Creative Summit | AI Ethics - which lessons to draw for the new economy - Imane Bello

[Applause] [Music] uh so friends we have now the very very last uh panel not panel conversation and input of the summit of this metabolism hackathon that we were putting on since thursday or since wednesday rather with eve global so full week of learning building interacting making friends maybe enemies even but this next speaker most certainly is not going to be your enemy unless you are an immoral sentient ai because imana bellow who i happen to have met on an event that zora put on a few weeks ago in paris at the epcc randomly got talking and her expertise is quite relevant currently and that is the field of ai ethics now one would say mikhail why would we talk about ethics in the context of engineering and technology building but technology is not neutral technologies biased technologies ideological and technology can have more and ethical implications and giving these emerging new ai ecosystems dali being a really good example a project that i really really like um there's quite a lot to think and speak about so um that's exactly what's going to happen and there might be some wi-fi issues because humana is on the road right now so um i do hope that she has a stable connection to you if not we might have to repeat it some other day but please um i would like to give the stage to you iman uh welcome welcome yes hey thank you for having me thank you so much for having me um thanks to it google obviously for organizing such a great event and as well as to orzara for putting all of this together so as mikhail was saying the discussion today is going to be about ai ethics per se we will focus on image generation models and the question is what lessons could be could be drawn from ai ethics for the future of the economy which is as you know but afraid so i'm not going to discuss much about about retreat because i don't want to bore you and i know you know that rape free has this huge potential it does represent a parenting shift and we have we've heard the keywords surrounded around financial inclusion governance decentralization so we all know web 3 does represent this this new tool this potential to use our collective imagination to design new um digital spaces on on changers so how exactly do we build this future together how exactly do we ensure that the expansion of the web tray actually serves humanity at large my key point here is in the title and it is that aa ethics per se can be a tremendous asset to know what has been done before in the field that also is related to paradigm shifts that is also related to doing things differently and potentially serving humankind at large and that is part of um its ethical considerations its consequences its potential harms on individual and groups and that we can just use that in order to gain some time in order to not repeat the same old mistakes in order to get ahead to actually fulfill what we are i think all here for it which is the potential website so very briefly we say that in ai ethics um there are two type of harms and just so you know before i deep dive into it i will today with you hopefully focus on um imagination models um notably as mikhail just mentioned dolly because i think um a those models do represent a paradigm shift as well and they will be used in their website to create images to create landscapes to create um multi-modal experience user experiences as well so it's important to understand what they are and what their potential consequences could be and how exactly we go around them so ethics in general um those represent according to researcher k crawford two type of harms harm a harms of allocation her b harms of representation harms of allocation are harms um where people affected persons are not receiving are not being allocated resources so here you can think of not getting a loan um applying um to get a loan an institutional bank see here we go with financial inclusion um and being rejected based on um either a decision that has been made by an air system um or a decision that has been pushed for uh by any system and agreed upon by by a human being either way so those are harms of allocations you are not being allocated a given resource um and then we have harms of representation which i think here we will also deep dive into because we are we will discuss um image generation models so we are discussing images representations of society um structures of power as well so harms of representation is when a given person or a given group are represented in a given manner either they are presented with indignity or they are being stereotyped upon um etc so here you can think of well we will discuss harms of position when we discuss the limitations um of some a models notably a mass generation model so if you haven't heard about them if you haven't heard about image generation model um i'll give you some examples so the first well-known imagination mode that we know of is called gpt3 or gpt2 but we'll discuss dipping free it's a system that uses languages language um [Music] to instruct a large neural network to perform a variety of text generation tasks so the input is a language and the output here is also text generation for a doll e we have double e and dolly 2 which is the newest version of dell e it's a system that was created by openai [Music] which simply needs a simple text description to generate an image so here you can think of um [Music] a chair that looks like an avocado a chair that looks like an avocado um and then so you input this simple text description to the system and the output that you have is well an image of the chair that looks like an abacus a what the system does dolly too it turns those simple text descriptions into and this is very important into realistic photos that have never existed before a raccoon that dance said basketball so it turns text into realistic images that have never existed before um it can also realistically and edit photos so if someone wants to take a very important historical building that exists somewhere um and very realistically a edit the photo and pretend that there is smoke behind um this building it is possible so take take a second to think of the potential consequences of of what it could mean solely in terms of misinformation for instance or of representations of reality as well how people