Rise of AI Conference 2019 · Berlin

Facebook’s Perspectives on an Ethical and Privacy by Design Approach to AI

Norberto Andrade · Privacy & Public Policy Manager at Facebook

At Rise of AI 2019 in Berlin, Norberto Andrade, Privacy and Public Policy Manager at Facebook, presents three perspectives on how Facebook builds AI: design, product and research. He introduces Facebook's design team's "four quadrants" framework, which plots products from small-scale pixels to full ecosystems, and from individual users to global society, arguing ethical risk grows as products scale up in both dimensions. He describes AI as foundational to Facebook, powering newsfeed ranking and translating over 6 billion posts a day, and previews the Codec Avatars project, which uses machine learning to animate ultra-realistic avatars for remote interaction in AR/VR, while flagging open questions on diverse training data and avatar authentication. He explains that AI removes 99.6 percent of fake accounts before users report them, and outlines Facebook's fairness process across people, data and algorithms, including subgroup calibration and documented decision thresholds. He closes describing the Facebook AI Research lab's open-research mission and Facebook's support for an AI ethics institute at the Technical University of Munich.

Key takeaways

  • Facebook's "four quadrants" framework maps products by scale (pixels to ecosystem) and audience (individual to humanity) to locate ethical risk.
  • Facebook processes over 6 billion translated posts a day using AI-powered automatic translation.
  • AI finds and flags 99.6 percent of fake accounts before users report them, per Andrade.
  • The Codec Avatars project uses machine learning to animate realistic avatars for real-time remote interaction in AR/VR.
  • Facebook's fairness process spans people, data and algorithms, including subgroup calibration and documented decision thresholds.

Detailed notes

Three perspectives: design, product, research

Andrade structures the talk around three perspectives on ethics at Facebook: design, AI product and AI research. He frames Facebook's ethical responsibility as amplifying the good and minimizing the bad, which in product terms means considering what the company is able to build (technical constraints and resources), what it is allowed to build (company policy and law), and, most importantly, what it should or should not build, given the impact on people across the world.

The four quadrants of design responsibility

Andrade presents a framework from Facebook's design team plotting products on two axes: scale, from a single pixel to the full ecosystem, and audience, from one individual to all of humanity. He argues designers are most comfortable in the bottom-left quadrant, where the audience and use case are well understood, but that moving toward ecosystem-scale, society-wide products increases unintended consequences that only reveal themselves at scale, citing Facebook's use both for organizing movements like the Women's March and for election interference and spreading hoaxes.

AI as foundational to Facebook's products

Andrade says Facebook could not exist without AI, which powers features including newsfeed ranking, using signals like reactions, shares and comments, and automatic translation, which processes over 6 billion translated posts a day so users can communicate across languages. He also says AI is central to Facebook's mission of bringing people closer together and to keeping communities safe, citing that AI finds and flags 99.6 percent of fake accounts before users report them.

Realistic avatars for AR and VR

Andrade describes Facebook's vision for AR and VR letting people feel physically present with distant others, requiring lifelike avatars that convey facial expressions, voice tone and gesture. He presents the Codec Avatars project, which uses machine learning trained on sensor data from prototype headsets to animate a realistic representation of a person's face in real time, illustrated by a demo of two people, Nadia and Dan, conversing through avatars, and an early prototype of two remote users playing a shared virtual soccer game. He flags open ethical questions, including the need for diverse training data across body shapes, sizes and skin tones, and using face recognition or fingerprints to ensure only the real person can control their avatar.

Building fairness into people, data and algorithms

Andrade outlines Facebook's AI ethics approach as investing in three pillars: the people who build and review AI systems, the data used to train them, and the algorithms themselves. He argues fairness cannot be solved purely technically and describes a three-step process of surfacing hard fairness questions, such as whether to optimize for procedural consistency or equal outcomes across groups, resolving them for a given product, and documenting the decisions for accountability. He notes bias can emerge at each step, from human annotators' labeling practices to algorithmic over- or under-prediction to the decision thresholds used, for example, in classifiers that detect false news, and says Facebook performs subgroup calibration to apply consistent standards across groups.

