 Welcome to the podcast of the robustly beneficial AI group in Lausanne, Switzerland. This week, we'll discuss emotional contagion through a paper that was published in 2013 by researchers from Facebook and Kornet. Yeah, so that was a very important paper, I think, that gets a lot of attention, but also a lot of controversy behind this paper. So you want to present it? So the title of the paper is Experimental Evidence of Massive Scale Emotional Contagion through Social Networks. And so what they did is they decreased or increased the, no, not increased, they decreased either the quantity of positive messages that people see in their news field on Facebook or of negative messages. So it was quite, it's quite long ago. So to detect positive and negative messages, they use the specific keywords and messages with positive keywords where classified as positive and with negative keywords as negative. And then they see that when inside your feed, you see less positive messages, it will make you post also less positive post. And if you see less negative messages, you will also post less negative messages. You will make less negative post. So that's what they call emotional contagion, like we are contaminated by the emotion that we experience. This was already some phenomena that was quite well known in real life scenarios that was very well studied. But people were doubting that through social network where we just interact with lines of text, we would also have these kind of effects. And with this experiment, they showed it. Yeah. So it's interesting that the paper was framed as a psychology paper. They were discussing this psychological hypothesis and how this paper is a contribution to this body of work. And it is, and it's interesting. To me, that's not the most interesting part of this paper. Like if I were to rewrite it, not necessarily to be published, but in a way that's most compelling, like the main takeaway, like I would phrase it more like in terms of like how much can you do by tweaking recommender algorithms? Maybe before we go to this kind of aspects, maybe just some basics for those who didn't read the paper. It's worth noting that this sample size in the number of people who were involved or rather used in this experiment is 680,000. Yeah, okay, three quarters a million or at least more than half a million. So it's a really massive number. So it is a bit far from the orders of magnitude that people were used to in massive psychological experiments. And back then, so I remember very well when the paper was published in 2014. So the work was done in 2013. Something also to keep in mind is that the algorithmic tools that were used in the work are now seven years old. So probably very, very basic. Facebook didn't have the machine learning AI capabilities it has today. There was no Facebook AI research team. There was barely any academic machine learning research happening then in Facebook. Yes. There was some data scientists like the main author of the paper but we have to put this in context that this was done with very basic tools back then but still the experiment was impressive for some worrying with its scale and what it showed back then. Yeah, I suggest we discuss more the results and interpretations of the results and then we can move on to the controversies because they are extremely interesting. I was not going to the controversy first. Just like the basics, the number of people in the experiment, the key findings and the tools that were used back then. Yeah. And so if you look at the results, so as you said, they observed like modifications of what people posted after the treatment. So the treatment being you see less positive contents or you see less negative contents. But so it's like statistically significant in terms of the p-value but the effect size is quite small that's something that's quite not worthy. Yeah, it was between 1% and 0.1%. Yeah. But given the experimental setup like maybe I'm framing this a bit as a patient question but you probably shouldn't, you like it was weird to expect a lot of huge impact because all they did was just to remove 10% of one kind of content. So like 10% of all the positive or the post with positive words or 10% of the post with negative words. Yeah. So this is like is really like a very small tweak. So you should not expect a huge, huge effect sizes but you do observe the effect sizes which is already quite impressive. And the other thing is that the experiment lasted one week. So it's like the effect of this treatment after one week. And actually next week we're going to discuss things that are about like modifications of human behaviors after several weeks. So one week is actually quite small, like in terms of human covenition and changes of behaviors. And so in the end, like even though the effect sizes were quite small, you have to take also into account that the experimental setup was also quite limited and it's already quite impressive that you see such an effect. Wow. Yes. And also this small effect size given the size of the social network if there is one billion users, 0.1% change means that it changes how one billion user would behave which is extremely significant. Yeah. Like another thing that was discussed in the paper is the fact that it was not only mimicry. You could imagine like just people were replicating the emotions that were given. But the, there's one subtlety is that like if they only did that, then you would expect that people who see for instance, less negative posts would put posts only less negative, would also post less negative posts, but you may not expect them to post also more positive content. And so that was discussed like it's also like people are actually being more positive by being exposed to less negative content which was like