 The next talk will ask the question, will algorithms solve the problems of climate change? More data, let us monitor the Earth, optimize solar panels and generally understand problems. But is it sustainable? Jens Aulig from Algorithm Watch and Friederike Rode, Institute for Ecological Economic Research will talk a little bit about this problem and I'm looking forward to this talk. So I give over to Jens and Friederike. Hello, good morning. I hope I'm audible and thank you so much for having us. So we will talk about our project called Sustain, the Sustainability Index for Artificial Intelligence. And we are very excited because it's about the first time that we are going to present this to a larger audience and I think this is a good way to introduce this to the world. So who we are? As our Harold already mentioned in the introduction, there are two organizations involved but there's actually a third one that is not present today. So there's the Institute for Ecological Economic Studies and there's Algorithm Watch, both based in Berlin. There's also the Technical University of Berlin or rather the Distributed Artificial Intelligence Lab there and we are running this project called Sustain, Sustainability Index for Artificial Intelligence where we want to find out over the course of the next three years how we can measure and index criteria for the sustainability of algorithms. There is funding for that, luckily. The Federal Ministry of Environment funds this project and this is based on the resolution of the German parliament, of course. So when we talk about sustainable AI, there are lots of questions and some of them were already mentioned in the introduction. So will we be able to stop climate change through better monitoring our planet? Unlikely so, probably there's no magical solution to all of that but on the other hand it is very clear that once we have more data and better data we can of course make more informed decisions. So looking at where AI is actually helpful for sustainability is one part of the question and the other part is how does it affect sustainability? So there's a lot of talk about energy consumption, how much energy do data centers suck up and there are already a few numbers that are usually thrown around in that context. I believe that Friedrich will at least mention that later there are numbers that say this algorithm takes about the energy of that many cars etc. Is that all true? Is there something that we can put into concrete numbers that remains to be seen? But sustainability goes a bit further. It's not only about doing something that is good for the environment. A sustainable world is a world that is worth living in and a lot of the magical artificial intelligence solutions that are proposed. Once you look behind the scenes you will notice that the emperor wears no clothes and what is behind all the magic of the AI is maybe good old precarious work conditions. People who are clicking through mundane tasks and maybe don't live in a very sustainable environment. We call these people click workers and a lot of the things that you can see maybe in your speakers that you have at home that answer all the questions are maybe not the result of some nifty algorithm but rather the work of click workers. So if we put this aspect also into the question of sustainable AI the question becomes much larger than what is good for the environment. It becomes a question of how can we live in a livable world and the question is very real and we see things affecting that every day. We see questions of discrimination racial bias in algorithms. We see questions of Google translating things on a databases that perpetuates gendered bias and sexist stereotypes. Is this really the sustainable world that we want to live in just with some AI? I think not. By the way artificial intelligence is a huge word and it sometimes gets used in contexts where people don't seem to understand what they're talking about. So it's easier to narrow it down. Algorithm is another word with an organization like algorithm watch. Of course we don't watch the performance of I don't know quick sort or something like that. We are interested in automated decision making systems. So every time we're based on data and statistical transformation of that data maybe in the form of neural networks or machine learning or maybe not and decisions are made based on that and they have an impact on society. We're interested in that and right now our interest with sustain is sustainability in that context. Sustainability is actually a really really old concept much much older than than a century. Funnily enough it comes from Germany. So nowadays you see it everywhere but in the beginning it just meant a very simple thing from forestry. To be sustainable just don't cut down more trees than you planned. But in 2015 the United Nations picked this up and now sustainability is everywhere in the form of the SDGs the Sustainable Development Goals that were adopted by the United Nations General Assembly. Basically a set of goals that will make our world more livable. We want to have no poverty in this world. We want to have zero hunger and of course life underwater, climate action and life on land are things that are all playing in this sustainable development goals. There are sub goals and everything around that. To make a livable planet at least some of these goals have to be considered and a livable planet also includes artificial intelligence and algorithmic decision making. So let's bring these two together. It is all linked and all brought together by these goals interacting. So there are questions of ecological sustainability. Let's not burn the planet. If we fail to do that everything else becomes less important because we don't have a planet to plan all the other steps about. But we also want to make life livable for human beings. Having something that is ecological but requires a lot of people working in precarious work conditions is not the goal. We want to have a society that takes out the full potential of every human being. And certainly if everything crashes the economy we have an ecological sustainable world and a socially just world. But economically we just get more and more poverty that doesn't really work. So everything is connected. Frederike will continue with this. Thank you very much Jens and good morning from my side. Now that we have the question is how do we act if there are conflicting goals between the pillars. For example if the protection of the environment is not appropriate for all social groups because it's more expensive or if there are conflicting goals between ecological social and the economics field which in our current economic system mostly is the case. Then we should come to perspective which allows for a kind of prioritization of the three pillars. And the answer to this question is the perspective of the nested dependencies. Without the environment the society and the economy cannot exist. So our planetary boundaries are the basis of our survival. And as Jens already mentioned there's