Open research and academic partnerships

Andrade describes the Facebook AI Research (FAIR) lab, then five years old and spread across seven locations, whose mission is advancing AI through open research: publishing peer-reviewed results, releasing data, code and models, and forming research agreements with academic institutions. He gives two examples of research directions: teaching machines to understand physical properties of the world, such as predicting how a falling object will land, and training generative models for creative tasks like designing clothing. He closes by citing Facebook's support for a new AI ethics institute at the Technical University of Munich, saying some ethical questions go beyond Facebook and even beyond AI as a technical discipline.

Quotes

Facebook without AI could not even exist.

Norberto Andrade

Fairness is something that you cannot solve merely through an exclusive technical standpoint.

Norberto Andrade

Questions answered in this talk

How does Facebook use AI in its newsfeed?

According to Norberto Andrade at Rise of AI 2019, Facebook's newsfeed ranking is powered by machine learning, using signals such as how many people react, share and comment on a post to determine its placement.

What is Facebook's Codec Avatars project?

Norberto Andrade describes Codec Avatars as a project using machine learning trained on facial expression, voice tone and movement data to animate ultra-realistic avatars, letting two remote people interact together in a shared virtual reality space in real time.

How does Facebook approach fairness and bias in AI?

Norberto Andrade explains that Facebook treats fairness as a process across people, data and algorithms, involving surfacing hard fairness questions, deciding on a definition of fairness for a given product, documenting the decision, and performing subgroup calibration to apply consistent decision thresholds across groups.

Mentioned

four quadrants of design responsibility, Codec Avatars, AI fairness, subgroup calibration, open research, fake account detection

Keep exploring

More from Rise of AI Conference 2019

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 4,022 words

Show the full transcript

well first of all many thanks to the organizers for for inviting me it's a it's a great pleasure to be here and to be able to share with you some of Facebook's perspectives on on how to build AI an ethical and privacy by design so I'll be giving you kind of a starting with with a recap three different perspectives on how we approach ethics at Facebook I'll start with not directly on AI but telling you a little bit about the design perspective what it means from a design perspective to build products with ethics in mind then I'll go directly into what matters which is AI for for this conference and I'll deliver the two other perspectives one will be on an AI product perspective and I'll tell you a little bit about what AI means today for the company for Facebook and I'll spend a bit of time also explaining how we build AI especially taking into account fairness and bias how we're building in a fair and unbiased way the third and last perspective

will be on AI research and here I'll tell you a little bit about how we're advancing the state of arts on AI through our Facebook artificial intelligence Research Lab and how we are promoting and supporting also external independence academic research so design AI product AI research so let's start with so if this is moving so so let's start with with the design perspective at Facebook we have an ethical broad ethical responsibility to amplify the good and minimize the bad but what does this mean when it comes to actually building products well it means three things one we need to think and consider what we are able to build and this has to do with you know what technical constraints what state of art of the technology we have at hands what are the available resources that we allow us to build that product secondly we need to consider what we are allowed to build and here we need to be aligned with our companies policies and practices but also being consistent with local and national

international policy policies laws and regulations but thirds and equally if not more important is what we should or should not be building and this is where ethics comes into in into play so we need to take into account the impact that the things that we're helping to build will have on people the good and the bad and this is specially important since we are designing products for millions of people in all parts of the world and sometimes in ways that we have not anticipated so let me share with you a very helpful framework that visualizes the main dimensions and components of this equation and this has actually been put together by our design team and basically it provides four quadrants that they call the four quadrants of design responsibility and we start with the x-axis so we have from left to right we have pixels all the way to echo system so and this is about the scale of the things that we are building and in increasing size volume we start with small pixels

we go to the broader product and we go all the way into the ecosystem in which that product will be operating will be interacting and here it's it's important to take into account how that product will actually be interacting with the ecosystem that will be surrounding it and this gets complicated because the more you move from left to right the more interdependencies constraints and complexities you will meet and you will face so take for example a simple product like a cleaning product and it does particularly well the task of scrubbing your kitchen counter it does that really well but what happens if that cleaning product actually is released into the water system you might have unintended effects there systemic effects because it might actually cause damages you might harm plant and animal so this is when it gets to understanding how the product interacts with the environments what systemic and network effects might be triggered now let's move to the to the other axis the the y-axis and here what we have whoops sorry move

