Nicholas Christakis: [00:00:05] All right, so human beings are embedded in social networks, and these networks obey very particular mathematical, biological, and social principles. And increasingly, we are adding artificial intelligence in the form of online agents and physical robots among us who interact with us as if they were social entities. And these sorts of agents that we're going to be adding to our systems range from driverless cars on the roads to checkout machines in stores, to humanoid robots in homes or in factories or in battlefields or in firefighting situations, to disembodied autonomous agents such as online bots and digital AI assistants in our phones or our eyeglasses or our workplaces. And these technologies interact with us on a level playing field as if they were human. And they will give rise to hybrid systems of humans and machines. And these systems offer opportunities for a new kind of social artificial intelligence.
Now, let me give you just a toy example of this. When you get, like, a digital assistant, like an Alexa, the manufacturer of that device is very concerned with the human-machine interaction. And that human-machine interaction is optimized. For example, you would never buy the Alexa if every time you needed something from it, you had to say, excuse me, Alexa, I'm very sorry to interrupt you. If you don't mind, would you please tell me the word? [00:01:35] weather tomorrow, right? This would be considered an absurd level of politeness. You expect to be able to say, Alexa, weather, and then the machine obediently responds. And that's fine until you bring that machine into your house and your children talk to that machine and learn to be rude. And then they go to the playground and they are rude to other children. So that machine that has been added into our midst, it's not just about the human-machine interaction, it's about the human-human interaction in the presence of machines. And so what I'm interested in is not, is that, is the human-human interactions in the presence of machines. And we can use an understanding of social network structure and function to assess the uses and impact of social AI within and upon human groups. with respect to factors such as trust and cooperation that are necessary for groups of people to work together, uh, and that affect the behavior of these collectivities.
So let me highlight some of the work that we're doing, uh, using several approaches to network experiments involving, uh, artificial intelligence. These experiments evaluate how AI might affect the structure and function of human social interactions. [00:02:45] Now there's a class of collective action problems of social interactions that we have that are known as coordination problems. And that is, these are problems in which we have to work together if we were to create something of use. And sometimes we solve these problems by creating centralized institutions like police or courts or governments. But often we are able to solve collective action problems that require us to coordinate or cooperate with large numbers of people in a decentralized way. We actually evolved to have this capacity. And one specific type of collective action problem that I'd like to start with today is, in fact, coordination.
So, for example, to avoid this traffic jam, people have to coordinate to do something dissimilar from their neighbors. So if everyone leaves their house at the same time, everyone is stuck in a traffic jam. But if, if they jostle their departure times and they leave over intervals, then nobody is in the traffic jam. Now, of course, you could have some kind of central authority that coordinated this. You guys leave first, then you, and then you. [00:03:45] But ideally, what you would want is some kind of decentralized, non-top-down way of human beings coordinating to solve this problem.
So here was our first experiment that was published in 2017 on such hybrid systems that explored how AI might help with such a challenge. And we explored the performance of human groups that were engaged in a coordination task. And our paradigm was to borrow from computer science something known as the graph coloring problem. This is a classic problem in computer science. But what we did is we took it and we put human beings in that situation. So what we did is we took 4,000 people and we put them in 230 online groups. So these human beings were dropped into these groups. And they were randomly assigned to a spot within an artificial network that we created, the structure of which roughly resembled real human networks. And they were dropped into these spots and they were told they were assigned one of three colors, [00:04:46] a purple, orange, purple, orange, and pink. And they were told that they had to pick a color dissimilar from their neighbors.
