, which were called statistical based statistical modeling, other statistical models systems, which basically looked at all the statistical data collected, and they said, "Okay, there's crime all the time at Friday night in the region of the clubs here in Zurich." Right. Right. And then it's basically like a weather forecast, it tells the police, "Maybe next Friday, it would make sense to allocate your main sources somewhere in that region." To try and prevent crime by being present, show a force, and that sort of thing. And maybe also catch criminals because they are most likely to be found there. And then the new statistical, not statistical, but the algorithmical systems, they are not based on these criminological theories, but they are based on the geological theories or whatever we want to call them. Certain mathematical, seismic, on the lying principles. Contagion theory from epidemiology or in all. And they can also be fed with several data that may or may not be relevant. And these are the more modern tools if I understood you correctly. So these are the two major distinctions in technology, I think. And then in the process, there's also these tools. You could also divide them in two groups, the ones that are more territory based, right, where they say, "Okay, there's more likely to be crime happening somewhere in the region of Alchdetten." And then there's the ones that are more individually based, where they say Claude Berger is more likely to commit XYZ. Right. Okay. Three main issues. So from a rule of law perspective, which one is the most problematic and why? Or are they all problematic? So I think we can see a distinction between the statistical modeling in one hand, an algorithmic tool.
the other and the main differences are of course that with the statistical modelling we can understand why the things are happening. It's explicable. So we have some sort of epistemic certainty and of course at least have a statistical modelling is subject also to the dangers of profiling discriminatory techniques perhaps but these dangers they're transparent. So we can say well if the police are sending I don't know 80% of their resources to the Langegrasse we might want to question that. We might want to say and the police might have a response to that so they might be saying well we've had a lot of crime there recently and that that kind of justifies the decision to do that. This is not like a racially motivated, this is not like socio-economically motivated and one can discuss it. In the context of the pattern recognition system so most of the algorithmic tools we have quite important challenges. I would say I would group them into three challenges. I don't know if that's useful. So the first time I say they pose like a chronological challenge. A challenge to our thinking about criminological theory about crime and the causes of crime and about how we're going to prevent crime. They even seem to these pattern recognition seems to call into question the importance of criminological theory in the first place. It also puts pressure I think and that's quite important that we think about that on notions of epistemic equality or understandings of knowledge. What it is we're doing when we're trying to work out where crimes are to be prevented. The second issue which I'd like to stress is the worry and this is not just these pattern recognition tools but connected to the potential for widespread surveillance that we have the worry that this central distinction between investigative policing and repressive policing has been called into question. To put this in a placket of sense you can't just like in the rule you can't just treat everyone like their potential criminal and that's a bit the direction we're going towards where we have widespread surveillance not necessarily through facial recognition technology which has been curbed a bit but through other forms of surveillance and number plate surveillance for instance reliance on widely available geolocation data and when these types of surveillance tools are combined with other algorithmic goals in the context for instance of predictive policing then there's really worries about these traditional boundaries between repressive and investigative policing. So if we think about that boundaries being essentially reasonable suspicion the availability of these tools are really putting pressure on the idea that police intervention can only happen when we have reasonable suspicion that someone has actually committed a crime and I think that's one of the problems we see also in the case that I mentioned at the beginning in Chicago of the SSL case is that we say oh it doesn't really matter if he's a potential victim or likely to commit a crime and in fact most of us it does matter it's the very essence of this distinction between investigative and repressive policing in the rule of law and then the third factor very important of course is the issue of accountability transparency and the potential for introducing but also reinforcing bias in the context of policing decisions and when you talk about accountability and transparency do you mean the police or do you mean the companies that invent those tools and bring those tools on the market? I suppose it's all connected in some sense so if the tools themselves are not transparent as to the data that's being used or to the the reasons for the decisions then the decisions are subsequently made by the police whether that be like just basic decisions about resource allocation on the one hand but also then more operational decisions they're also then impacted by this lack of transparency and accountability in some sense to the extent that if like the data is racist then the operational decisions are likely also be to be tainted by these initial problems with the tools so I'm not sure that's something that one can really divorce because the police as we've just said there's no way of them