​Michael Bernstein: Welcome to the future of crowdsourcing

From the campus of Stanford University.

People are worried about data. They're worried about their privacy and their security. They should be. We need secure systems.

This is The Future of Everything.

But we can't have a system that closes that data off. It is too rich of a source of inspiration, innovation, and discovery for new things in medicine.

With your host, Russ Altman.

Today on The Future of Everything, The future of crowdsourcing and social computing. So in the last decade, an idea has emerged that difficult problems can be solved by combining the efforts of large groups of people. Crowdsourcing is the word we now use to describe outsourcing— there's where the source comes from— a task to a crowd of people. Crowdsourcing. We hear people talk about the wisdom of the crowd when referring to a wide range of tasks. One of the first examples I remember hearing about was in the context of guessing the number of jellybeans in a large jar. So it turns out that any individual guess by somebody at the county fair is not likely to be very good, but if you take an average or some sort of combination of multiple guesses, you can also often get a very good guess about how many jellybeans are in the jar and therefore win the prize. Similar stories are told about guessing height, weight, and other quantities. Of course, the wisdom of the crowd may also extend and does extend to more profound societal questions. For example, how we choose our elected officials. But crowdsourcing most often these days is discussed in the context of the internet where it is possible to get the assistance either for free or for money or for goods and services of many people to solve tasks. Now Amazon has a service called Mechanical Turk that allows you to package work into small bundles, farm it out to hundreds, thousands, maybe millions of people. And in this case, you pay them a certain amount for each bundle, and they complete it, and you can collect the results. One example may be labeling photographs with words that describe their contents, like captions. You send 10 pictures to a worker and ask them to write down a few keywords describing what's in the picture, and then you pay them a nickel, a dime, a quarter, a dollar for their effort. If you're worried about quality, so you might say, well, Maybe they're going to type in bad captions. Well, if you're worried about that, you can send the same picture to other workers and get like a little bit of a vote within the crowd of what are the best words. And you can do this for thousands of photographs. This has been done and you can get captions for them. And you can also rate the performance of the workers to find out who did a good job and who in retrospect maybe was not doing such a good job with captions. One of the things that Amazon can do is follow workers across many tasks. To figure out, as I said, who does good work and who doesn't, and they can pay high-quality workers more or allow customers to choose how good should the workers be for their task. Obviously, it might be more expensive to get really good workers, and so you have to decide, do I want a few high-quality workers or hundreds of lower-quality workers? What's the best match for my task? The issue of how to create a community for crowdsourcing, how to incentivize it, how to reward the workers, and how to understand the ways that they can be organized for more complex tasks, more complex than writing captions, are non-trivial. How do we get the crowd to solve complex questions that are not easily done by computers but need to be solved? My guest today is Professor Michael Bernstein from Computer Science at Stanford University. Michael is an expert in human-computer interactions. crowdsourcing, and the related area of social computing systems. Michael, what are the social— what are social computing systems, and do we have a good understanding of when they're the best way to solve a problem and when they're not? Sure.

So social computing is an area that's focused on asking how we design online, often computer systems that bring together large groups of people. Maybe that's for work, maybe that's for play, maybe that's out of interest. And these things are both social and technical. You can't just sort of build a playground and say, hey, I built a system and now it's a playground. It actually involves the interactions of the people, these emergent social effects, and the design of the computer software itself. You can almost think of it as a computational version of architecture or urban planning. You're trying to create environments that will nudge people to engage with each other in ways that are positive. Positive or fulfilling.

And so it requires, in order to do this, my guess would be you need a somewhat profound understanding of human psychology. The normal ways we've evolved to interact by looking at each other in the face and having conversations with words in our mouth is different obviously in the internet. How much do we understand human psychology enough to make effective social computing systems, or is that one of the open questions about these systems?