And they were given five minutes to do that. So these people would look around at their neighbors. Each person would look around and they would say, like, this guy here is pink. He sees he has a pink neighbor and an orange neighbor. What he should do is switch to the purple color. And the red lines here indicate color conflicts. In other words, if the two colors of adjoining nodes are the same, they get a red line. And the background purple lines indicate the structure of the social interactions. So you put the people in here. They look around at their neighbors every second or second and a half. They make a switch consistent with the objective that all of them have to pick a color dissimilar from their neighbors over the next five minutes. And if they do, then and only then will they be paid. I'm going to pay you guys to work together. And if you all work together in a decentralized way to solve the problem, you all get paid. Otherwise, you get nothing, okay? So here's what happens in this experiment. Here on the x-axis is time in seconds up to the five-minute mark. [00:05:47] The game lasts five minutes down over here somewhere. And here on the y-axis is the objective function or the number of color conflicts. And so here at the beginning, there are 12 color conflicts. They're randomly assigned their initial colors. And that's shown here on this little histogram right there. And so the people start looking around and switching their colors. You know, they switch and they switch. And they get to this point here where they've now reached the situation in which there's a color conflict between these two people. And this conflict, however, is what we call an unresolvable conflict. So the light orange lines are resolvable conflicts. That's this guy here. He can make a move to purple that resolves the conflict. But these guys here that are orange, there is no move they can make that reduces the number of color conflicts in their neighborhood, right? This guy can't switch to purple because, actually, if he switches to purple, he'll have more conflicts. He's got, like, four purple neighbors. And he can't switch to pink because he's got two pink neighbors. So he looks and he says, well, the least conflict I have is to just stay orange. So now this group is stuck, right? [00:06:49] They have an unresolvable conflict and nothing can happen. No progress can be made in the solution of the collective action problem until one of these two people makes a counterintuitive move, switches colors to purple or pink, and increases conflicts temporarily. And that's, in fact, what happens. And then time goes by, and the human beings at 245 seconds solve the problem. The machine detects the solution, stops the game, and pays them, okay?
Now, because we were sneaky, what we did is we did some experiments where we surreptitiously replaced some of the human beings with bots. And we evaluated how the addition of AI-endowed bots to create a hybrid system affected the group performance. Is it possible to add some bots to human groups and improve their ability to coordinating when facing such a challenge? And what we did is we added three bots, and we experimentally varied two axes. Where the bots were placed, where they randomly dropped into the network, where they put into [00:07:51] the center of the network, or where they put into the periphery of the network. And we randomly manipulated their sort of AI capacity here in a very trivial and simple way. Namely, we manipulated whether the bots acted with perfection or acted with a little noise.
In the perfection situation, every second and a half, the bots looked around at their neighbors, and then they picked the color that had the fewest conflicts with their neighbors. What you might think of as irrational behavior. In the 10% noise situation, they did that, but 10% of the time, they picked a random color. And in the 30% noise situation, they did that, but 30% of the time, they picked a random color. So we made the bots, let's say, more and more error-prone, more and more noisy.
And then we looked at, we had, I think, something like, let's start with a control group. We plotted here on the x-axis is time, and these are survival curves. On the y-axis is the probability that the group as a whole has not solved the coordination game. So here, if you look here, for example, at the beginning, at time zero, [00:08:51] 100% of the only human groups, the sessions with only humans, are in orange. At the beginning, 100% of the human-only groups have not solved the game. And then as time goes by, more and more of the human groups solve the game, so that by the end of the five minutes, maybe 60% of the human-only groups have solved the game.
Well, what happens is, is if you put 10% noisy bots into the central position of the network, you get discernibly improved performance. Here, substantially more groups of the people, when the bots that had a little bit of noise were added to the middle, were able to solve the problem. In fact, they reduced the median time of solution from 232 seconds to 103 seconds. And there are other findings also in these data. Perfect bots and overly noisy bots were both unhelpful. You needed some calibration. It was the 10% noisy bots that were the most helpful. and also the position of the bots had some impact as well. But crucially in these experiments, we also found that human beings who were not connected to the bots, [00:09:52] who were further away in the network. So in the graphs, there were some people who were connected to the bots and then some others who were not. We found that even those people started changing the way they played. So there was a ripple effect, a cascade effect. The benefits of how the bot was interacting with its humans rippled through the network and then began to affect human-human interactions further and further away in the network. In other words, the bots helped the humans to help themselves, and the benefits of the noise dispersed within this social system.