knowing why these decisions have been made they can't explain it we can't expect them also to explain because there's no explanations which are available in the language of criminological claims or the usual types of things that would the police experience at the end of the day. And when we come back to the second point the widespread surveillance and the problem of the blurring between the preventive policing and the repressive policing what could be possible solutions there do you have any idea or one of the several solutions so one solution is just to try and prohibit these or reliance or restrict the reliance on surveillance tools by the police that one would strictly regulate in an effort to turn protect this distinction between police intervention repressive police intervention is dependent on the moment in term where someone is suspected of having committed a crime. But some people say that this is a lost cause so we'll see it's interesting to see in the future whether it's actually going to be possible to sort of stem the tide of the technological advances particularly in the context of surveillance because the surveillance opportunities are so prevalent that it's perhaps unrealistic to assume that they won't be utilized by the police but I think it's very important to be very aware of these dangers and to think very carefully about the potential for legal regulation to protect fundamental concepts in the relationship between the individual and the police and the rule of law. And I guess surveillance is one issue but the other issue is the data then used by AI tools to come up with certain patterns as you already explained which is something that I think you mentioned before we started recording the podcast that is prohibited in the European Union by the AI Act. Exactly so this type of internet scraping for photos and data to build a database of like to enable visual recognition that's been prohibited now and the UN that I think that seems an appropriate type of legal regulation the type of thing that would be very important at the moment. And it remains to be seen we'll see what the federal council decides to do in Switzerland we're expecting a report in 2025 so eminently on the broader Swiss approach to regulation of AI. So it'll be interesting to see whether they follow sort of the skeptical or cautious, more cautious lead of the of the European Union or whether they're attempted to look to America where it's a bit more wildlife it seems. So one could say that the data protection laws are a bit of the fig leaf that protects us from being constantly surveyed by this police or so if we have stricter data protection laws they can prevent or prohibit collection analysis of all this. Well I suppose there's always been an exception to the typical principles of data protection in the context of the criminal justice system so I mean we wouldn't expect that to be to change in any sense and what I think what has changed is the willingness of people to surrender their personal data not just to the police but to private companies right so we're all I mean I'm guilty of that too I have my location data turned on I like to use it I like to buy my ticket on my phone to to take the tram it's the willingness to surrender personal data to private companies and that it's perhaps at least as much of a subject as then deciding how we're going to regulate the use by the police of this data and also very important I think which firm and from legal perspective or regulation perspective is then the relationship of course between the firms responsible for developing surveillance and algorithmic tools and the police so between private firms and the police that's something that I know that is of a real concern to also the police authorities and how they should best regulate that in their to those questions about independence about increasing privatization of what we would see as a public service and so on so worldwide if you look at the situation there seems to be a bit of a spectrum if I understand correctly at least I hope so and so we've maybe all heard of this Chinese system for example where this with the social scoring system where you basically tracked in certain cities at least maybe all over the country you're everyone each and everyone is basically tracked by the state with cameras everywhere facial recognition software and what not they track everything when you jay walk or cross the street when the lights turn red instead of green and and maybe also have some predictive policing tools in the background running yeah we've also in in another episode talked about AI judges that that issue ruling also all automatically and you also before mentioned that the US seems to be a bit of a wild west in that area in the EU the the protections seem to be a bit stricter is that also true for Switzerland are we are closer to the EU situation or I think we'd like to hope so I think there are I think there are differences between like even between let's say continental Europe and the United Kingdom in the context of how privacy is perceived and the importance of privacy rights which was certainly taken more seriously here and and continental Europe was in than in the United Kingdom or probably in the United States so that's for sure I think played an important part in in the scope of the regulation and the EU level and we can just hope that the Switzerland this the proximity to the EU and in view of the fact that it's a it's a joined up approach to the regulation of artificial intelligence which doesn't just focus on predictive policing I think we can assume that the this was approach is unlikely to be too far removed from that of the European Union would be my expectation but it seems to maybe historically also be a little bit connected to
