and something probably everyone agrees on is that ai should be designed following ethical values and moral principles but there are no practical and clear guidelines of how to do that yet so how should we implement ethics in ai and how can we bridge the gap between philosophical discourse and practical application in companies these are questions that our panelists of the next panel want to tackle carla Hustedt is the head of the ethics of algorithms project at battlesman foundation she has consulted the german parliament's ai on cat commission on the issue of algorithmic transparency also on the panel is hitrana she's member of the group board of management of volkswagen responsible for integrity and legal affairs dr andreas davis will be one of the participants he is the co-founder of kip project software for secure and privacy focused data science and machine learning and they will be joined by professor dr marcus gabriel he's the chair of epistemology modern and contemporary philosophy and the director of the international center for philosophy at the university of bonn
so quite an exciting panel it will be moderated by ai consultant author and fellow podcaster peter zebak and let me hand the stage over to you peter with something rene descartes almost said it is not enough to have a good ai the main thing is to use it well peter the floor is yours thank you very much dania also a warm welcome from my side to the ethics of algorithms panel hosted by k.i park germany ki park germany is an open ai initiative that promotes the application of ai in germany and if you dear listener have a specific question about how to apply and industrialize ai or you want to learn more about the initiative please visit wki park d e or join the keynote tomorrow at 5pm for now we are going to address our ethics of algorithms from a theoretical discussion to practical application topic we're going to do that along a guiding question which is our algorithms biased and i'm going to start with you marcus for a starter please tell us
what is ethics and therefore what is ethics of intelligence well in general ethics is the philosophical discipline which studies the nature of moral facts and principles moral facts are facts about what humans insofar as we are humans are to do or not to do what we absolutely ought to do is what we call the good what we absolutely ought not to do is what we call radical evil and the rest lies in between and so the discipline of ethics tries to figure out what that is so how universal value can be discovered in tandem with different disciplines so ethics in the 21st century can only proceed in a transdisciplinary fashion including of course politics society and the business world so that's what we do in the discipline of ethics now ai needs an ethics for the simple reason that ai systems are used in the context of human machine interface systems such as companies news feeds etc right we have all these applications and these applications structure data in a certain way and therefore right
make it more likely that a certain decision will be taken or some other decision this is why ai systems are subject or ought to be subject to ethical considerations and the future discipline of ethics of ai has just started thank you for this introduction marcus carla what are the sources of ethical problems occurring in ai systems when we talk about algorithmic decision making i actually tried to avoid the term ai because i feel like it gives a wrong impression of what these systems can do and what they cannot do we also call them social technological systems for the reason that it's not just the technology that is the cause of bias of mistakes but it's mostly the human decision decisions that lie behind it so it's the goals of the system of course it's the data that goes into the learning algorithmic system it can be the code itself or it can oftentimes also be the way a system is implemented the way it's integrated into an organizational structure because in many application cases in
the health sector education society relevant fields it's still humans at the end that interact with the machine thank you for that we just heard in the hour before already from Merantix as the example of the medical applications moving to hilton do applications exist that should be excluded from the beginning for ethical reasons from the involvement of algorithms i could think of in the case of miners autonomous drones elections or in the pre-selection of applicants oh hildefoot i see you still have your mute on yeah you may want to start answering from the beginning we assume that this is the case although i believe it needs to be decided individually for each specific topic for the automotive industry for example we do have dilemma situations in traffic in which such codified rules as you as you just explained must not be applied at all very good and moving to our fourth our participant andreas do algorithms appear biased here and there because the responsible data scientist has selected the wrong or the non-representative characteristics i.e the
features from the data that's certainly part of it i mean if you look at for example supervised learning where you want to predict a certain outcome a classical example would be hiring so we have a machine learning system for example that trans tries to reproduce hiring decisions by people and we train that system by giving it historical data for example so like hiring decisions that were made in the past and uh maybe there were some part of these hiring decisions that were made made because of discrimination for example against a woman or ethnic minorities for example and the algorithm basically doesn't know about it it just tries to reproduce the decisions made by the humans in the process and it can do that the better the more data it has that um can predict the outcome and if this data includes for example so-called sensitive information like the gender or information about ethnic yeah i think of information ethnical data then um this algorithm can do that better and can of course also reproduce this
bias and this discrimination that was inherent in the training data so in that sense yes um but it always depends on what kind of data we feed to the algorithm and how easy we make it to also infer on these sensitive attributes from the data that we have provided to the system thank you very much dear andreas this concludes kind of our first round as you as a listener will realize it's not that easy in an online fashion to do a real discussion in between the members but i do propose that with the upcoming questions that we have that you do try maybe to answer or if you have a different opinion in your answering of what you just had so moving back to marcus uh will we ever get a strong ai as according to john searle the chinese broom that will think like humans and will eventually overtake us in terms of human intelligence question mark and if not when it comes to the ethics of algorithms should we maybe concentrate on the