Now let me further fix ideas regarding this collective challenge and how these simple AI agents might help with another analogy. So imagine you have a plane. So this is like in gradient descent in machine learning, for example. So you have a plane, and you have hills and a mountain. And you have different hills, okay, of different heights. And you have a mountain way up here that's the tallest mountain. So you drop, I'm going to take four of you, and I'm going to drop you somewhere here, and I'm going to handcuff you together, each of you looking in a different cardinal direction, [00:10:54] and I'm going to blindfold you. And I'm going to say to you, find the highest mountain. So you guys talk among yourselves and you say, well, why don't we each take a step in our direction and report back to the team? So you take a step north, and you say it's uphill from here. And south says it's downhill from here. And east and west say it's lateral from here. So you all agree, let's take a step north. And you keep doing this iteratively until you reach a point when all of you say it's downhill from here.
Have you found the highest mountain? No. (...) What have you done? You found the nearest hill.
Now will you ever find the highest mountain? No. You will never find the highest mountain. You are stuck. You are locally optimized, but globally sub-optimized. And in order to globally optimize, you need a little noise. You need occasionally to allow this group of people to make a counterintuitive step down the mountain or step down the hill. So they take a sequence sometimes by chance of a sequence of steps until they get back onto the plane. And then they navigate all around this fitness landscape, exploring all these peaks until they wind up on this peak. [00:11:56] And this tall peak tends, the global optimum, tends to be a receiving state because it takes much more noise to get off that peak than the other peaks. And so now you oscillate around the global optimum. So in our work, we've been exploring this kind of simple programming inserted into social systems to see if we can improve the performance of human beings in addressing diverse kinds of collective action problems.
Now another collective action problem involves a different challenge, which is cooperation, not coordination. Humans often have to cooperate to produce what is known as a public good. And a lighthouse is one of the canonical examples of a public good. Public good has two canonical features. First of all, it's what's called non-excludable. And that means that other persons cannot be prevented from using it. If you build a lighthouse for your own personal sake, because you're navigating the seas and you don't want to crash into the shore, that's great for you, but you can't stop anyone else from using it, okay? Non-excludable. [00:12:58] And also, it's non-rivalrous. That means that consumption by one person does not reduce consumption by others. If I use the light from my lighthouse, there's no less light for you to use. And this is unlike, for example, a piece of cake. If I have a piece of cake, it's mine, right? I can prevent you from eating it. And if I eat it, there's none available for you, okay? So public good has these features, and it's these features that in turn make it very difficult to produce public goods. Because when it comes to building a lighthouse, it's very tempting to pit your individual interests against the groups. If you don't contribute to building the lighthouse, you can still benefit from it. And so everyone is tempted to do nothing, and then the lighthouse doesn't get built to the detriment of all. And it's also worth emphasizing that public goods are useful, because you can actually produce things with them, such as safe sea travel. And as such, the underinvestment in public goods is a serious problem in our society, and has also come to be known as the tragedy of the commons. These public goods, for example, norms of trust that we maintain among us are efficient. [00:14:01] Think of when you were at a high school. Some of you went to a high school where students trusted each other, and that meant you could leave your backpacks alone in the hallway and not worry that anyone would steal them. Others of you went to high schools when there weren't the same norms of trust, and now you had to lock up your backpack or keep it with you at all times. In which of those two environments do you think you would have better learning? Right? In the former environment. So that norm that's collectively maintained is productive. In that case, productive of learning. Here, productive of safe sea travel and so on.