with terrorist activities, for example. I mean, the UK had been victim, quote unquote, to several attacks, I believe, in the past, maybe in the US, with the 9/11 attacks, and the aftermath of that, have they maybe-- are they more proud? Because of that, too, maybe less emphasized the privacy of the individual, and say, the ends justify the means, sort of. It's difficult to say. I mean, there might be a bit of that. I think the recognition of privacy, right? My feeling is that this traditionally has been viewed less importantly in British political life, let's say, than in other countries. And I can't really speculate, and reason for that. But I think it has a longer that these-- in the weighing up of privacy rights, against, for instance, freedom of expression, for instance, traditionally, freedom of expression, took considerable-- what I was seeing as weighing higher, I suppose, in the balancing act. And what's interesting, too, is I think we see, at least in the United Kingdom, following the coming into force of the Human Rights Act, the turn of the century, that had a big impact, then, not just on privacy rights, but also on the regulation of the police, because a lot of police activity at the time was not subject to legal basis and law of the sort that we would understand as absolutely essential, let's say, in Switzerland or in other European countries. So a lot of the police activities were regulated in cause of conduct, but not in formal legal statutes. And with the coming into force of the Human Rights Act, following incorporation of the Convention of Human Rights, that caused absolute upheaval in regulation of policing in the United Kingdom. There was a major revision of the whole law there, I was necessary on the basis of that. So I think the problems go beyond the newer issues with terrorism are problems, if you like. The way that privacy rights are perceived has just a different constitutional path in the United Kingdom, would be my feeling. So if I understood you correctly and I agree, then the ends do not justify the means. So let's say hypothetically, there were a tool to prevent or to foresee, similar to minor to report to infer to foresee you committing a crime somewhere down the road three years from now, I don't know. Then the use of this tool would still not be morally acceptable. Or epistemicly acceptable. So I would agree that for me, these things are absolutely connected. It's not just about instrumental idea of preventing crime. And it's also living together in a community where we can accept forms of power and authority that we have. So yeah, we definitely subscribe to your-- Because now, for example, if I look at Germany, currently they have this Ashaffenburg and all those sad incidents that had happened, and maybe in the light of this, the public, or parts of the public, might be proud to say, I don't care about the privacy laws or anything, I would rather the police would prevent that. And they would use tools and find these culprits before they even do something. But as we see, it's not as easy as that. So if we go back to the initial example with Robert McDaniel, they tried to utilize the tools to propose the reasons and the outcome was completely different. So I think the message here is very much that we have to think about-- I also have to think about processing but the ways and means which we want to do that. Achieve the results. Good results. But it was still strikes me about this case you were just going to be gaining. If they think he possibly could be a victim, how come they did not protect him? Well, perhaps they can't be there monitoring him 24/7. So they just kind of said to him, we think you might-- we're worried that you might be a victim, or in fact you might be just involved in some type of gun violence. And that should be enough, you have to take measures yourself. But it's quite unrealistic. But what you can see this too, I would say, an issue to following Osmond, a fit around for his case at the court of human rights on the article two, on obligations on police to inform people who may be at risk to life. So in these cases, two of the police in Britain following this judgement in Osmond, the police have to, if they have some sort of reasonable reason to suspect that someone might be at risk to life, they have an obligation to go and tell the person. But they don't generally do much more than that. It's not like they will move them into what this protection or put them in the-- there's no ability to do much more than warn them. You mentioned prior the despise that it has a self-fulfilling prophecy. So I mean, we've distinguished these different systems. And one can be easier justified, so to speak, with criminalistic, is probably the worst. Criminality theories? Yeah. But still, even with the old tools, even with the statistical tools, the statistics still have an underlying racial bias, for example. Yeah, to the extent that the bias exists in policing, and then this possible that it's reinforced. And of course, the idea that if the police go to certain locations more frequently, then they're likely to uncover more crimes. Or this seems to be very likely. I mean, this seems like a reasonable hypothesis. And I think that's why a lot of people do suggest that this type of hotspot policing might in fact just turn into self-fulfilling prophecy. For by more resources are focused on a particular area, and because of that, this area, in fact, seems particularly-- I don't know, hotspot, crime, or criminal activity, or something like that. And it's also questioned then about allocation of resources, I suppose. But whether one wants to allocate resources to crime, which might be easier to uncover than other types of crimes, or the idea that one might be more willing to police to start an era that's then to focus on white color crime. So has there been studies about the efficiency of such tools? For example, you could say, if we use this tool in Zurich, and the tool says you should allocate more resources in Alchtheten than in the Reis Eis. And then maybe tell the policeman not that, but