insight into the fact that it's not just mimicry. Yeah. Kind of was very interesting. Yeah. So these people like I often use it to show that algorithms have an impact, they have a huge impact. Sometimes it's even debated like people sometimes argue in the end it's the humans that we share the content and that's true humans play a big part. But these kind of people I think really shows that especially in social networks where you have curation by some algorithm, the algorithm plays a critical role on what people feel. And maybe that's an extra operation from this paper but probably it also has a huge impact on what people believe and what people are actually doing that they do practices. Yeah. I was thinking that I would be really interested to know when let's say we would do the same experiment on YouTube where for some we would remove some of the content that is quality information and for other users we would remove the content that is false information and low quality information and then later on ask them about their belief. So I don't know what this study would give but I expect that it would be even more than an even larger effect than for emotions. Yeah. And just to take examples from the books that Mindy and I wrote like you can imagine that if the YouTube algorithm stops recommending anti-vaccination videos you could have a huge impact of this small tweak in the algorithm on global health and issues related to vaccinations like for instance the World Health Organization reported that around 100,000 people died because of measles which is a disease for which there is a vaccine. And so you can have huge impacts by tweaking these algorithms but you have to do it in a robustly beneficial way which is not always easy but there's something to be done here. Yeah, another interesting point that is good to look at in the paper was that so they had a control group where they remove at random messages not specifically removing a message with the emotions and they found out that the groups for which messages with more emotions were removed ended up interacting less with the platform. So they saw a difference and even in one case there was a difference of up to 3% difference of the interaction with the platform and this goes back to other things we have discussed about ASFT such as polarization the fact that people interact more with comments that are more charged with emotion such as anger but also very strong happiness that makes people engage more with the platform and we did use from that that if one algorithms tried to find statistical pattern in how to do recommendation he would find that recommending this kind of messages this kind of post is worth it and is doing it so if we let the algorithms get smarter and continue to recommend based on any metric engagement then this is what we should expect to happen. Yeah, and this is interesting and important in terms of like if you want to impose ethics on algorithms then there may be a cost involved for the company that's probably going to be a cost involved for the company and it's much easier to encourage companies to implement some ethical values in the algorithm if you can show that or strongly suggest that these ethical values will not be too harmful for instance the use of the platform so typically this result, particular result would suggest that if you want to encourage companies to put ethical values in the algorithm probably suggesting them to have purely like no emotion content would not be a very good idea because it's very bad for the company whereas if you try to argue that by promoting more happy content engaging but with positive emotions then maybe the companies will see that it's not that harmful to it and can actually be beneficial for the company in terms of like the reputation of the company or in terms of users of precision of the platform and overall I think this is a key element to take into account when you're designing when you're proposing ideas to put ethics into the algorithm but I'm not very confident that this would be efficient I think even suggesting to promote more happy content I think we can expect it to hurt the business a lot because things like fear or anger has been shown to propagate much more and get much more engagement like just recently we see that there are a whole contagion of fake news about the coronavirus that are going on and they play they just play on the fear of people online that because they're afraid we can kick and release articles aren't there topics on which we could reach easy agreements that say topic A should be propagated a bit more topic B shouldn't be propagated as much I think this we could easily agree on that but it would also be a result that it would hurt the business there would be less people watching YouTube or less people scrolling on Facebook so you have business side effects but you don't even have side effects for the intentional purpose you had initially you might want to decrease the propagation of topic B and then end up having worse than B appearing yeah there's another paper maybe we can discuss for another session that studied a bit this like what are a few other papers like what are the emotions that create the most amount of engagement and yeah fear and anger on top and this is but actually happiness is quite good as well in this list so I think it's interesting to investigate all these things and maybe it's also like it's like clearly also very contextual it depends a lot on the particular person it depends also probably on the overall mood of the platform like especially like the coronavirus like you have a piling up of this fear and so people are already within this emotion and I think emotions are very addictive in a sense like when you're angry you want to stay angry in a sense like you're