Subprime Attention Crisis and the Timebomb at the Heart of the Internet — Interview with Tim Hwang transcript
--
Author’s note: this interview was first published in podcast form on December 10, 2020, and mostly deals with the book of the same name. I had it transcribed by some software so some minor typographical errors may remain.
Thomas:
Hello, and welcome to Physical Attraction. This week taking a break from climate, we have a special guest on the show, Tim Hwang. Tim Hwang is a writer and researcher and he’s the author of Subprime Attention Crisis, a book about how online advertising may have become a bubble. He is currently a research fellow at the Center for Security in Emerging Technology at Georgetown University, the former director of the Harvard MIT ethics and governance of AI initiative, which was a big philanthropic fund and research effort trying to advance the development of machine learning in the public interest. He’s also served as the global public policy lead for artificial intelligence and machine learning at Google.
Thomas:
We had a fascinating conversation that started with a little word on deep fakes and misinformation, and then gets into the heart of this subprime attention crisis thesis. Is online advertising a bubble? And if so, what does that mean for the internet that depends on it so much?
Thomas:
Hi, Tim. So first of all, thanks very much for agreeing to come on the show and being so generous with your time, I want to start with a little bit of background, you’ve had a fascinating career, which is involved being the director of public policy for AI and machine learning at Google and research surrounding the ethics and governance of AI at MIT. In that context, you’ve done a great deal of research in some of the issues that generally come into this malicious use of AI, which includes things like deep fakes, disinformation, algorithms used to amplify particular messages, and so on. The main thrust of what we’ll be talking about is your book Subprime Attention Crisis, which is about advertising but I wanted to just ask about this first, especially after the 2016 election, this has been an area with a great deal of heat, and maybe not so much light. So would you like to talk about some of your research in this field? Should we be worried about deep fakes, machine learning being used to influence people’s opinions? If we should, how should we look to combat this sort of thing?
Tim Hwang:
Yeah, absolutely. First off, Thomas, thanks for having me on the show. I would say yeah, on this specific issue of the kind of relationship between machine learning and disinformation, my point of view is that it will be a threat, it is something that we need to be concerned about, but frequently, not in perhaps the ways that we often expect. Or the easiest kind of doomsday scenario to come up with, is maybe not necessarily the scenarios we need to be most worried about. So my background is very much as a software engineer, as a technical person. I’ve been basically brought more and more into the world of kind of policy and regulation and sort of national security strategy over the last few years.
Tim Hwang:
It’s really with that technical lens, that I look at things like deep fakes. In some ways, I think people see the coolest and latest GAN out of NVIDIA, and they say, “Oh, my God, we’re going to soon be awash in deep fakes.” But having come from the technical side, and if you’ve ever played around with machine learning before you know it can actually be quite expensive and difficult to use. GANs are hard to train, you need to acquire a lot of training data. So one of the pieces that I recently worked on in terms of research is a paper that just came out called Deep Fakes: Grounded Threat Assessment.
Tim Hwang:
Really, what it attempts to do is basically say, “Let’s look at machine learning, and from that can we think a little bit about places where the sort of operational constraints of things like deep fakes, make it easier or harder to use the technology?” I always joke that I think about the middle manager at the internet research agency, the kind of like mid level bureaucrat and the Russian propaganda operation. They really have to think about costs and benefit, right? They have to weigh how expensive it is to use a tool against how much disinformation or persuasion or confusion it’s going to create.
Tim Hwang:
I think what’s interesting is if you think about that, it becomes pretty clear pretty quickly that it’s actually unclear if we really will see massive use of deep fakes in this kind of political and election domain. That’s largely because there’s lots and lots of cheaper ways of spreading disinformation that don’t rely on sort of like the latest technologies. So I would say, that’s a long way of saying I’m a little bit skeptical about its use in the political domain. Now, I think what’s interesting is that to go to the very thing, first thing I said is that I do think that these technologies can and will be used in a way to harm people, but I think it has been a lot more evidence in the world of say non consensual pornography. Where people are kind of creating explicit images of people without their consent. That’s a use case where I feel like this kind of cost benefit analysis maybe doesn’t apply so much.
Tim Hwang:
So I do think that like this analysis is helpful, like getting us to a better sense of where are the problems and where aren’t the problems, and I think that’s really more where I sit. That it isn’t necessarily the doomsday technology that will eliminate our ability to tell what’s true and what’s not. I think really more likely it will be producing harm, I would say in places that are really not in the kind of political domain that people have been so worried about.
Thomas:
Mm-hmm (affirmative). I think that’s an interesting perspective because so often our scenario for this has been, for example, there’s a fake video of a presidential candidate which creates a massive scandal. It almost seems like in the last year, at least, the scandals have generated themselves and people haven’t really needed evidence that that’s convincing to silo themselves into different information, ecosystems and so on. You wonder whether the problem with deep fakes is the fact that people could dismiss something that’s real as a deep fake, if they wanted to. It essentially undermines our trust in evidence itself.
Tim Hwang:
I do think that there’s a lot of things that spread disinformation that don’t really rely on the image being incredibly sharp or believable, right? There’s like these poorly photoshopped images that are widely circulated through the web and widely believed. I think that just goes to show that if you’re a real pragmatist about these things, there’s ways of spreading disinformation that really don’t rely on the latest greatest technologies. So I do agree, I mean, the effect that you’re pointing to, which is the reverse which is not necessarily will people believe deep fakes? But rather will people believe or sort of treat video with a little less credibility? Some people call the liar’s dividend. I think that again, is like a much more realistic type threat than the ones that you’ve mentioned. This fake video, the interrupts in election, the day before the election. I would say, in some ways the American election was a natural experiment, and I think the fact that we didn’t really see it is a major influence is a pretty important data point for us.