tell them like, OK, the tool said, city center is likely to be a hotspot. So I think for the early statistical modeling tools, determining efficiency or efficacy of the tools is quite simple in some sense, because the police have a lot of experience themselves, and they probably have a good feeling for what's working in the city of Zurich or not. And with the later tours, that's a real question about how with things like a bread, Paul, and other tools, the real question about whether there are any use at all. How good actually are they at predicting crime? And they're the juries most certainly out. So there was a recent study that was sort of late 2023, I think, by a wired that suggested that the bread, Paul, it was really, really bad. Had a 1%-- I don't know how bad it was, but it was really bad in the context of robberies, for instance. So there was really questions on the basis of that, I think, after the study, a lot of police departments in the states stopped using it. So there was obviously the police work also had it were not entirely comfortable, or didn't think it was particularly useful. And there's been other academic studies, a whole lot of attempts to do randomised controls of the use of these predictive policing tools. But the evidence is slant, I would say. It's very unclear, but also for their use. So that doesn't really necessarily speak for them at the moment. But you never know what I suppose the hope is, particularly with the machine learning tools, and that they would become more effective if you like. So the fact that the argument might be early days, they're not effective yet, but they might become more effective. At the moment, I would say they don't seem to be very effective. That would be the feeling of the academic studies that have been conducted. But as I say, I think these studies are very difficult to-- it's probably very difficult to get a whole of the data to start with, and it's very difficult to conduct rigorous studies on the effectiveness of these tools. Hopes and worries. Maybe from an attorney's perspective. I'm also thinking about the process, because in my education so far, I've not even been really made aware that these tools exist, that police can have access to such resources, to such tools. Which poses the question, is there a way of me even knowing that the police have used such tools in the process of finding my client? And is there a way of maybe countering this? In a court of law, maybe question whether or not these tools, the use of such tools, have been legal? I think in the individual cases, I think, rather not. So I don't think in an individual case you would have much luck as a defense attorney trying to challenge these off pre-cops in an early stage in the investigation. But I think the more effective methods of remedies have been through a lot of the NGOs, and so it's not quite active in drawing awareness to the use of predictive policing, also facial recognition technology. Algorithm Watch is quite active, has a lot of good information on this sort of like policy-making level. I think that's perhaps the more likely way of having some sort of redress then in that context. Are the political level of regulating police activity, allocation of resources and so on? What use of which tools is allowed or not allowed, according to Swiss law currently? Well, the regulation of investigative police conduct is within responsibility competence of the Cantons. Cantons law, police law regulates the use of these tools. And there's a discussion, even within Zurich, but also in other Cantons about the extent of the legal base, whether that's sufficient for reliance on different tools. I think the feeling is that certainly not, there isn't a special legal base at the moment for the use of facial recognition technology, for instance. about the police officers.
Obviously, our opinion, for instance, the stop-related I've been using some types of statistical modeling tools, and they are for sure there's a feeling at least that the cantaloupe provides us efficiently for that type of activity. So, the future, what do you think the future might look like in respect to policy-making and predictive policy tools? I think in the short time, I think there would be very surprised of Switzerland to a different approach to the regulatory approach of the EU, which obviously means that the use of predictive policing of the type that I've been discussing, the algorithmic predictive policing is likely to be so really cartel, so it's not going to be at least in the short term, it's not going to be used. Nestle and Switzerland, but at the same time, I think that we should be very much aware of the issue of surveillance. That seems to be the issue of availability of data. This seems to be the issues that are going to be very important in the future, and there's certainly also going to be putting a lot of pressure on this distinction between investigation and repression in the context of policing, so that something that will have to be looked at, I think, on the political level. So, Sarah, what are your hopes or worries for the future of predictive policing? I do quite strongly feel like in liberal democracies, commitment to the rule of law, to the idea that certain processes must be respected by the state, the prevention and investigation and punishment of crime, I think that's very important. It's not just about the accuracy of decision making. It's also about the way in which decisions are reached. The process is absolutely central to ideas of fairness or justice or accountability in the rule of law. So, I think when we think about things like crime prevention, we should definitely think about values such as dignity, community safety, promoting a system of democratic policing, which is accountable and responsive to citizens in society. My hopes with the future is just that we try to maintain a commitment to the rule of law. I think that's the essential