a codependency of all the three pillars. So in other words this nested dependencies model acknowledges the inherent value of the environment and prioritizes the health of our planet and the viable society over economic gains. Next slide please. But what are we actually talking about when we want to connect A to sustainability. From my point of view there two differing perspectives. The first one which makes up the main motivation of this project is the question of how can we assess kind of all existing AI based system or ADM systems from a sustainability perspective. So with an AI based system as Jens already mentioned we mainly refer to machine learning. A certain area of weak artificial intelligence because not everything is artificial intelligence what is sometimes called artificial intelligence. And we especially refer to deep learning based on artificial neural networks. I will leave it like that for the moment. But what is the aim of our project. We want to develop a framework for assessing the impact of existing AI based systems. And the application with regard to their social ecological and economic risks. That is we want to develop a holistic perspective for assessing AI based systems and showing which screws you can turn to develop drain and apply them in a more sustainable way. That is one perspective. We call it sustainable AI or sustainable AI based system to make it more precise. Another perspective which has recently come up and which Jens already mentioned as well is the question of AI for sustainability. So how can we use develop apply AI based system in favor of our climate or in favor of environmental goals in favor of social goals. Sometimes it's called AI for earth or AI for good. But that's in our point of view that's a kind of different question because these kind of applications only make up a minority of all AI applications that are existing in the world. Next slide please. In a way we want to kind of integrate those two perspectives. The first aim is to create sustainability criteria for principally all AI based system to or at least to bring us a little bit closer to the goal to develop this kind of assessment criteria. And on the other side we will examine three case studies in sectors which are of high relevance for the sustainable development goal such as energy mobility. And we choose which we chose another one which is online shopping because we think the impact is very high and there's very much AI based system used in this domain. So we will examine in the case studies what what can we expect from AI based system for the energy transition the mobility transition and so on for the sustainable for reaching the sustainable development goals. The results of the sustainable criteria and assessment and of the case studies will collected in sustainable AI imports. If our criteria work out very well perhaps we are able to create an index showing like what kind of how is the the the certain AI based system how can we assess it in terms of sustainable sustainability and perhaps we can this can be displayed on a website that dashboard or an app and we want to develop policy recommendations and guidelines for sustainable AI development which mainly are referring to the community who is training developing and the organizations who are implementing and using AI based system. The next slide please. It's a little bit small perhaps you cannot read it but it doesn't matter that much. Our first thought within this project was how can we find a structure out of the universe of impacts related to AI one can imagine. Our first answer was this kind of figure which shows three levels which we consider as appropriate for answering the question on how to measure or to to to look at the the impacts of AI based system. The first one is the algorithm itself or we call it algorithmic system because it's not only the algorithm it's also about not only about how the algorithm is designed but also about how the data is gathered what kind of data is used what is the utility function of the AI based system what in our perspective is a really important question so that's the most inherent impacts you can imagine when it comes to AI then we have the application level here the question is in which sector in which domain is the AI based system used and what social ecological and economic impacts or risks are related to the application of this AI based system so you can for example imagine using AI for medical diagnostics or for advertising you can use it for predictive maintenance or for selecting job applicants and as you can imagine the impact related to this certain area is very different might be very different so that's a no nothing so that's a important question and the the third one is the like the kind of big question of the societal level the systemic level it's like what impact does the application of this AI based system in a certain area does on our societal goals and related to this question is the question of how can it can it help us to change structures practices for sustainable development so I think the question on how AI can help our society to get more sustainable is mostly related to this third level and not to the level of like how do we design the algorithmic system so our first step was to to collect the various literature related to the impacts of AI and work out some preliminary criteria and we tried to find out of the existing literature and we also created some new criteria especially in the domain of economic dimension in the social dimension we found those criteria you can see here it's about transparency traceability fairness and justice harm avoidance non-discrimination a quite important issue for every AI based system with which is interacting with humans or which is making decisions about humans privacy and data protection for sure and respect for human autonomy and freedom human oversight and robustness and accuracy and at the moment we are about to trying to develop indicators for those criteria or how you can on the basis of which you can assess some systems such like such like were there did they make pre-tests with the AI based system and stuff like this next slide please the economic dimension in our perspective it's the most difficult dimension for criteria because it relates to very different levels of society so in our perspective we collected well we tried to collect some of the criteria which might be feasible foreign assessment which are like texas so is the organization paying texas data protection data collection practices but also aspects like market power anti-competitive behavior privacy policies anti-corruption labor market effects which are like more on the structural level income distribution working conditions as Jens already mentioned the click worker issue and stuff like this and impact on competitive supervision and robustness and accuracy which we had already before but perhaps we can also see it as kind of economic criteria and then there's also going on a kind of growing discussion about the ecological impacts of AI based