So cooperation in human groups to produce public goods is challenging, and various mechanisms are required to sustain it. And we've done many experiments that involve dropping humans into network groups and asking them to play diverse kinds of public goods games with their neighbors, manipulating many structural and other features over the years. So, for example, we started years ago with an experiment in which people were put into a network like this. They were introduced to their neighbors, and they played a public goods game from, like, behavioral economics, [00:15:03] where, like, I could give a little bit of money to each of my neighbors. Like, I take a dollar, and I divide it among my neighbors. And then the scientists double the dollar. So let's say I have four neighbors. Let's say I have three neighbors, and I give a dollar to the group. It becomes two dollars. That two dollars gets divided amongst the four of us. So the whole group gains wealth at two dollars, but I only get back 50 cents. The two divided by the four, I get back 50 cents. So I have to make a sacrifice for the benefit of others. So naturally, everyone says, I don't want to sacrifice. Let every other sucker give the money. I'm not going to give anything, and hopefully others will contribute. But, of course, if everyone does that, you get a collapse again. And the best behavior is if everyone contributes maximally. So here in this situation, we start the game. The blue dots are the, the blue people are the nice cooperative people. They give maximally to their neighbors, and they're creating public goods, like building the lighthouse. And the red dots are the exploiters who make no contributions, what's also known as the defectors. And what we find in this experiment is we reproduce a result that's been known for 30 years, [00:16:03] which is that cooperation collapses in groups, right? By the end of the game, multiple rounds later, pretty much everyone becomes a defector, except these little blue people here on the side, keeping civilization alive, you know, among themselves. You can also think about, again, to invoke high school for the undergrads. Remember that situation in which your science teacher assigned four of you to do a group project, and you were going to get the same grade. And you get assigned to four, three other losers. So now you've got two choices. Either you do all the work, and they also get A's because you want an A. Or you say, that's ridiculous. I don't want these lazy guys to benefit from my hard work. And you say, I'm not going to do anything either, and you all get F's, right? That's a terrible dilemma. Well, what happens here is human beings eventually pick the latter option. They all choose to get F's because they don't want to be suckers and keep contributing. So cooperation collapses in social systems as a general result. But what we did is, once again, we added some albeit different kind of bots to the system. We added some bots that were endowed with very simple A. [00:17:05] And these bots were like little marriage brokers. They brokered social interactions. They looked around them locally at who was interacting with who, and they gave suggestions to the people in the system. You know what? You should cut the tie to that defector that's taking advantage of you and form a tie to this nice guy over here instead. And this gentle rewiring advice that only use local knowledge. There's no teacher there ordering people to be nice. No police. No court. No centralized authority that's executing this. Just acting on local knowledge, we found that these bots could, in our experiments with over 1,000 people in 64 groups, we found that not only could cooperation be stabilized, but for the first time ever, we showed a result that cooperation could actually increase from the baseline when these types of bots were added. And DeepMind subsequently replicated our results and extended them in another paper that appeared about a year or so or two after this one.
Now in still another experiment, we explored how bots might affect group creativity. [00:18:07] because finding new ideas is hard. And theory and experiments suggest that groups may be better able to identify and preserve innovations than individuals by sharing their discoveries.
But innovation within groups faces its own challenges, including groupthink. Right? If you put a group of people together, they may prematurely converge on a suboptimal idea.
Or you might imagine, you should imagine, that a group of people might collectively have greater wisdom, might be able to come up with more ideas. For example, if you give a group of people the task of perfecting a fishing rod, you know, the first person might say, well, why don't we put a hook at the end of some string? And then the person is holding the string with a hook. And someone says, why don't we put a stick add to the string? Oh, that's a great idea. So they combine their knowledge and they do that. And someone says, well, the bait with the hook is floating on the surface. Let's add a wave so it goes down. Well, now it goes too far down. How about we add a bobber so we know where it is and so on. And so people innovate, share knowledge across themselves, preserve knowledge across time, [00:19:08] and you get these cultural artifacts that are the compound product of multiple people sharing ideas and being creative.
So we wanted to create a game in which groups of people searched a landscape for an optimal idea. And we decided to use nouns as a proxy for ideas. And we took 20,000 nouns from the classic computer science word-to-vec corpus. So we took 20,000 nouns. And the distance between these nouns could be defined by the cosine similarity metric. You can imagine a hyperdimensional vector space in which cat is more similar to dog than it is to desk. And the way they did that is they looked at how often the words cat and dog co-appeared on websites. So they had a universe of websites and a universe of 20,000 nouns. And they said these two nouns oftenly co-appear and these two other nouns don't often co-appear. They created a 300-dimensional vector space. [00:20:08] And now you can describe how similar are any two nouns. And we decided to use nouns as a proxy for ideas. We took these 20,000 nouns. And then we picked a set of nouns. Imagine we picked one noun, but we picked a set. One noun at random from all of these, like braggadocio, for example, was a noun. So we picked braggadocio. And we say that's the perfect idea that we want this group of people to find. And then all the nouns that are near braggadocio fall away in this vector space. So you have the peak noun that gets the most points, 20,000 points, and all the other nouns to the nouns farthest away. And we put human beings into this system and we say, find this word. We don't tell them the word. And then we tell them the point value of the words. So they start guessing. And as they guess, they get feedback. And they say, ah, this word has more points than another word. And they get closer and closer and they start sharing the knowledge with each other. They're trying to be creative to solve the problem.