too fast here we have from the it's the same idea but the scale here that we're talking about are the audience that you're designing for so we start at the bottom with one individual human then we go up and we're actually designing for our user product base and then if we go and we go all the way until you're actually designing for the whole society and even humanity and here what is interesting is that the more you move from the individual to the humanity the less you will assume and you will know about the audience that you're designing for you will know less about their motivations about their needs about how that product is actually impacting how the product is actually impacting that audience so we have Facebook has been used for in a many different ways so you can have people that have actually organized social social change movements like the women's March but we also have people that have use and messaged Facebook to interfere with the elections or to spread Oakes's so

here we need to take into account how that letter from the individual and humanity will will play and the more you will go up the less you will know and the more you need again to take into account and Internet affects and miss reduce miss users so here here at the left bottom quadrant this is where designers feel particularly comfortable this is where they can really craft their art where they can push their pixels and create something that you feel very comfortable with because they know exactly what is the audience that they are intentionally designing for and they know what tasks that product will do and how important will be for the for the users but then when we move up then we move into the space of unintended and anticipated consequences and we only know about this intent and anticipated consequences when they reveal themselves at scale when you have a lot of people using it and the same thing applies to to the top quadrants in here but it's interesting to understand is

that when you're designing for the world for society there you are going to encounter different challenges because Facebook does not work in the same way everywhere in the world and people in different parts of the world use Facebook in different ways so we need to take into account also that that complexity of how the the social context the political context the social norms the economic conditions of the users plus public policy laws plus economics plus sustainability design will all work in that top quadrant and will influence how that product will be used and will resonate with our community so these top quadrants are the place of unintended effects of how our products are being used is the space where perhaps social bad actors or societal forces will collide with the good intentions of our designers so what we need to do is actually to design for the four quadrants so this is the space where the ethical responsibility design lies this is where we take into account the need to think expensively about the impact

that our technology that our innovation is causing on our community how is resonating there and this is where you need to think not only about what you're able and allow to build but what do you actually should be building so that was the first perspective let's move quickly to the second and here we're talking about a AI and AI as you can imagine is increasingly important to the company in fact Facebook without a I could not even exist AI is foundational to everything we do is behind improving existing products and it's about creating novel experiences to our users across our platforms and most importantly AI is also deeply connected to our mission into our values so bringing people the power to build community and bringing well closer together is all what AI is actually helping us to do so let me give you an example ai being at the core of our engineering efforts is actually behind some of our most used features one example is the newsfeed so the ranking of the newsfeed of

the content that you see is powered by machine learning so we take into account signals like how many people react how many people share how many people comment posts that determine how high up that post will end up in your newsfeed now let's talk about how AI can help bring people together and here I have an example of natural language processing of how we use that technology to help people communicate even if they don't share the same language and how we do that is that we use that technology to actually process our automatic translated posts so you have people that do not speak the same language that I can actually don't worry about how it's being translated because AI can do that for them and now nowadays we're doing overly 6 billion translated posts a day then there's another example so when we talk about bringing people together bringing the world closer together this is all about defying distance and there are a couple of ways to do that you know there's what I would

say like the traditional way which is about enabling people to communicate and to share content instantaneously I would grab my phone I would take a selfie I would send it to my mom and I'm sharing an experience timeline with her doing that that's the traditional way you know about sharing content but then there's there's another way there's there's a way about actually enabling people to meet and be in the same shared space even if they in reality are physically distant so what we have here is the future of AR and VR that is being that's which AI is underlining that will allow us to connect with anyone anywhere so just imagine that you'll be able to live anywhere in the world eliminated commuting you can be at the birthday party being truly present there even though you are all the way across the globe just imagine that so for this to build this we need a deeper sense of connection and Co presence we need to make people that they are feel to make them