non-thinking the non-understanding i.e the weak ai well yes i think you know like there i sighed for very different reasons the arguments are flawed in john john's chinese rune argument but uh so i don't think that there will ever be such a thing as what john searle has called strong ai so you know human level or even super intelligent systems the reason that i give for this it's a very different reason than on that room argument which is outdated given by contemporary ai systems but the real reason is that intelligence is arguably a feature of a certain living system and systems intelligence has evolved as the capacity to solve a given problem in a finite amount of time for instance you're hungry and therefore you need to solve that problem and you have evolved in such a way to find to select from a potentially infinite solution space the best solution and you're more intelligent than some other system if you find the same solution faster right so we can measure that property and build
models that replicate intelligent behavior in animals in a technical fashion that's i think what ai systems are doing ai research right and we have advanced way beyond the semantic kind of or the syntax the syntactic engines that john cell had in mind but as long as our you know uh systems are not literally uh based on neuronal biological material which they're not they're not going to be intelligent however in the strain in this sense however the real problems do not lie there the real problems are in weak ai right as hilton vanna has mentioned certain diametric situations which are not solvable right because they're strictly dilemmatic are therefore to be avoided right so we have good reasons for instance not to go to level five uh automated driving in certain conditions for those ethical reasons right so automate as far as possible but there are limits to automation ethical limits right and i think this is what we ought to discuss so the real impact of weak ai for instance pre-selection algorithmic buyers new feeds
etc and in order to do so we need to understand the integration of a weak ai systems in the human life world because our intelligence has changed in the last decades we have become more intelligent due to the use of those systems because that's the real intelligence explosion and we have yet to find a way of dealing with the fact that we have made our ethics explicit right i mean we now know much better how we judge because we can see patterns in those data thanks to the use of weak ai systems this is i think where the action is thank you very much for putting that perspective which i personally can support but nevertheless i do propose that as you did the other answers members do maybe give their perspective as well carla which organizations and which persons in the organization are responsible for the implementation of trustworthy ai is it the producers of ai system the operators is it others um all of them actually like as like i mentioned before the sources
of the problems are multi-fold and so are the solutions so i think there are certain technical issues of course technical basis that lie in the hands of the organizations and individuals that build the technical parts of the systems and of course those organizations and individuals also need to communicate how the systems function what they can do and particularly what they cannot do towards the people who then use the systems in the organization to implement the systems because we frequently analyze cases um where things have gone wrong with good intentions and what we see in these cases is oftentimes that it's actually a communication and misunderstanding problems between different stakeholders and that the people who understand the social problems that need to be solved like in education sector healthcare security police forces they understand the complexity of social problems but because of the narratives that arrive from silicon valley from the tech sector overall of technology being a simple fix for complex problems they do not take the necessary measures to implement the systems well things
like competence building among people using the system things like transparency which is not just a technical issue but it's also about communicating towards the people affected so what we really need is better communication between those different stakeholders and all of them understanding their responsibility and the role they play in the ethical implementation of the systems and of course we need politics to kind of create the right framework to make all of this possible okay so everybody takes their responsibility within the ecosystem and needs to communicate uh hiltrude how does that work trustworthy ai working with more or less bias algorithms we're talking about that today what needs to be done so that we can get towards this trustworthy ai as desired by the eu commission yes i mean we have heard from from marcus and andreas already that trustworthy ai is based also on those uh who developed the algorithms so so some factors that need to be considered and and caller mentioned that already also is that we need national or better supranational institutions
that set ethical standards for these algorithms but what we also need is ethical training for developers and programmers of ai related software and we also might need especially for developers and programmers of ai related software in critical situations some background checks and we need pairing concepts that in the quality checks of ai look not only for technical but also for ethical quality good thank you for that andreas today in europe it's my personal feeling we experience a kind of a biased zone between the law on one side and morale on the other side uh we are legally obliged just as one example to differentiate between men and women but if ai systems will be going through certification in the future will that mean that europeans will become more thoroughly streamlined in their thinking and actions i would say almost similar like you know let's take the example of china but only on the basis of other values so will that mean that eventually uh europeans also would feel like big brother is watching we are
not necessarily i think i mean what will definitely happen is that as we move to digitize some more problem problems and more and more processes and use machine learning for any kind of decision systems really to um govern them we will like find or uncover a lot of the problems that are hidden now because so many processes are still done being done manually and there already is discrimination today there already is like buyers and a lot of problems but it's still largely invisible i would say because there's no data that is easily available to analyze then i think with like the further use of machine learning and data science we will see more of these problems and that will also allow us to think in a more structured way about how to solve them you know and i think if you look at data privacy we can um also take that as a good example how we could do regulation in machine learning or ai because data privacy or privacy in general is a human