shouting at others and they're not angry upset you if they get angry also like makes you even more angry yeah I think it's also an interesting research in terms of psychology to know yeah how much people value these different things and how to deal with this kind of addiction to certain emotions and getting people to reflect on whether this is really beneficial to be in this state of emotion I'm not saying it isn't useful to be angry maybe sometimes it is useful to be angry to provoke a change or because but yeah a better understanding of the effects of these different emotions is very relevant actually to making algorithms more robustly beneficial another interesting point at the time the paper came out it was very political and one of the reason was people saying that this is not a ethical experiment yeah so there was one of the big controversies I think it's both discussing it I think because the size of the experiment is extremely unusual as you said for psychologists you bring psychology don't have that many that many subjects and so you like just doing the experiment feels like you already have a global image of so many people and in a sense it can be harmful to the subjects especially and there's this other problem which is like did people really give their consent to participate in this experiment so technically it was written at some point in the paper that it's part of what you sign when you sign the the agreements on Facebook yes but yeah there's this debate like is it actual consent to sign this because well who reads this kind of agreement and the other thing is about the opting out like usually one of ethical one ethical rules rule of psychological experiments is that people can opt out whenever they want and here it was not even clear that if they could opt out they were not even aware of this so yeah these were a few discussion points a lot of the yeah I think people don't like this this kind of experiment because somehow when deciding to show more negative content to users we have we see that some this kind of user also posted slightly more negative status updates like to maybe 10 000 users changed their status somehow did we hurt these 10 000 people and in the other hand on the side when we were showing more less negative comment and more positive did we create a large amount of good to this more than 100 000 users that were slightly more happy in their life because they were able to see this kind of message and then it raises so some people do the mistake of thinking that the right ethical thing to do is to not do experiment and change nothing keep the algorithm the way it is but then I don't I don't I don't buy this and I find that if we if we are if we think that one solution is much better than another to make people feel happy then I would really want that the algorithm goes for this kind of solution yeah maybe just to give you a presentation of what you just said like the experiment did not involve showing more negative content to people was only the removal of some kind of yes yes sorry it's only a small fraction so yeah so there was probably a lot of thoughts that went into this when they prepared the experiment but yeah it's definitely something both wondering about like because they did choose to remove 10% of positive content in the news feed for some users and this actually created those users to write more negative stuff so putting them in the probably worse mental health condition during the experiment so is that justified I think this is a difficult question but you have to compare it with the gain you get by learning this information about this experiment I think this experiment gave out a lot of insights into what can happen at least on social media does this insight is this insight more valuable than harming and like when I say harming I should insist on don't harm them a lot like it wasn't like torture as we said the effect size the effect sizes were very small now the question is like were they going to be much bigger would they have stopped the experiment and that would be an interesting question but it seems like the harm that you caused here is very very small but the insight you get is quite quite big and interesting I would argue yeah and also this is the harm that is done by this kind of experiment when you are in the wrong condition of the experiment it's very similar to the harm that is done by having the social network used by 1 billion people every single day and it's not it doesn't make a big difference but as you say the insight we learned from this experiment that the slope is positive if we show less negative things then the user will by themselves post more positive emotions that's very interesting to know and yeah the key point of this that's not mentioned in the paper but what we hear from it is that yes this social network have a huge impact on our emotion and can really change the way we behave yeah yeah and then there's this I guess all the question is like should you now what should you do now now that you have this information like should you actually go forth and change the algorithm for the better and what which I guess it's still no other question like from is going is Facebook actually going to do this because we like if this is what we want them to be doing we also need to make sure that they will be doing this which is also another problem like I feel like given the power of this algorithm we need to make sure that they are going to be robustly beneficial and maybe an observation that we need to start with is that they are not robustly beneficial at least not all the time right now and the status quo setting here is improvable I'd say at least and well I guess that's what we why we're doing this podcast is that we think that these algorithms should be improved