feel that are physically present in the same space through someone else and being able to communicate their expressions and their ideas without effort effortlessly so and this has to do with facial expressions with voice tone with gestures so what we need basically it's lifelike avatars so developing realistic avatars is the next stage of augmented reality and virtual reality that will allow us to have those technologies being part of our daily lives and we have a project called Kodak avatars that actually allow by building real ultra realistic avatars you have two people sharing the same space and interacting in real time in virtual reality so let's let me give you an example I don't know if the sound is working but let's watch this playback of Nadia conversing to Dan through there they're ultra realistic avatars exercise I like AI so I do yoga and then I do boxing too so what we have here this was very short so we what we have is that the sensors in Nadia's prototype headsets are measuring her facial

expressions in AI translating them in real-time to Dan who can see a perfect picture representation of Nadia's face facial expression voice tone of voice and this is done in in in in real-time and it's about really translating not only who you are but really conveying that how what you look like what you sound like but obviously and the important part is that AI is underlining this process so the Kodak avatars project is it's a result of a process that that uses machine learning to collect learn and recreate social human expression is about training neural networks to map movements and expressions to voices and obviously we are achieving very good results in terms of realism of capturing the faces but we all know that in order to capture genuine human communication you need to not only look at the faces but the rest of the body so you need to require a full body language so let me give you an example how we're doing this and obviously this is all very early stages we are

still years away from this but this is where we're moving and this is what AI is allowing us to do as bringing people together so this is an early prototype of a simple live interaction so what we have here is that two people that could be very distant there could be remote one could be here in in Germany the other could be in Sydney and they're actually animating their avatars in the shared virtual space which in this case is the soccer field so what we're doing and if this is playing you didn't play let's try again so what so I want to show you this simple live interaction of how immersed people are hey I go back but I so you can see they're warming up and they're actually going to play soccer game and you can see how immersed they are now with the play with the ball going back and forth like it was actually there and like they were actually in that shared virtual space well well there could be really distant

so this is about really getting people immersed through that representation of avatars so the point here is was really to understand how AI can can allow this obviously we're still years away but we're thinking about important ethical questions about this and one is how to make sure that body models will work equally well two different body shapes sizes skin tones so we will need diverse data training sets to make this happen we also need to take into account that there might be important privacy and security questions so we need to understand like is that avatar authentic is is it you is it unique to you how can we prevent people from accessing your avatar so technologies like face recognition and fingerprints can actually make sure that the device you use to animate your avatar you only have access to it the other thing about a eye that is there is important and we're still in the AI product perspective is the is how important AI role in keeping our communities safe so here we are

through AI we are becoming more and more effective at removing bad content from our platform and keeping non violent ink platform in saying we've been reporting publicly this efforts across different categories of contents and what is important here to notice that through Yahoo through AI we actually able to find and flag content that we would then be actioning on before users reported to give me an example on fake accounts we're actually able to find 99.6% of fake accounts before users actually report it so only 0.4% of a council we act on we act on because users reported all the rest is actually being done through a through AI and this brings me to our AI ethics approach and here we are investing in three main pillars in the people that are building the algorithms in the data that we use to train the algorithms and any algorithms themselves because they ultimately represent what AI learns in terms of people investing in people that build the technology in people that review the technology so I'm part

of a review team back in in the headquarters where I review artificial intelligence and machine learning features so we also invest on the people are doing that review and then finally you also invest in people that are represented in the data sets ai is a product of the data that is trained on that data comes from our community from people so people do play a role at teaching AI and we need to understand who these people are because if AI only learns from a small set of people it will only have a narrow point of view so diversity here is fundamental the other important point is data we have important processes in place to make sure that the data that is selected and annotated is done in in the proper in the privacy friendly way and then ultimately we have the algorithms and here is that even if you annotate the data even if you select it flawlessly perfectly algorithm might still learn or propagate biases that were present in the data training set so