riot um so we have to derive european law european regulation from this principle that privacy is really the right to every human and then from this law of law so to say the requirements that we need to fulfill when we build systems for example that process personal data and from these requirements which are legal requirements then we have the technical requirements that are implemented by the engineers and i think for the ir machine learning i don't see a reason why it shouldn't work like that you know working from fundamental human rights to a legislation to a law that can be applied then and from this law thinking about how we can do the technological uh steps to ensure that it is actually um complied in here in europe okay so working on gdpr as a good european example so to say marcus i come back to you a little bit of a different topic i believe very important as well shall the police patrol more in areas where there have been more crimes in the
past shall crime be correlated therefore with residential areas ethnicity and eventually the human face and even more general shall we as humans involved in pre-crime i.e in predictive or preventive police work well these are some tough questions so like we know that predictive policing and complicated ways where you know say clear the uai which is a prominent example in the united states has been in use for almost a decade i used to live in san francisco and you know even back in 2013 uh or now you already have good apps for instance also for new york city where which trace the you know your probability of getting uh you know marked in a certain part of the of the hood etc right according mostly to algorithms and earlier data so this is already in use not everywhere of course and this is highly problematic to a certain extent because we do know that existing correlations between you know a certain population and socioeconomic conditions and the probability of illegal behavior right that we increase the
tie between those correlations so and this might then lead to forms of for instance racism so racism let me just say what the mistake is i mean there are so many things wrong with racism but the fundamental mistake is the that the racist confuses a correlation between a the phenotype of a human animal and a certain kind of behavior right with a genetic explanation that's what racism is if you think that you know white or the hispanic or whatever person does a certain thing because she has that phenotype that's just a bad explanation for what happens in reality however there sometimes are statistical correlations between the phenotype of a certain population and a certain kind of behavior but the explanation for this correlation is purely sociological right this is how you dispel this is how you refute racism now the problem with predictive the ethical problem with predictive policing in particular the use of clearview ai which is also muddled with the american alt right so there are some particularly bad aspects about clearview but
be that as it may the general problem is going to be that you reinforce right these correlations rather than overcome that but we shouldn't just criticize ai there right we should rather invent positive ai so this is a business case for a system which precisely does the opposite right so why not use the the you know the method in order to go into a specific neighborhood where you measure these correlations and take appropriate ethical uh matters right decisions so you there's always a positive use so predictive policing as it's now practiced is ethically harmful but we can use the same algorithmic systems in order to do the morally good and i think that we need to change the discourse right from risk assessment discourse which is what we're mostly doing in ethics of ai right now to applied ethics of ai meaning ethical training in and by design this is how we prevent the bad things from happening not by you know criticizing them after the fact but it is sorry marcus need to jump
in here just because of time did you want to finish your ascendance at least sorry go go to the next person i made the case yeah carla do we need new norms new laws for the digital age and if so what role can the the algo dot rules that you coordinated developed for battlesman foundation play um i do not think that we do need new fundamental rights for the digital age because we're actually quite well set on this regard however what we need to question is whether the mechanisms we have currently in place to protect fundamental rights and freedoms whether they still work in automated processes and oftentimes that is not the case for example because oversight bodies do not have the the rights for example to access code in certain situations or they do not have the competences they do not have the right amount of people to deal with increased efficiency and we need to analyze whether these oversight mechanisms still work we need a strong civil society that can take on the
role of watchdog organizations that can spark a societal debate about all these questions that were just raised i mean it was said many times in this discussion now that the use of automation kind of forces us to have this discussion again what's between the super evil and the very good but what we see oftentimes is that when something goes wrong with technical systems people say okay it went wrong let's stop using the system and then the debate is over and we failed to actually take that opportunity and in order to do so i think it's really key to um yeah have the debate focus not just so much on like the economic questions but really on societal questions and to build a strong basis for that we need competence building in society we need civil society to become stronger and also interact in the digital sphere and now i didn't answer the second part our arguments they're actually very aligned with what i just said because those are not rules that define what fair is
or what's unfair in a certain situation because that depends so much on the application case but they rather give certain recommendations on how to deal with possible mistakes for example by making sure that systems that people are informed when a machine is making decisions about them or that the decision is explained to them and that for example there's mechanisms for appeal questions of again i.t security of privacy are part of the aggress so those are mostly process standards rather than standards that describe the system itself thank you very much carla i had my beeper go off here that means that we need to slowly but clearly move to our closing statements and i i want to start with you there hilton um i think you've all thought about and you have your personal views that you're going to share with us now one question that i want to give each of you with is that as i've been listening this morning to the rise of ai speakers there was kind of a general uh almost