Both of these are true. We have a really, really deep understanding from social psychology, from communications, from sociology, and others, how people interact with each other. But a lot of this does get translated by these different channels, these different media when we go online. And just within online media, having video versus audio versus text chat versus emoji all produce different kinds of emergent results. And as a designer, this is the interesting question. Given what the group of people wants to achieve, what's the right combination of techniques that will sort of get them where they want to be? And that is a big open question. There's not a direct science of saying, oh, just like we can say, oh, you want the bridge to hold this many pounds or this many tons? Well, this is what you're going to need. If we say, well, you want your people in this group not to troll each other? We're not quite at that point.

So the principles of engineering for these computing, social computing systems are still in the nascent phase of discovery and kind of reducing to practice.

Yeah, you know, there's a great phrase which is drawn from urban planning known as a wicked problem. A wicked problem is something—

That sounds good.

Yeah, it's, well, it's terrible, but it's great. A wicked problem is something where essentially, first of all, you can't know whether you've succeeded, And second, there may be many competing goals, and by trying to improve one outcome, you actually tend to make another one worse off. And you can imagine this when you're planning a city. If you're trying to make the roads work more smoothly, you may end up, in fact, just getting more people on the road, producing more traffic. And the same kind of thing can happen with social computing systems. Look at all of the crazy emergent effects and how our Facebook news feeds work, or how people are manipulating each other. intentionally and unintentionally in terms of their news consumption, their emotional states, and other things. We are social animals, and it's a very, very challenging design problem to think about how you balance and find that equilibrium that sort of doesn't mess with everything. Facebook, I don't envy the problem they're trying to solve.

I was just gonna ask you that, because I think that they might not have realized that they were building a social computing system when they built this. This was just gonna be Let's say hi to our friends, let's share our feelings, and let's share some pictures. And all of a sudden now the second and third order effects of these systems are both powerful and somewhat anti— it sounds like a wicked problem to me.

Yes.

So is it in fact the case that you would consider Twitter or Facebook examples of social computing systems, or is that not appropriate use of the term?

That's right, yeah. Social computing has, I think, as many different definitions as people and colleagues that I would talk to, but broadly I would consider it a computing system that is bringing together groups of people. So that could include Wikipedia, although that can also go by other phrases like peer production, which talks about people working on things together.

Right.

You could think about even things that are happening out in the physical world. You could imagine, you know, apps on your phone that might bring people together. Those are also social computing systems. Really, at some level, you know, as long as there are multiple people interacting, through this technology-mediated channel, it's social computing in my mind.

How about, so just one little question about that is, there's these citizen science efforts.

Yeah, absolutely.

The bird, the annual bird count, the where are all the bees count. Would you count those as inbounds for social computing?

I would. I don't think that everyone would, but because there are these, the Venn diagrams move and shift around and overlap in really funky ways between social computing and crowdsourcing and citizen science. I personally tend to take the easiest, back-off route, which is social computing being sort of the broadest category of anything that involves social interaction between people as mediated by technology, and then crowdsourcing as being a more specific goal where you are actually making an open call for help with something.

Yes, yes, so let me, let's think about, a little bit about crowdsourcing. Maybe one question more about the social computing. It's clear to me that in the future we might be looking at hybrid systems where the social, where the individual beings who are participating in the social computing, only some of them are people. So I'm imagining a future where intelligent agents, AI, because of the great advances in AI, might start participating. And I'm wondering if this is an area of interest. Is this what people are looking at? Because I could imagine that the rules of the game, if we know that everybody else is a human, it might be a different set of rules if some of the other people are humans but some of the other people are computers. But Maybe I'm way off. So what is the state of that, of our understanding of those kinds of hybrid systems?

Yeah, that one is fascinating. Some of the most foundational work in this space was actually done here at Stanford by Cliff Nass and Byron Reeves through a book called The Media Equation. And what they found was that they could basically take a given social psychology experiment where they found that person A interacts with person B in the following way, Replace person B with a computer, and surprisingly to them, person A would then treat the computer just like they would have treated person B.

Wow.