and should be made more robustly beneficial but it's also very hard to make them more robustly beneficial and you need more insights to understand what are to be done what was what is the priority for instance in terms of ethics to be done for for algorithms to be more robustly beneficial and also I think the the idea of prior rightization which is an important one in effective algorithm like trying to find what are the what are the ethical problems with algorithms that are most important to tackle right now is extremely important then maybe there's not enough thought about this prior rightization and this probably also has to do with we don't have sufficient insights into what the impacts of the algorithms are and so this kind of paper is actually I'd say very critical to show that yeah yeah there are problems for instance with data privacy and these are really important as well but I feel like these kind of impacts into our large scales are extremely strong and maybe neglected compared to other concerns that that may have for algorithms so you would be happy to see more experiment of the type run on other big platforms yeah so that's I guess the big question because what the controversy at that time led to was essentially companies what companies stopped doing this kind of or rather they stopped making public this kind of research because like pretty sure I would bet easily that they're still doing things that are much worse than this experiment in practice they probably are doing a lot of A-B testing and seeing people's reactions and probably they're testing more for user engagement than being visual impacts so clearly they have a lot of insights into these impacts of the algorithms that from the outside this was very hard to have so we definitely encourage them to make them more public at least what they are already doing and it's a bit of a shame I'd say because I feel like because of the controversy was such information was no longer made public and we have less insight than we could have into what these algorithms are doing and what is the impact of these algorithms hey it's a it sounds a bit similar to what we discussed two weeks ago about algorithmic accountability and understanding black boxes so here if such company do experiments to to measure metrics like user engagement or whether user users are happy or not yeah it's it's sad that today they do it in a not transparent way that we don't know what they are trying to do and no one here is able to sufficiently study the recommendations of YouTube because it's black box at Recording Choir so yeah it would be really amazing if they make at least part of this transparent and accessible to us by publication or just opening opening more APIs to to be able to study the recommender systems yeah yeah and it has to do whether they're going to do it or not has to do with the incentives that we give them and I feel like unfortunately the backlash that occurred because of the after the publication of this paper like a lot of people complain about the ethical grounds of the of the experiment and I agree the experimental grounds are are disputable but by it's like don't know where I heard this but I think was do I get or something that says that whenever your mother calls you and blames you for whenever you call your mother and your mother blames you for not calling her enough she's actually doing something very counterproductive for how it goes yeah I feel like it's the same thing like Facebook released something publicly like they got blamed for this while they're not going to release anything publicly so of course we have these strong beliefs and we want to express them but I think it's important as well to think about the implications of making these concerns to be able to to violent to yeah to feel good yeah but yeah I do feel like there should be more studies about this because maybe very controversial because I think these social networks are getting more and more complex and in fact of the of the algorithms are more and more complex because there are more and more personal personalized the architect with more and more data and they have more and more clever sophisticated algorithms that are used to analyze and to do recommendations and it's back it's a very black box so far at least at least medium black box to use the terminology we introduced two weeks ago from out for outsiders but I feel it's also a black box for like very dark box for insiders as well like I'm not sure YouTube how much YouTube understand what the YouTube algorithm is doing because the YouTube environment is very very complex and if you don't give yourself the resources and to do these kinds of experiments then it's like having your head buried in the sand and you just move along and you just blinded to all the possible side effects that your algorithms have and I don't find it obviously beneficial yeah and then also people continue to think that this become a system of a small impact on on our social life but they actually have a strong one yeah maybe another thing we can discuss is about the the fact that this experiment so had impacts negative impacts on some of the the subjects of the experiments which is something that's undesirable but that unfortunately often happens in psychology experiments in general and the approach that was used in this paper was a very classical or statistical test approach meaning that you first choose a sample size here and was very large and then run the experiments on all of the subjects and you have to go all the way through the experiment and then analyze the resource this is like the very classical setting of classical hypothesis statistical testing and and well and you can see this as well it's dangerous like if the treatment was much more harmful than it