like a red line which is always about we europeans we germans we europeans we make norms and we do ethics while uh our american friends or the chinese asian people uh start first doing things so maybe if you do your closing statement if you can but you don't have to uh refer to this um this theme so hillthroat uh starting with you yeah thank you very much i i think um what we have learned today is that we have to incorporate every stakeholder in our decision making and competence building of ai and yes the european union has a very special chance challenge here whereas a lot of development in china is driven by the government or in the us it's driven by corporations we can drive everything that we develop based on the needs of the people and i think um we have a good um a precondition for that we do have many ethical stakeholders in our society from philosophers to theologists and and also a good uh supranational institutions and um i'm pretty
optimistic that our ethical standards uh will serve the people thank you very much andreas moving to you with your closing statement so for me um yeah as i said we had a lot about how privacy and ethical concerns are holding us back in europe but i think we shouldn't really think about it in that way because um it's really not the case that we can only have massively successful products and platforms like for example facebook and twitter if we accept that they very like our privacy and it's also not true that we cannot have ai systems machine learning systems that automate a lot of things if we accept only surveillance or you know other paradigms like the chinese model for example for exploiting and using personal data especially i think um what we have here in europe is really a very good foundation and data privacy and human rights and we should really make use of that and see it as a strength because we can see that not only in europe but everywhere in
the world people are actually craving for these things for privacy for being taken and being in control of the technology and not the other way around and i think we can if we can deliver on that we would have a big advantage in uh in terms of the competitivity of our products compared to the us and to china as well so i would see mostly as an advantage than a disadvantage what we have here in terms of privacy and our german angst as some people said this morning thank you andreas marcus your perspective well i think of ethical value as constitutively universal and global so i you know sometimes some chinese philosophers are morally more advanced than some other oh i think that the global challenge for humanity so i don't want to think about this as europeans if we come up with european values they are therefore false no value that is local can be an ethical value so torture camps are as evil in china as they would be in whatever baden-wurttemberg and
that means that ethical value if it exists is universal anyhow so if there's anything that we can do specifically here it has something to do with societal training etc and let's not forget there are way more professional ethicists of ai in the united states of america and all those big corporations then we even have professors of philosophy in germany right so or china i know for instance people in the ai world in shanghai they're all fully trained highly professional philosophers so i think it's an illusion that we are leaders in ethical training we can become those leaders for the reasons that are typically given but i think it's also very crucial to see our own mistakes for instance we do have racism in our basic law in the grunt gazettes which is why the jurors are now changing it and i think we should also start by criticizing ourselves rather than fantasizing about evil china or evil united states of america they are not as evil as we think the chinese communist party is evil
right but this has nothing to do with our chinese friends thank you marcus uh final words for you carlan yeah thank you i think i actually have some thoughts that connect very well to what marcus just said i mean first of all um i think what andrea said before the the ethical development of technology is not just something that's morally right but i think it's actually the only long-term sustainable way to move forward with digitization because if we do not do this without the trust of the people it will definitely backfire in the long term and secondly i think when i look at the ethics debate and i've been following the development of this in germany in europe for the past four years and i'm happy that it's becoming like raising on the political agenda but it's so urgent that we kind of move from all these past words like transparency that can mean so many different things to concrete action and that of course means that we do need to find ways to deal
with mistakes that will always occur in these systems that we find regulatory mechanisms that allow transparency oversight accountability but it also means that this is connect connects to what marcus has said before that we need to find a vision for what digitization means in europe that is not just in contrast to what's happening in the us and what's happening in china because right now digital policy making on a european level is often very reactive it's always like one step behind and the reason for that is mostly because we're kind of looking at the technology and then trying to fix the mistakes that happen but we do not have a societal vision and this is really key we need to use technology kind of as a mean to dream like we do with science fiction but in the focus should be the change we want to see in society not the technology in itself we always need to see it as a tool at the end thank you very much carla facebook twitter and google have
all switched to algorithms that now moderate the content of their platforms instead of people i sincerely hope that you dear listener and a participant as well have enjoyed me as a human being moderating this panel although as i suggested in the middle it's not that easy to have real discussion uh come up um let me say thank you to our participants and of course the ud listeners i've been asked to make one final call to action by our sponsor ki park germany to share with ai practitioners and especially startups to participate in an ai ethics lab so ki park want to make ethical questions tangible for developers and bridge the gap between actual source code and normative ethical guidelines a transparent open source project should help to shape best practices for ai practitioners towards an ethical by design framework you want to join please drop ki park a message or join them at swap card where you are at the very moment i say thank you to our sponsor k i park germany as we
presented by deloitte investor and salonis as well as to fabian and veronica from rise of ai for making this panel possible looking forward to meeting and seeing all of you the next time