So each noun is related to others in a semantic space or a fitless landscape. [00:21:10] And we had 18 different target nouns, as if we pulled a peak up from the landscape of these 20,000 nouns in 18 different locations in 18 different landscapes. And these nouns were deliberately unusual, like fratricide, shoehorn, sarcoma, cartography, and so on. And then we performed experiments involving several conditions where people were alone trying to navigate the landscape, where people were in groups working together to navigate the landscape, or where people were in groups but also had some bots in there trying to help them be creative. And the bots worked because they could pass information from one region of the network to another by communicating with each other. So for example, bot number two could pass the word sky or car to bot number one. So here's a network example.
We drop the people in. The square dots are bots. And this bot has four interactions and this bot has, I don't know, six or seven interactions. The human beings at the beginning start guessing. [00:22:12] They don't have any basis for a guess at all in the first round of the game. They guess sky, car, rabbit, rat, dog, cat, desk. Like if I asked you guys, pick a noun to guess, you would guess house or sun or moon or cat or dog or some small typical noun. And now they are told the point value, like how similar is the word rat, dog, cat, desk, and so on, to sarcoma. And they are told the point value of those nouns in relation to the target noun, which is sarcoma. And then those point values are announced to them. And then this bot can look at the point values of the humans around it and can relay either a random choice to this bot or the highest point value or the lowest point value, kind of a backdoor channel of communication spreading information from one region of group of people trying to solve the problem to another region. So let's look at some examples to kind of fix this, because this can be hard to understand. The people played this game for about 25 rounds. And here's the cosine similarity with a target noun, which is the word fratricide. Okay, so they're 20,000 nouns. [00:23:13] The fratricide is 20,000 points. And the other words have other points. And here's a solo person, a person by themselves that is guessing. So their first guess is bit. And they're told how similar is the word bit to fratricide. And then their next guess is birth. And they get a big bump in cosine similarity, because you can imagine the word birth is closer to the word fratricide than the word bit. Okay? And then they guess. They try money next. That's worse. Then they try monkey. That's even worse. They check. They do a sanity check. They try birth. It hops up again. Then they try baby. That's not a bad guess. Birth to baby. But baby takes them further away from fratricide. And then lady and so on. And they navigate. And they're guessing. And they're guessing. And they're guessing. And they get nowhere near fratricide by the end. Now in this other situation, we have a group of people, but no bots. And now the people can, in addition to making their own guesses, see the guesses of the people around them. And build on the ideas of others. Okay? Created. Like making the fishing rod together. [00:24:13] So this person, their first guess is dog. But their next guess is shield. You can see shield is closer to fratricide than dog is. And this person is guessing and guessing and getting input from his neighbors. And by the end, he gets closer and closer. He winds up with the word foe. Okay? Actually, soldier was his best guess. Had the highest point value during his trajectory.
So how did adding the bots matter? So this, again, shows the summary of the results. On the x-axis is the round. On the y-axis is the mean cosine similarity in this hyper-dimensional vector space, which is a measure of group performance. And here is a group of humans that are acting, guessing solo. So the humans that are guessing on their own don't do very well, right? They, they, you know, it's just chance whether they can somehow, or, you know, some innate ability in each individual human. They get a little bit better with time getting closer. All of the groups outperform the solo. And this is an old result that's known. A group of people is more creative than an otherwise similar in size set of solo practitioners. But what we find is, is that if we add the most similar bot, the bot that looked at its neighbors, [00:25:17] and found, what is their local consensus about here? My humans seem to be thinking that this is a good word. And then passes it to the bot at a distant part of the network. That bot substantially improved the performance of this group of people to make a discovery. You should be able to imagine how this could work in a group of engineers or any other knowledge workers. How, like, you can distribute knowledge in an efficient way, avoiding groupthink and fostering creativity, by designing bots that help the humans to help themselves. The bot here isn't, doesn't have a brain. It's not, it's self-suggesting ideas. It's just helping the humans to spread the ideas among themselves.