actually was it would have been the experiment itself would have been causing a lot of harms and that's really not beneficial so what instead a lot of people have been proposing was to have more on different kind of testing approaches at one sense based on multi armed to bend it so the problem of multi armed to bend it applied for instance to medical drugs like the efficacy of medical drugs in this setting what you're doing is you have subjects coming in and for each subjects you still choose randomly either to give him the drug or placebo and and then you will go to the next one and so on but the trick of of many multi armed and it protocols is that you're going to change the probability of giving the treatment to a new subject depending on the performance of the the drug on the previously tested subjects so that if the drug is starting to kill people which is not good you should decrease very quickly the probability of giving the drug and you can actually prove guarantees so it's a very classical setting to do this and more generally this is the problem of safe exploration so the problem of a safe exploration is that we want we have this very complex system say social videos and you don't fully understand the effect of different treatments on this system or different actions in general and so you need to understand it better so that's the exploration phase but exploration can have a cost and so if you find one solution it's also it seems so tempting to do what's so called the exploitation meaning that they just do what you think is best not what you think will let you acquire more information and this is like very discussed and it's a very interesting problem and so one of the most fundamental problem I think in decision making in general and apply here to ethical problems so I guess what I would like to see if there are such studies what as such studies are done by companies and hopefully hopefully according to me published in the future is approaches that are more along these lines I think in the case of Facebook when they did the experiment they just cared about gaining the information and that's why they they are this perfectly balanced sample of as many participants on the negative condition as many participants in the positive condition as many participants in the control condition so but if they on top of caring about gaining as much information possible if they also found it very important that there are less negative messages that's what they so that means that throughout the experiment as they as they learn more and more that the negative condition leads to more negative messages they would change the way they handle participants and and change them with the change that condition from them yeah yeah and this is also valuable from the business point of view if you replace ethical part of the values by profits for user engagement as opposed to maximizing if you just want to know more you just want this big experiment well this we're just running the experiment is going to have a cost on profits or on user engagement or however and so you don't want when you want to make sure this is not going to be too costly but you cannot do it ahead of time so you can learn this as you go and if you quickly discover that one action is actually very harmful to your business so let's hope more for ethics then we should stop doing this yeah and if there is a if you measure a large effect you will need very few examples of this large effect to be able to make a right decision but when in case where they effective very small like maybe in the case of the Facebook Facebook study then it's a case where you you would still need a lot of participants to detect it yeah yeah I think one difficult thing about doing these ethical reasonings is that the effect size really matter like you have to put numbers into these things and to say well this is not too harmful but inside food so we should do this and this like is very very hard to do I can totally understand that people feel like this is something we should not do but I do feel like in the end this is something if you want to make algorithms robustly beneficial especially in the long run we need to have this kind of thinking between weighing different things and doing so quantitatively is is I think an important way to go go I understand that this may be controversial maybe you wanted to bring the the question of public health so that this experiment teaches us is that though people are sometimes in denial of the impacts of platforms for people's not like not now less and less in denial but we tended to to hear a lot of denial in the real impact of platforms and what they recommend to users what this experience shows us is that you can have a significant impact with relatively small experience given the size of Facebook now because of the backlash these kind of experiments are not reported anymore or at least like Facebook apologized and stopped publicizing this kind of experiment and like in the opposite for the purpose of public health and public mental health we actually need to have some oversight in the kind of routine A-B testing that is ongoing because who knows what kind of impact you are you are going to have just what seems to be innocent A-B testing just testing this versus like A versus B and then looking at what the user engagement be recently there have been calls by Sundar Pichai the CEO of Alphabet so Google for actually more regulations so now we even have people from these platforms telling the public authorities please regularly well this is a bit stretched but we are almost there like please regulate us and I think it's like it's time to think about so about what could what could public health authorities bring and actually they should they should be involved in the