But then deny categorically that they had done so. So people are nicer to computers, they like computers better when they're on the same team, even though a computer has no notion of what it means to be on the same team, these kinds of things. And as this moves into an era, a more modern era of, say, human-robot interaction, this has to be taken into account that robots can create emotional states in us, and they can come across as bossy or as collaborative, and we need to design them intelligently for tho— for those reasons. Even in my own group, we've been thinking a lot about what, how computation and AI can even make us better collaborators, even if it's not a member of the team per se. Uh, as an example, when you're on a team, should you be flat or hierarchical? Should you make decisions democratically, or should you, uh, should you, essentially have it delegated, or should you be a bunch of friends, or should you argue with each other?

A very fundamental decision that anybody who's putting together an organization has to make.

That's right, and it turns out that organizational behavior, and my colleague Melissa Valentine in the Management Science and Engineering Department here was a collaborator on this, has an answer for this, which is called it depends. The name for this is structural contingency theory, and the problem is that—

Oh my goodness, that's a mouthful.

Yeah, so contingency meaning it depends, and structural means The structures you use will, it depends.

Gotcha.

And in social science, often the answer is it depends. That's sort of the general way it works, and then the interesting question is on what?

Right, and can you give us guidance so that we can actually build a system that solves our problem?

Right, so as a computer scientist, I see the it depends as an opportunity for AI to actually start playing a role. So what we do is, for example, have an agent that would help the team explore various options. Should we, again, be hierarchical or flat? Should we be critical or welcoming of risky ideas? And over a series of rounds of trying this and getting feedback, the system will learn what actually works well for the team. And this winds up with something like teams that are 40% more performant than having a manager, largely because people are very risk-averse. Again, there's a bunch of work here at Stanford talking about how people are risk-averse. And so they're like, well, it's working pretty well. I think we'll just keep going with it. The robot's gonna go, you know what, let's try that. And it turns out it works.

This is The Future of Everything. I'm Russ Altman. I'm speaking with Michael Bernstein about social computing systems, crowdsourcing, and right now, how to design your team for effectiveness. So let's go to this 40% increase in productivity, 'cause that will get everybody's attention. Anybody who's a manager, that's like getting some free people on your team who are actually effective. So can you give us some insight by going— give us a little bit more detail about the kinds of things that the computer recommended that maybe a person would have, as you said, would have been hesitant to recommend because it might have disrupted the team or appeared like they were doing a power grab. What do these AI systems, what do they watch, and then how do they make the decisions about how to allocate the team resources?

Yeah, and let me start by, you know, caveating here that this is, of course, early work. It's, you know, we've found this experimentally in the lab, and the next step is, of course, to go out into actual organizations and see how well this generalizes. But essentially, this system uses a technique known as a multi-armed bandit, which in fact is the same kind of thing that any web property uses for doing A/B testing and figuring out what's the right color of blue or what's the right link.

So wait a minute. Sorry, I just got to make sure I follow. So a multi— does this have anything to do with Las Vegas?

Actually, yes.

OK, because I think of the one-armed bandit as the thing with the 3 lemons and the 3 cherries.

Yes.

So now what is a multi-armed bandit, just to make sure I understand?

So the terrible mental image that people use is that of an octopus pulling the one-armed bandit. So imagine that you have a row of different one-armed bandits, some of which have higher—

Which is how they appear in Vegas.

Right. And some of which have higher payout rates than others. And what you're trying to do is try pulling various of these bandits to figure out, wait, is that one actually the one that has the highest payoff rate?

Oh, they might not have told you which ones.

No, total mystery.

Okay, so it's an octopus trying to figure out a row of one-armed bandits That's the multi-armed bandit problem.

Right, so what you're trying to do is develop an algorithm that without knowing which thing is the best, try them all and figure out which one's the best as efficiently as possible.

Which is similar to the problem of figuring out of these possible organizational structures for our team, which one should I pick?

Right.

Ah, got it.

And essentially what this means is that you need some sort of feedback function.

Right, did I do well? Did I win?

Right.

And how often did I win?

That's right, or how much did I win, or what's the probability I'd win? And what we end up doing is, in the case of teams, you could either take performance scores, right? How well did this team do on whatever they were trying to do, if there's some way to measure that? And if not, you could have some combination of the subjective feedback of the members of the team. How well's it going?