So simple AI agents with interpretable behavior can enhance the capacity for creative discovery in human groups by sharing ideas around which there is local consensus in one part of the group with people in a distant part of the group. And as a result, the group can perform better.
Now, we've also begun to experiment with physical systems. And we've added humanoid and non-humanoid robots [00:26:19] endowed with simple AI to face-to-face groups of humans and showed how they can make it easier for groups of humans to work together by helping them overcome friction or an inability to cooperate in their interactions. One of my favorite examples of this is an experiment I did with my former grad student, Maggie Traeger, who's in the back and now is an assistant professor at Notre Dame, is this experiment. In this experiment, we took three real humans that came into the laboratory and a humanoid robot. And we designed a little game, a little railroad track lane game, that is played on a tablet computer, and this group of humans and a robot had the task of laying railroad track from point A to point B, like a little Thomas the Tank Engine railroad tracks. And then we gave them some pieces that they could pick from on the tablet, like straight pieces and curved pieces. But we occasionally contrived, although it looked like there was a mix of pieces and that in principle they should be able to go from point A to point B, we fiendishly designed it so that they weren't the right number of curves to allow them to get from point A to point B. [00:27:21] So they couldn't do it. They couldn't solve the problem unbeknownst to them. Okay? So first each person would take a turn laying a piece of track, then the next person and the next person. They're working together to link point A to point B on their tablets.
And what we did in this experiment is we manipulated, and they played 30 rounds of this game in this virtual world, and we had 51 groups, we manipulated the conversational style of the robots, specifically whether the robot expressed vulnerability by owning a mistake. So the robot said, you know, I made a mistake. Or whether the robot told dad jokes. I'm assuming everyone knows what dad jokes are. Okay. So we also had the robot tell, by the way, that's a cultural universal, like anthropologists in the Amazonian jungle have looked at indigenous peoples and the dads there also tell dad jokes. And the kids are like, I can't believe dad's stupid jokes. But anyway, and actually there's a theory about what dad jokes are meant to toughen up the kids in a way. This is a theory. But anyway, that's a whole other tangent. Anyway, we had our robots tell dad jokes or express vulnerability. [00:28:21] And what we were interested in finding is whether shifts in robot speech had the power to not only affect how people interact with robots, but also how the people interact with each other. And once again, this offers the prospect of modifying social interactions by the introduction of artificial agents into hybrid systems of humans and machines.
So here's a little example of, oh, and I didn't tell you the results. So here are, here's when we have, so these, the thickness of these lines, we set video cameras up to monitor who's talking to whom and how much are they talking. And the thickness of these lines indicates how much person two is talking to person one and so on. So person one doesn't speak to the robot very much. That's a thin line. And when you had neutral robots, you get this pattern. But when you had the vulnerable robot, all these lines get thicker and they equalize. So we found that a vulnerable robot increased the equality of speech among the humans, increased the volume of speech among the humans, and actually, in separate results, [00:29:22] increased the satisfaction of the humans in that kind of environment.
And here's just one clip of two different rounds that illustrate the robot speaks in a neutral way first, in the passive voice, which doesn't much affect the human communication. the robot says, in a very Reagan-esque way, a mistake was made.
But in the next round, the robot says, I made a mistake. And you can just watch what happens here. Let's see if we can get this to work. (25 seconds pause) So across many, many dozens of runs, this is the kind of pattern we find. So a simple manipulation, simple manipulation in the robot speech pattern changes how the humans interact with each other. [00:30:25] And it doesn't, I'm assuming, doesn't take much to imagine how the whole way we design our chatbots and everything else might be affecting not just the Alexa example I gave you at the beginning, how we treat each other. The humans seem to trust each other more and to have more fun in this situation.