system and it's mainly about the energy consumption and CO2 emissions for example the share of renewables the ecological but also the ecological impacts on consumption patterns resource consumption or for example waste and recycling do we need ever more devices sensors and stuff like this for applying all this magical AI systems and does this really make our world more sustainable or does this infect harm our environment more than the saving environment at the moment we are thinking about how we can distinguish between different levels of AI deployment because they might entail very different impacts some of you might know as Jens already mentioned at the beginning a study from Emma Strube and colleagues and the results often are cited like this the carbon footprint of training a single AI is as much as 284 tons of carbon dioxide equivalent five times the lifetime emissions of an average car but in fact this number is not true that way because this number refers to neural architecture neural network architecture search so it's not the same as training a common model but it's like the thing people only do for research purposes and which costs like two millions of dollars which no normal organization would do to train like a single model for like showing me what kind of genes I can buy if I because I bought this one so recommendation systems or stuff like this so we really have to look a little bit closer on how do we really want to assess the impacts of AI based system and that's why we are trying to distinguish between like three levels and the one is the network optimization or network architecture research with which mainly refers to the research area then we have the training of the artificial neural network and then we have the inference process so the process where the model really is applied and makes decisions or helps to make decisions so we think we should especially when it comes to ecological dimension it's important to distinguish between these three levels and with our project we would like to open the black box a little bit more and create awareness among amongst the involved actors about the risks and we about the risks we buy into when we try to solve like every optimization problem no matter how small with complex deep learning algorithms and what can we do to avoid those risks and improve AI based system in a sustainable way thank you very much okay okay sorry thank you for your very interesting talk so I think we already know but we know it much better now that the world is really complex so and I think this research is really important for society but for me when I listen to your talk and I think this could be perhaps the first question before people hopefully started to put questions in the pad you can find it in the Fablan and you can also put your question in German there if you want so but for me it was like but what can you know when I listen to these complex things what can I do as a person depending on your research well in the end I think you will have something that for you as an interested person will help you to understand the whole situation better we we started out with a with a very ambitious goal to be honest Frederike said maybe in the end there will be a website or an app or some kind of easy to look at number and we're not really sure if this is the end result what we are sure is that we will produce information and that we will produce both policy recommendations so we want to influence policymakers and we want to hand out something for the hacker community for the developers with the AI development guidelines so this will be something that comes out of that and if you develop software and care about livable planet this this is something that is very concrete for you but I agree yeah that's this is a huge problem and there's no easy answer to that and we are just starting this journey of three years yeah but thank you for starting it so this would be my second question is this the first time that people think about this problem are you the first ones doing research about this or is this is there already a base also a base which is also practical used in in algorithms or the internet well there is research about this issue but it's not like from the sustainability perspective it's we have a really really huge discussion this discussion on the ethical issues which we refer to as a social dimension and we have a discussion on the ecological or a growing discussion on the ecological impact energy demand and stuff like this co2 emissions the economic dimension I think it's a little bit more connected to other discussions like yeah platform capitalism and stuff like this but I think the perspective like bringing all these issues together is not yet really well developed and the other question how it is like how we can see it in reality I think not that much because it's not like the question you you you you said before was like what can I do as an individual but the problem is as an individual you even don't know if you're interacting with an AI based system and that's the first problem so but perhaps you can do stuff like ad blockers or something like this because we we see that AI is really is often used in advertising and that creates really huge amounts of like CPU load and stuff like this but from a personal view I would be happy to have more you know transparency in this and I think your research is could be part of it to find out more as a as a user about these problems and then adjust your own behavior to this to these things if you're interested in so it's a pity but I think your your talk was kind of you know it's so complex and it's kind of overwhelming but really really interesting so we we don't have any questions perhaps the people are still at breakfast and it was like oh I have to think about this so perhaps you have your contacts on on the pipeline or am I right let me see there are the slides and inside the slides you will find contact possibilities are well I will upload something with the contact contact addresses yes so I think that would be this would be great so if people you start thinking about it this and can come back to you later and perhaps there will be also some nice input because there are a lot of clever and smart people at the ccc events and perhaps somebody is coming up with ideas or have another questions and so ask these questions later via email or something like this and I thank you very much for this really important input and I'm looking forward for your research results perhaps last question how long will this research last when will it be finished when will you start publishing papers and stuff like this we will be publishing things throughout the course of the project but it is the the funding goes over three years I think the first sustainability report will come in about February 2020 stuff like this yeah and we will have little events yeah so we're looking forward to this and perhaps you will come back and have another talk and tell us more about your project and yeah thank you very much thank you thank you