Yeah, you just gotta figure out good or bad, because that's what you're gonna try to optimize.

Right, and we've developed some algorithms that can sort of mesh this, because these things weren't originally intended to work on humans per se. I guess octopuses have more high-bandwidth computing skills, and so that they can smooth over the amount of change that happens at once, because imagine that if this computing system is saying, try this and this and this and this and this all at once.

No, we turn off the computer at that point.

Yeah, it turns out, yeah.

So it has to understand the rate of change that a human organization can tolerate without having people freak out. throw their hands up and say, I'm out, this sucks.

That's right, and what's more, we're more open to certain kinds of changes early in the process or late in the process, and you can model this, and all of a sudden, it starts working a lot better.

This is The Future of Everything. I'm Russ Altman. I'm speaking with Michael Bernstein about computing social systems, social computing systems, and the possibility that AI systems will be able to organize kind of large crowds of workers in perhaps optimal ways.

Yeah, I would hesitate to say optimal, because who's optimal, right? For the workers, for the manager, this is all gonna be contentious.

It's a wicked problem.

That's right.

Okay, so what about the ethics of all this? Could you imagine that an AI system decides that slavery, I'm just gonna go right to the, hey, I've looked at this and I think that slaves would be a very cost-efficient way to run this business. So how do we think about imposing human operating norms? I'm assuming that slavery is in general not a good thing, totally not a good thing, with computational algorithms that are being programmed to optimize things.

Right. I would argue that in many cases, slavery is almost sort of the de facto state of the world with, you know, work happening. It's— I think that, you know, really, this is something that needs to be that we need to, as designers of these systems, step back, understand, and say, what are we really trying to create here? It's easy for a system to optimize just the number of widgets you produce, but in doing so, you're going to end up burning through the people. They're going to be unhappy. You're going to be unhappy, et cetera. And really, I think the challenge is to be able to state clearly, what is it that we actually want to achieve, right? Back in the days of— what organizational goals do we have? And it's not just, Profits, right? It's gonna be something that's—

So you can build these other considerations into the way you decide who, what's winning and what's losing.

We absolutely have to.

This is The Future of Everything. I'm Russ Altman. More with Michael Bernstein about crowdsourcing and social computing systems next on SiriusXM Insight 121.

The Future of Everything with Russ Altman from the campus of Stanford University.

Welcome back to The Future of Everything. I'm Russ Altman. I'm speaking with Michael Bernstein of computer science at Stanford, and we're talking about crowdsourcing now. Michael, so crowdsourcing is a thing. What is it good at, and who's using it currently in serious ways as parts of companies or other kind of serious institutions that need to get problems solved?

Yeah, so let's start with, that's right, where it's at right now. And I would say that basically every major internet property that you're probably using is using either internally or externally some form of crowdsourcing. The biggest use of it that I come across is people who are training machine learning algorithms. So what is machine learning? It's an area of artificial intelligence that's trying to essentially make guesses like a human would. So how does Facebook decide what to show you at the top when you show up? How does Google decide which search results are good or bad? How does your map software know what businesses are open and when? All of these are things that have been labeled by people.

Absolutely.

And this, you know, can either happen through, you know, internal crowds within these companies where maybe a company has contracted a large group of people around the world to work with them, or in these open marketplaces. You indicated Amazon, like the bookseller online, has an entire marketplace called Mechanical Turk. Well, workers will do these kinds of tasks for payment. It's essentially, think of it like a big, almost like eBay, where instead of selling stuff, you're selling your own time.

Yes, so these folks, how complex do the questions get? So in my intro, I talked about, here's a picture, please tell me, you know, it's a red bicycle in front of a fire truck, or it's a little boy playing with a ball. So that's pretty much a universal capability among humans. How complex do the requests that these internet companies are making to these workers, how complex do they get? 'Cause some of the things you said, like you might require a degree in marketing to know what's the best thing to show somebody on Facebook right now so that they press a button and buy something. So does it really get to that level of complexity at this point?