Now here is still, and I think this is the last experiment I'm going to show you and then I'm going to wrap up, in still another experiment, we developed a novel cyber-physical platform to test such social and indeed ethical effects of simple types of AI. Because given the nature of collective action problems, the involvement of AI in human groups could paradoxically and unintentionally suppress existing beneficial social norms in humans, such as those involving cooperation and altruism that we have evolved as a species to have. So we have hundreds of thousands of years of natural selection working on us to make us capable of solving collective action problems. So the question is, well, if we delegate some of that agency to machines, will we lose the capacity to work together to solve those problems? [00:31:32] Will we come to rely on these machines, and so now we degrade our innate ability to cooperate and coordinate and create and so on? So in this experiment, collaborating with Hiro Shirato at CMU, another former grad student of mine, we built a platform which involved two little Raspberry Pi endowed little mechanical vehicles, and we connected it to some software we have that allows us to organize online experiments on a vast scale. So people were in their own homes, and they were assigned one of these cars, and they're driving these cars towards each other,
and we had them play the game of chicken. So in chicken, you know, like, whoever gets to the other side fastest wins. So you're incentivized to not yield to the other guy. But if each of you chooses not to yield, then you crash, and you both get the worst payoff. So what humans would do in this situation, if you're playing an iterated chicken game, is you would quickly learn to take turns. This time it's your turn to go straight through, I'll pull over and let you, but next time you pull over and let me go straight through. [00:32:35] If we're selfish, we just keep crashing into each other, like destroying each other time after time after time, or stupidly, we both swerve off, and neither gets the benefit of going straight. So in this situation over here, the yellow car decides to pull over, the blue car just proceeds unimpeded all the way to the other side.
And we used 300 participants and 150 dyads, and they were paid depending on how fast they got to the other side. And then we added some AI. We added auto braking assistance, where when you had a proximity alert, when you got near to the other car, it braked and gave you a chance to decide, gee, I should pull over and let this other guy through, or vice versa. Or we added auto steering assistance, which is, the moment it came near the other car, it just swerved off, okay? And we added a minimal communication function, where people could say thank you, or something like that. Just very minimal communication. And first, we showed that auto braking assistance, where the cars stop at a fixed distance before colliding, increased human altruism. [00:33:38] That is, giving way to others, as the yellow car does here. So adding a little auto braking assistance AI made it easier for human beings to work together and cooperate in the situation. Moreover, allowing the humans to communicate further helps them to make mutual concessions in the auto braking condition. On the other hand, auto steering assistance, where the car simply swerved, completely inhibited the emergence of reciprocity between people in favor of self-interest maximization. The people just surrender their moral agency. They don't bother anymore. They just let the machine repeatedly swerve, and they give up, okay? So all the innate ethical abilities that people had have now been removed by the addition of AI in the auto steering condition, but enhanced in the auto braking condition. And this also should give you pause. You should be thinking, oh my god, every little ditzel we do when we program these AI agents might have good or bad effects on people's natural tendencies.
People's ability to cooperate and take turns and to act altruistically can atrophy, leading to worse collective and individual outcomes in the end. [00:34:46] And in fact, in short, AI can lead to a kind of moral laziness. (..) Here's one final example of lasting change after exposure to AI in hybrid systems. And it's also an example of how the presence of AI can change human-human interactions, even after the AI is not a party any longer to the interactions. So in 2016, DeepMind developed AlphaGo, and in that same year, this AI agent played against Lee Sedol, the remarkable world champion from Korea.
I watched the match. I can't play Go, but I recognize it to be a magnificent game. My son plays Go. (.) And I was really rooting for Lee Sedol. Lee Sedol is like a hero in Korea, like the way we would have, you know, great athletes on our Wheaties boxes and stuff. Like, his is on little noodle packets and on cereal packets. Like, I think it's magnificent that in Korea some brainy nerd like me gets, you know, seen as an important person, okay? So he's very popular in Korea. And he comes out for the first match, and he's too cocky. [00:35:47] I can tell he's too cocky. And he loses to the machine, and then he apologizes to his fans. He says, I'm so sorry. It's five games. It's the best three out of five wins. Then he plays a second game, and he loses again. And now he's getting serious. Then he plays a third game, and he loses. He's lost the competition.