point that we focus attention on the importance of democratic policing in the rule of law. Okay, Sarah. Thank you very much for being in the podcast today with us. It was very interesting. Thank you for having me. What's our pleasure? What's the interesting interview? All the things, different aspects that we might have to get to know now. What do you mean? What's the hang? Yes, I've been hanged up with self-fulfilling prophecy. What's the problem? I think that's the right thing to have the right to understand all tools and understandings. Also, when you look at a more controlled way, the statistic says that there are more professional deputies, for example, you also get to understand more professional deputies because you're in front of the way. And that's something that maybe remains with the new AI-based tools. It always tells me that it's a bit different from the weather. You have a tool that says here and here, it's about what's going on. And I thought it was interesting that you said that it's not so much about the detail, the usefulness or where it's going to be, is this little bit of a certain kind of twelfth. That also has a bit of a different weather. Actually, the question is the weather is about accuracy. For example, in Switzerland, the weather is bad. Or so, or a person who says here, it's in the left knee, tomorrow it's raining again. How do these tools look exactly? That doesn't really seem to be really present or not present in the detail for everyone. So exactly as the experience that the politicians have to do and on which they have to support their own interests. Yes, and what's also in my problem with this tool is that it's often called a black box problem, which is what I call a "opacities problem", in terms of the actual opacity. Exactly, because it remains opaque. So, under-suitable? Yes, and you can actually evaluate which data you can implement. Depending on the system. But what the system does then? Why does the system come up to the bottom of these data? The muster comes up to the one that comes up to the one that you can't always follow. That's what we can, for example, from ChatGPT. So I can't write a question in the end and then come to the answer, which sometimes is true, sometimes not, but we have the answer to the answer that is what is not present. I don't know. You have ChatGPT, when there comes more spontaneous interests, so ChatGPT and Weather Forecast, and at these two points we get the idea that it's generally a-eight-holes, the problem is that they don't start to be so good with their results and they have the time, but then they are really good at getting to know themselves. Because they learn themselves, so that we can be very curious whether or not they can or not. And something else we already talked about in the intro is that the relationship with preventive to repressive police, because the problem that Sarah also mentioned, that we have a little danger that somehow everyone and every potential leader, once seen, will be basically so that we then somehow have a problem of the unsolved attitude of the presumption of innocence. Exactly, we have to maybe listen to the lines and hear the speakers, maybe something specific, or concretize what you are talking about. So currently it's the least in our right-wing system. There is this two-function of the police work, so you have said that one is actually prevention, that you say security police officers, the police are your friends and helpers, how this is so colloquially being done. And on the other hand, the police have extended the right-wing police officers, so the police officers have started to have a job. There are these trainings, so the basic idea is not everyone's attention on the street, potential attitude, or potential, but it will not be spoken by hand, but it will also be the unsolved attitude. Exactly, that is part of the principle, right? Exactly. Every person is first of all unsolved and not as potential attitude. I don't know about the word potential, because potential can be argued. Are they all? But you are right, so you don't go from there, it has something bad to do, but you go from there to there. Of course, there are private rights and so on, the police are not controlled, they are not at all violent, they are forced to start or are forced to. And there it is clear that this project of policing or this un-adjusted danger that these modern tools, these border connections, the whole time, the people are waiting although they still haven't done anything yet. That is probably also the reason why the EU is very careful with AI-protective policing tools and the basic is forbidden in the EU AI Act and as high risk AI tools, I think that I really have memories of it. My mind is also like that. Exactly. What I still wonder was the observation, so to speak, the end of the pyramid development, then a tool to have that actually make the whole time about everyone and everyone of us can do it. Claude Bertschinger becomes at 25th March, 23rd, by God, over the Amplifan. Because he is a big 44. Exactly. And green eyes. And then you have this statement, or with 87.5 percent, probably over the God Amplifan. And what do you do with it? And it's funny and important to be aware of what he has said. It also goes to the process of this result, at least exactly as important. So the tool is not suitable for the middle, you can't even, if it is possible, a such tool to program and implement. You can't do it, or just with big conditions, because you have the rights that we have spoken of, the fact that you actually have it. That's a perfect end word. I thank you for listening. Yes, thank you too. Thank you. Feedback, we are very happy to talk about our email address,
[email protected]. Follow us, for example on Spotify or Apple Podcasts or the Podcasts.airerwall. Nadine, thank you very much for taking this episode with me. See you next time. Or feed the herd.
PodcastTranscript AI