understand information how people understand realities so it can turn text into realistic photos it can realistically edit existing photos and it can also a start with an image and create variations um of that image based on the given image so think here of the case by clipped a the input here is the case by gustav clipped and the outputs are variations of the kids so not the kiss but variations of the kiss same colors a different positions but when you see it and if you're interested you can absolutely take a look at those examples that i'm giving those are real examples of photos that are public you can find them on the open ai website if you simply google um dole e2 open ai pictures and i wish i could have done a ppt um but i'm in vacations and and traveling so i don't have my computer so i'm just like this is totally me and my words um but if you if you want to take a look at the pictures you absolutely can and they're very easily found to be found online so just to to to sum up and then we'll dive into the limitations and what we can learn from them for the web3 um you have an image generation model that can turn language into realistic picture that have never existed before edit realistically existing pictures um and take pictures and and just like do variations out of that one given um picture in terms of art in terms of landscape creation like for instance for the for the metabolism it could have tremendous applications and we'll discuss the web free applications of image generation models but right here i do want you to take a second to think of um the limitations of those models because there could be errors in um the label data so that is a if within the system a picture that represents a car has been labeled as a plane and you ask the system for a plane it will give you a car so there can be errors in the label data um there are gaps and it's training um as well so it's very important to have in mind that the system is limited um and it's not solely limited in in terms of um technical limitations which is what i've just pointed towards it also has limitations because it has this potential to harm and i think it's very important to take this into account if we are to actually embrace disparity shift that webster could be the potential to harm individuals or affected groups is notably by reinforcing stereotypes um by erasing individuals by erasing groups by denigrating them or treating them with indignity i'll give you some examples there are plenty um but uh but i i do wish to discuss with you if we have like five minutes at the end of this uh discussion so i'll give you some examples of assumptions um that are reproduced um by the system um it does over represent um [Music] generations of people that are female passing for some a job occupation or a male passing also for other job occupations so if you do ask um the system to represent a builder you will have images of male blessing persons if you ask the system to represent um a flight attendant you will have images of female passing persons so there's this um [Music] there's also what i mentioned before in terms of erasure of certain groups or of certain individuals because the system does um [Music] over represents as well a person's with lighter skin tone or due to it's a potential to harm notably by generating a what we call legally explicit content a [Music] some images are not um cannot be reproduced so certain experiences for instance um [Music] if you're sick um if you've been in the hospital if you're a person with disabilities and some experiences have a less chance of being represented so those are just examples um of limitations this particular image generation system um when we take into account the added um capabilities of that system which i mentioned before um so which are a um variations you know the kiss of klimt and then you can do dozens of variations of the kiss um so when you take into account variations and when you also take check out what is called um in painting so a you know taking a picture and then adding something realistically to it or just editing it but very realistically um you also have all the paths of the pathways other possibilities and to harm to repeat use um bias stereotypes at scale so here the questions are always okay what is the model what does it do um how do we wish to deploy it and because open ai i thought as really far ahead how exactly they could deploy dolly and they did not release it for the world to use um how do we wish to deploy it and which harms are there which can be avoided which cannot which harms do we measure um and those that cannot be avoided how do we mitigate them how do we minimize the risk um so taking all of those limitations into account uh taking this potential harm into account i think it's um it's best to have in mind that by taking some tools that exist that i can create a lot as well obviously that's not the point um it's also it's also best to have their limitations in mind what it means um for the web3 is that you know we could think of using um image generation models for nft creations for um profile pictures avatars or even and that's what i mentioned earlier um for you know creating land creating um environments entire environments notably in the metaverse um so we could create landscapes based on landscapes images that exist or ask for landscapes that are that have never been um existence before to be created via such systems and so it has this potential to be used um to generate not only um images but also entire multi-modal experience um or worlds so based on the limitation that we have seen a or in general i would say um when wanting to align what's work with the paradigm shift that does represent 123 um plus wanting to a use this a this potential for creativity that also does represent image generation models um it's also important to have a those limitations in mind in order to do things differently and in order to actually align the promise of webfree with what we actually uh do um [Music] so one key example that i could that i could just like throw up in the air obviously um i do not have um solutions i only have questions that one can have in mind for a all stages of creating a and design new places whether it's at the stage of addiction or at the stage of team formation or whether it's at the stage of pitching and thinking the project um what does the project