It can. I think the complexity has been the challenge. So anything, you know, that where it's a classification problem at the end of the day, like yes or no, this is a dog or not.

Something that a human does Kind of almost intuitively.

Yeah, I would say that essentially the biggest use of crowdsourcing to date has been essentially for anything that any human could do. Sort of generic cognitive perceptual equipment gets us there. And so, is there a dog in this picture? Or what would you describe as happening? These are all things that basically anyone with language skills can probably do. And this is important because actually AI can't do that yet. Doing this, we're training the AIs how to do it. But it can be really, really honestly demeaning work, especially think about, you have entire groups at places like Facebook where they need to actually be moderating potentially very adult or very scary or very underhanded content, right? And so there's a lot of that right now.

So if somebody hasn't filtered that out, you could be showing images like it's not just a red ball in front of a fire truck, it's a terrible image from war.

That's right.

If you're not careful, you've thrown that up in front of some poor person who's getting a quarter or 50 cents to label it, but you've just damaged them.

And in fact, that's their entire job, is to look at just the images that we're worried about. Oh my goodness. And this can lead to huge psychological trauma, as you might imagine. So these are the kinds of things that crowdsourcing is being used for now, mostly, I would say. Sort of data annotation, data labeling, and a ton of companies are doing this.

So now in my field of medicine, there's been some very high-profile papers you might be aware of where, for example, systems have been trained to look at moles on the skin to figure out if they're likely to be melanoma. A hugely important thing. And the initial results show that the computers are doing pretty good, but they needed to get that information about is that mole normal or not normal, probably from a bunch of dermatologists.

That's right.

Now those dermatologists, as you may know, get paid pretty well. So is it just that you have to ratchet up the amount of money that you pay to get a dermatologist into the crowd, or are there other strategies that people are approaching for when it's requires deep human expertise.

Right, so sometimes the engineering question in my world is actually exactly that one. Is it possible to scaffold a number of non-experts through the design of the task and through training to be the equivalent of an expert? And often the answer can be yes in certain domains. I would hesitate to say—

So kind of bad news for the dermatologists, that we might be able to train— fellow humans who didn't go to medical school to help with this task? Is that what you're saying? I wanna make sure I—

Potentially. So think of it, I would think of it this way. If you can contain everything that's necessary to know into the context of the task, so I can look at this thing and only this thing and figure it out, and you can tell me everything, all the decision points I would need to know, then yes, absolutely. But you know what, doctors and dermatologists take in a lot of other information, and that's where it actually starts to get, I think, out of the hands of what we know how to do with sort of standard, what would be called microtask crowdsourcing.

This is The Future of Everything. I'm Russ Altman. I'm speaking with Michael Bernstein about crowdsourcing. On this same issue of the dermatologist, this actually clearly relates to our earlier discussion because you could imagine a hierarchy where untrained people look for very simple things like big brown spots or spots that have the following features that are very well, easy to describe, and then they pass up some of the cases to more expert and more expensive workers who then make a final call. And so this is where you were talking about hierarchical structures versus flat structures. You could imagine that in the dermatology example, your AI system might figure out, because I have less dermatologists and a lot of non-dermatologists, maybe I need to create a workflow that takes advantage of their expertise and doesn't make them tired or bored by doing simple things. I'll just give them the hard cases.

Right, and that's exactly what we're seeing. And if we expand out, the kind of work that my colleague Melissa Valentine and I have been doing actually starts to think, well, okay, but at some point, these dermatologists themselves are actually going to be connected via the internet to these online labor markets. If you zoom out from Mechanical Turk to things like Upwork and Freelancer.com and all sorts of things like this, all of a sudden, imagine that you could have any expertise in the world, you know, just whenever you need it.

Right.