And you could sort of see the audience, and the commentators were marveling at the machine's ability to play Go, making these weird and beautiful moves, some of which they later discovered had been played because we have records going back thousands of years of Go matches played in the Chinese Imperial Court. And they could find, oh, my God, this is a medieval move that the machine made. We hadn't seen it in so long.
And then Lisa Dahl comes back on the fourth game, and he wins. (..) And I wept. I was so happy. (..) Because he had the machine, you know, he had come back for my species, and he had beat the damn machine. Like, in a heroic, like, I couldn't understand, like, the brain power that must have been required for Lisa Dahl to do that. [00:36:49] I was ecstatic at that moment, and so proud of him for still trying, even after he had lost the match.
Now, what's interesting is when Lisa Dahl was interviewed afterwards, Lisa Dahl said that his own game playing changed after the match. So once again, the AI is helping humans to help themselves, okay? He changes how he plays because of contact. And subsequent investigations by other scientists looked at professional Go players and looked at the median decision quality. There's some standard in Go for looking at how good a move is, and median novelty, like how unusual a move is. And they find that when AlphaGo defeats the human world champion in 2016, all around the world, Go players start making better moves that are more innovative, okay? So all the humans playing Go among each other have changed because of AlphaGo has been added. The AI helps the humans to help themselves. So we're continuing to build on our work to design and add simple bots to these and other situations involving social dilemmas and collective action problems. [00:37:55] We're looking at how bots can affect coordination, cooperation, communication, creativity, trust, navigation, sharing, and evacuation. And in our lab, we are not focused on super smart AI, like LLMs or AlphaGo, to replace human cognition, but rather on dumb AI to supplement human interaction. We're not trying to invent super smart AI to replace human cognition. We're inventing dumb AI to supplement human interaction. And our AI can afford to be dumb because the humans are smart. Our AI is like platinum added to an organic chemistry reaction. It's just a catalyst. All we need is the catalyst to help a group of people be better.
And of course, it's important to acknowledge that the reverse is also possible. Social AI can be used to harm groups of people. But our approach offers a number of other technical and conceptual advantages. First of all, these simple bots are intelligible and hence clearly illustrate broader powers and opportunities. Unlike LLMs, which are a black box and you don't know what it's doing, I can tell you exactly what our bot is doing. [00:38:58] It's noisy. It's brokering introductions. It's passing messages in this very specific way. And second, our controlled bot experiments can also provide insight into how human behavior could beneficially also change. In other words, I can take this from the lab and I can teach a group of humans to do what our bots did. In a way, you can't easily teach a group of humans just do what ChachiPT did here. We don't know what ChachiPT is doing, but you know what they're doing in our situation.
So I'd like to close, this is my last slide, with a metaphor. Consider these two objects. They are both made of carbon. If you take the carbon atoms and connect them one way, you get graphite, which is soft and dark. Take the same carbon atoms and connect them another way, you get diamond, which is hard and clear. And there are two key intellectual ideas here. First of all, these properties of softness and darkness and hardness and clearness are not properties of the carbon atoms. They are properties of the collection of carbon atoms. And second, which properties you get depends on how you connect the carbon atoms to each other. Take the same carbon atoms and connect them one way, you get one set of properties. [00:40:01] Connect them another way, you get a completely different set of properties. Similarly, the nature of our connections affects the properties of our social groups. It's the ties between people that can make the whole greater than the sum of its parts. New properties, such as cooperation and violence, innovation and productivity, trust and mistrust, truth and falsehood, wealth and poverty, health and happiness, can emerge and spread because of the connections, because of the ties between people, and not necessarily solely because of the people themselves. In fact, our experience of the world depends on the structure and function of the networks around us near and far. And our species evolved for this to be the case. And it should not surprise us that we will respond to AI in our midst. Thank you very much.