do to whom how do we deploy it etc etc so i leave you with this um examples that i really like of publication methods in ai effects research a few years ago um [Music] there was absolutely no requirements as to how exactly ethical considerations could be taken into account by researchers [Music] and i think there was a momentum four years ago and as of now the major ml the machine learning deep learning conferences research conferences do um ask their participants if they wish to be published to take into account not only the ethical considerations of the project but also um analyze their potential unintended consequences and suggest as well um risk mitigation strategies so a first step to do things differently could be whenever you use um the tools that we have um to create new ones to think ahead of um the potential harm um [Music] that the project per se um could embed so my point here uh and i would be happy to take any questions because there's a lot deep dive into at many stages um my point here is that it's great obviously to have it to have in mind the capabilities of the model um and those should be taken into account but it's best to also um take into account the larger context in which um these images can be used or shared and to do this you know in the present times and also in the future because it it needs to be a flexible approach and i told you at the very beginning that um ms generation models could be used to create worlds [Music] and as of now we don't we don't know for sure how exactly what what exactly the consequences will be um [Music] so based on this recognition you know this humility of we don't know exactly what all the consequences are going to be but based on the knowledge that some consequences are known some consequences have been attached to existing systems to existing technologies there are fields out there that are specialized into um you know centering lived experiences of affected communities to actually um imagine globally a future that works for everyone and so yes i will leave you with this i will leave you with um the possibility to to imagine further and based on a given understanding of how social hierarchies work based on an understanding of how the relationship between language and power works um and just with this you know with this plea that there's so many um people out there that would be thrilled to work with you so that you can also interrogate your work and enrich it and with ways to render it less harmful uh to whom in what ways you know and and just do it you know do it better voila uh that's what i can give you that's obviously a very brief introduction on um ms generation models on you know the type of harm that we have highlighted when it comes to ethics um and then you know i'm also very eager uh to learn about your projects and see what's what we can do you know uh in the web free that's a very bright future we could build together but if you have any questions i would be happy to to take them i'm not sure how you can ask them but uh but i would be here for it um thank you so much iman um i think i will you know like um we'll step in in terms of questions we have to wrap it up unfortunately very soon but nonetheless um i do think that this will spark a couple of follow-up conversations in particular with artists also i do think that we might might want to workshop this um in a different forum because i find the conversation and its implications for complex technological systems or as i was earlier when maybe you have not been tuning in yet been describing your web three is a techno-social space because it never can you know the technology can never live outside you know social context and certain things are just impossible to to achieve certain neutralities of technology there's a couple of conversations in web 3 about you know credible neutrality what does that look like is that even possible from a ethics point of view from a legal point of view so there is quite a few questions um that need to be addressed and don't get addressed enough in my opinion in work three in particular but i'm wondering you've been talking about guiding principles and since we do have quite a few builders here that you know try to build projects that they want to be equitable you know that want to be um driven by agency you know but very often you know certain experiences and knowledges or you know ideas are missing in the process you know you can't build anything taking like you can't build taking something into account that you don't know of what would be like your ai ethics quick guide to building with outcome to build with with ethical accountability like are there anything any pointers and even if it's a book that you would say that like is incredibly insightful when it comes to these things okay um i would say um at the stage of ideation and information um it's very it can be very insightful to look for other skills then so will we take related skills and look for um ethical philosophical experts um i would also say that it's very difficult to know what what one cannot know um so as well i would you know um [Music] interrogate the technology per se and try to see exactly um what the consequences could be and then to give an idea of um you know one resource that i find very useful to get started in jai ethics i feel like the podcast um called radical ai is a very good place to start um yes i think it's a good place to start radical a podcast um okay most certainly will will put that onto my pile of things i want to know and i mean there's also you know like we could get into very very in-depth conversations also about about ideas of of ethics you know there's of course like schools of philosophy what does like an ethic look like that is rooted in you know like cosmological principles of knowledge formation you could get very very deep unfortunately we do not have the time for this um but um imanabello please follow her on on on twitter and as i understand you just started your web 3 journey more or less current yes so i know we will see much more of you in our space this was just an intro to this weird and wonderful world that is our community around the global and zora and i can't wait to have another in-depth conversation with you um as well thank you enjoy enjoy your trip your holiday um have a wonderful evening thank you see you soon take care see you soon bye