And that is a very different proposition. And what we've been working on are systems that can, for example, convene entire organizations in, say, 15 minutes, allow them to pivot as soon as they realize that something's necessary, basically have online organizations that are far more adaptive and quick and rapid, and can accomplish basically anything that, you know, a small organization or nonprofit could. And that, I think, is going to lead to all sorts of differences in both production, how we get stuff done, but also other work that we've been doing in terms of what collective action is going to need to look like. You know, when you're changing the modes of production, you need to ask, what's life going to be like for the workers and how are you going to empower them?

Yes, I'm struck that this future vision, which looks very possible technically, has profound social questions. So, and I guess my question to you is, do you worry about those or is that somebody else's problem?

No, absolutely. I think that, you know, as As a humanist cum engineer, we have to think about these kinds of things. In the United States especially, our social safety net has actually been bundled up very tightly with the organizations that we work for. Stanford pays for a lot of my healthcare. It's responsible for part of my retirement, and so on. And if all of a sudden we're shifting to a world, and there are a number of people who are already there.

They're already in that world.

Where, okay, all of a sudden, Uber, as another example.

The 1099, so to speak, economy.

That's right. And now all of a sudden, we need to think about how we create, what are the needs we have, and then figure out how we actually encourage this. And I've been working, you know, speaking to a number of people in a number of economies. The film economy, like in LA, how do writers and actors and people on sets who film, who make movies, they have a very, you know, strongly unionized way to make this happen, and it's very successful. Construction workers as well. I think there are models out there that we need to, in my opinion, either design into the system or just let things go wrong, which I'm not in favor of, and eventually we'll get sort of policy moderated.

And patch it up, but after the fact.

That's right.

So when you talk to these workers, these Mechanical Turk workers or other kind of crowd, the crowd.

Yes.

What are the issues that they are worried about? Are they worried about ownership of the product? For example, do they sign away any future profits that Facebook and Google might make from their work? And is it, that's just the deal, you're gonna get paid a certain amount, and then we own the results? Or, I'm thinking about this because in medicine there's been some question about whether patients should, who participate in clinical trials, should be participating in the profits from the drugs.

Yeah, so under current law, the freelancing means that if you hire me, you get the IP from whatever I produce. But that obviously not only has financial implications, but in terms of my own motivation—

Exactly.

That changes things. So now imagine that you and I wanted to work together online. Maybe we need to rethink, what does this all work? Think, how does this work? What does IP look like? How do we actually take ownership of these kinds of things? And I think if we want to be creating the next wonder of the world in these kinds of models, we're going to need to produce new kinds of sociotechnical infrastructure. That is new computer systems, new policy, new code, That makes it happen in a productive and a pro-social way.

Now, I'm almost afraid to ask this final question, but who is appropriately thinking about these social issues related to the 1099 economy and the workers in this new age of work product? Who is it appropriately coming down to? Is it the federal government? Is it the institutions who are using these workers? Is it a mix? Uh, or is the conversation actually right now not happening?

You see it happening in different ways in different places. So in academia, we're starting to see a very robust conversation across my own field of human-computer interaction, people in, in economics, anthropology, and many others starting to think about what, what's going on here. At the policy level, you see, and, and in the courts, you start to see some cases asking about, yeah, the, the, the 1099 economy and, and whether that's really a, a good thing or not. And then you do actually start to see some companies, many of which are— many are not waking up to this, but some are. You do see some companies that are starting to say, well, our entire workforce is going to be essentially, you know, say, crowdsourced designers.

So we need to take care of them.

So we—

yeah, we like— in the same way that organizations used to do this, we need to come up with a way that they're happy to be working with us because you know what? The friction is so low that they could just go work for someone else through a click.

Right.

And so we need to be the best place for them.

Thank you for listening to The Future of Everything. I'm Russ Altman. If you missed any of this episode, listen anytime on demand with the SiriusXM app. If you'd like to ask a question about this episode or a previous episode, please email us a written question or a voice memo question. We might feature it in a future episode. You can send it to thefutureofeverything@stanford.edu All one word, thefutureofeverything, no spaces, no underscores, no dashes, thefutureofeverything@stanford.edu. Thanks again for tuning in. We hope you're enjoying the podcast.