here let me talk about a little bit about fairness and fairness is something that you cannot solve merely through an exclusive technical standpoint it's not about having a group of engineers that will come up with a fairness library will that will upload algorithm bias module and through that will just extract all biases from their data from their models it's much more complicated than that because fairness is actually a process and it's about three following a best practice in three of them surfacing hard questions about fairness we need to understand who are the people are being impacted by automated decisions and how we we subdued them we also need to understand what perspective of fairness we should follow is it about ensuring consistency procedural consistency to everyone or is about striving equal outcomes across different groups I were talking about fairness in terms of treatment in terms of outcomes and then third second we we need to resolve those decisions that will resolve those questions and come up with with decisions understanding for a given product

taking into account its goal what is the best definition of fairness and third we need to document we need to record those decisions for the sake of transparency accountability and build case studies that will ensure consistency so we don't need to reinvent the wheel and next time we have another challenging proud that we could going back and understand how we've dealt with that product and here we need to do this at every step of implementation process it's about how we're defining ground truth about how we generate labels that train the machines how those machines are making predictions and what they say what decisions and interventions we make based on those predictions and in every single step you have potentially different biases that might emerge they might emerge from the labeling practices of our human annotators having their own prejudice or or biases and that's why it's so important for us in the data annotation phase to have a team that ensures consistency of labeling practices they actually draft guidelines that are approved and that make

sure that human annotators we will minimize their bias that go into this exercise we can have bias in the algorithm as I've explained before you can have over predictions or under predictions so we need to calibrate the scores there and we do a lot of subgroup calibration in fact and then we have in intervention bias and this is about deciding the decision thresholds of when to intervene when do we for example building a machine learning classifier to detect false news we will get a score from zero to one when is the decision threshold this is 0.7 0.8 and what does that means for the different groups that are uploading information to Facebook so you need a decision threshold that will be applicable to all groups in the same standard way so we hold content equally is through the same standard it respectively from where it comes from and in this process is super important obviously to leverage the external community because they will have the questions that we don't so another where anytime that third

perspective is the AI research and here I'll just talk a little bit about the Facebook intelligence Research Lab and which mission is really to advance the state of art of AI through open research they are five years old they have over six 160 people spread in different locations seven of them but what does it mean actually to do open research it means a couple of things it means first of all to publish their results peer-reviewed journals and conferences and making it publicly accessible to everyone it means contributing to the open source community by publishing data codes and models and he means also to entering to research agreements with reputable academic institutions but what is next for AI and n and here I'll just give you what are we interested in and what problems we want to solve and I'll just leave you with two examples one is actually to try to make a machine smarter but not in the narrow sense of just having a machine being really good at doing a specialized task this

is actually want machines to understand the world where we live in we want to make machines understand the world that surrounds them so if you take into account a baby that just came to this world he will not have a good idea of the physics now if I if I drop this clicker it will fall down vertically you will not simply go up or go horizontal or if I if I let it fall it might stay still the same solid or I might break into different pieces this seems very intuitive for us and makes a lot of sense but machines do not know that so actually we have a project in place to make machines understand the physics of the world so here what we have is the this cubes understanding where they would fall the the one at the bottom the machine actually is a very naive model does not know where that where the cube will fall so he thinks will be a bit everywhere while the other is actually getting it right

getting the location where it fall and understanding the location of that cube will also allows us to predict the future more more accordingly then the other one is the creativity so we actually training machines to generate photos of clothing and as the model learns it actually stabilizes and generates more and more creative pieces of clothing and you can see this in the future being used by artists to actually leverage this technology to create new pieces of Arts in new domains like painting fashion composing music or creating animation videos last I will finish with sorry I'm going down I'm going back apologies for that so the last slide I'm promised I'm done so apart from all the in-house research that we do we also very invested in promoting and supporting independent academic research this is what we've done recently when we announced supporting the creation at the Technical University of Munich of an Institute for Ethics on on AI and why we do this we do this because we believe that we do not have all

the answers and neither we should have them and because some of the questions go way beyond Facebook and they go beyond even AI as a technical discipline and academia is best equipped through into this interdisciplinary research to seek those answers so with that and I thank you for your attention and I apologize for going a little bit over time

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.