routine testing that is ongoing because we don't know what the impact would be on people's health and the larger effects so maybe also like that's one of the important consequences of the of this experiment and the way it was done and maybe the way it was rushed to be to be done or published or both Yeah I really agree with this like yeah I talked to to people at some point from the World Health Organization and in particular people at World Health Organization working on AI and applications of AI to to public health and this was neglected and when people were not talking about I had I had the same experiment exactly what we do with the World Health Organization and they tend to like some of them still think in terms of television and newspapers and they don't realize that I think if you want a vaccine like a flu vaccination campaign to succeed besides going to campuses and running ads in hospitals you actually have to try to have YouTube and Facebook on your side and nudge people through their platform to just take the flu vaccine early in the year would have much larger and maybe they're doing it I hope but some of the interactions I had with with this well the few organization like the few interactions I had I could have from with the World Health Organization was the same feeling that there's too much focus on on all media and a lot of overlooking like platforms that still not taken seriously like for them like YouTube is still for fun and entertainment and they still don't realize or maybe some of them still don't realize how much gains they could make for public health if they focus a lot of attention there yeah I said there are two different aspects of it like one of them is that when people hear about AI they really think they like a lot of people from my understanding is they think mostly of like diagnosis using some machine learning algorithms for like majorly or whatever or robots for surgery or robots for surgery but they don't think of YouTube recommender system and yeah and they don't think as as this as something that has to do with health and yet like vaccination is a huge part and information is key to good real vaccination and information goes through these social media as a lot these days and the other thing is that like the more I think about it the more I feel like the biggest changes in health will be more and more about mental health or at least it's it's a big part of it and and yeah in terms of mental health like information is critical and what information people are exposed to on a daily basis is critical and again like social media has a key work yeah let me know if I'm wrong but I thought there was studies about detecting a depression from a mouse movements or scrolling patterns of people so if this platform can with the higher accuracy detect a depression and help the people that are in it but still this is not either this is a field where it's a bit it's a bit complex because you might you might have yeah you might from starting from a good intention have side effects that are even worse than you're okay there's a need to be sure I guess so but but there are like other sub fields of public health where we could already have an easy discussion like the seasonal flu okay like the seasonal flu could like many improvements could be done like thousands of lives go each year because of the seasonal flu and hundreds of hundreds of problems yeah sorry thousands like regional talking like regional national scale and and the seasonal flu vaccine is something people still don't think of as a necessity and or if just a larger fraction of us took this this flu the spread would be much more controlled and much much much more lives would be saved each year and so the like fields were like you can already have an easy discussion and then of course you can go beyond and and to more more complex public health concerns such as mental health yeah yeah I think for mental health I think these are very encouraging paper but I think it's harder so we also need a lot more research to be done to understand what treatment what kind of videos you propose that is going to be robustly beneficial for for the half of the mental health of the user this is very difficult and unfortunately we lack data about this and if we want to get more data I'm going to have to run experiments on people because unfortunately we don't have the the means to simulate people so far this is where the oversight the oversight of the oversight of public health agency is necessary because nobody wants a private company to to run and then and even the even the company itself Facebook itself wouldn't would refuse this this responsibility yeah and then it's like it's a necessity to have some oversight from public health authorities if you want to over to undertake these kind of experiments or just like oversight for what's already happening just the testing that is routine testing as I said being done on the algorithm should have some oversight yeah so I'm going to end with a call to medical doctors or people like there are a few opportunities and things to investigate especially if you want to do research about this and I think there's a lack of it right now so yeah that's one area of impact to design more robust artificial algorithms thanks for watching or listening I hope this was productive and it would make you think even if you maybe disagree with what we've said and you know we're going to welcome also a criticism about what we said I think it's a difficult topic next week we're going to discuss something related to this we're going to talk more about the long-term effects of exposure to different kind of information and especially in the context of algorithms and the internet and until then I hope you're going to be here next time and we'll see you