Thomas:
Part of the issue with misinformation and disinformation is people’s willingness to believe it takes them most of the way there, rather than necessarily how convincing the evidence is. I mean, is that what you see when people discuss disinformation and how it spread through algorithms and on social media?
Tim Hwang:
Yeah, I think that’s correct actually, and it will go to something that we’ll talk about a little bit later, which is that it’s unclear the degree to which, now we’re moving on from deep fakes to the kind of you brought up the kind of case of sort of like algorithms influencing people. It’s interesting, I mean, the evidence for that is a lot less clear than you might think. I do think that we always have to separate what is the impact of the technology versus these deeper kind of social forces. Because again, I don’t think that the belief in a deep fake is going to be based on how sharp the image is. A lot more of it will have to do with does that person prime to believe the narrative that’s being expressed? Do they get something out of sharing that kind of content with other people? Those forces are the ones that really make a difference I think in the spread of disinformation more than this kind of cool machine learning demonstration that someone can do.
Thomas:
Then it becomes a harder problem for the tech companies to solve in a way because their inclination, I think, is to try to engineer their way out of problems that they might have engineered their way into and to say, “What we need is more machine learning and more AI tools that can recognize these fakes or automatically prune out this disinformation.” I often wonder whether it’s intrinsic in the nature of having these algorithms that optimize for attention on a website, or optimize for clicks on a website, which is intricately linked to the advertising industry, which is what we’re going to come on to talk about that is almost really producing this ecosystem where things get siloed into highly engagement content and so on. It’s not even a question of all we need is a better bot to weed out the disinformation and patch the problem so much as, well okay, what aspects of the fundamental structure here are influencing these things to keep happening?
Tim Hwang:
Right. I think there’s a concept that comes up sometimes in sort of defense and national security circles which is the notion of asymmetric warfare, right? I do think that disinformation and media manipulation really is asymmetric in the sense that basically people who are trying to do sort of negative things online, they have lots and lots of low cost options to choose from. You can imagine a world in which we come up with the world’s greatest deep fake detector, or even an algorithm that and we can argue with whether or not it’s even possible, but like an algorithm that tells what’s true from what’s false. I guess I tend to believe that it’s going to be easier for sort of bad actors to be able to get around those defenses faster than we can create good defenses against various manipulation tactics.
Thomas:
So from the US perspective, I guess you could say it’s less of a digital Pearl Harbor and more of a digital Vietnam.
Tim Hwang:
Yes, unfortunately, I do think that is unfortunately the case.
Thomas:
So I want to move on and talk about Subprime Attention Crisis, which is Your new book about the online advertising industry. It’s a really great book, I recommend everyone get a copy. It’s essentially a dire warning, as the title suggests that the industry may be in a bubble similar to that which the housing market was in, in the run up to the financial crisis. Before we come on to that thesis in detail, we need to emphasize how cool advertising is to the way that the internet is structured. It accounts for massive percentages of Google’s and Facebook’s revenue, nearly all of Facebook’s revenue. How did advertising become the cornerstone that the internet is based on and how has the landscape for it changed over the years? I realized that’s a big question, but a sort of potted history of that, I think would really help.
Tim Hwang:
Yeah, no happy to talk a little bit about it, and it really is a fascinating story. In many ways, I think in 2020, I think it’s easy to forget that in the early days of the internet, it was never clear that it was going to be an enormous moneymaker. In fact, there’s all these amazing quotes from people in the early ’90s, being like, “It’s an interesting technology, but it’s unclear if it’s ever going to make a real buck.” I think there’s this amazing kind of nascent period in the development of the internet where people are experimenting with different types of business models, and in some ways advertising became a powerful business model largely because we invented a specific kind way of doing advertising online that’s known as programmatic advertising.
Tim Hwang:
Essentially, this is basically the use of algorithms at very large scale to buy and sell attention online. Really, it was a technology that was necessitated by the growth of products like Google, like the search engine. Where suddenly you had lots and lots of people looking at this website and you had to figure out how to make money off of it. It turns out that programmatic advertising was just a way to make it extremely scalable. One of the things I call sort of programmatic advertising in my book is, it’s a kind of financial jet fuel. That it really is an incredible way of scaling fortunes, incredibly, incredibly quickly, and really is responsible for powering the growth of the defining companies that really shape our kind of day to day experience of the web today.
Tim Hwang:
Really, I think that was sort of an unexpected development. But I think the main thing that really allowed advertising to become a dominant model was that it was scalable, it was cheap to deploy, and you were able to make a lot of money really quickly and it really has become sort of like the cornerstone for how all these companies fund themselves.
Thomas:
I think it might also be help to just get a little bit of an overview about how ads are bought and sold in the modern day internet. I mean, how do these things look from a sort of buyers perspective and who are the big sellers here?
Tim Hwang:
Yeah. So this is actually one of my main reasons for writing the book. You mentioned a little bit earlier that I spent a few years at Google. Google for a company that raises more than 80% of its revenue from advertising. It’s kind of amazing how little people talk about advertising on a day to day basis. It’s kind of a rumor. People say, “Oh yeah, we know, we’re funded by advertising.” But you can talk to engineers at Google where you say, “So how does that whole ad system work?” And it’s just like a magic box, essentially. So one of the missions of Subprime Attention Crisis, whether or not you agree or not with the argument I’m making, is to really describe in detail how sort of modern advertising works and it really is a fascinating and amazing kind of engineering feat in a lot of ways.
Tim Hwang:
So the basics of it are pretty simple to understand. Essentially, as you browse around the internet, you click on websites that are triggered to deliver ads. What’s interesting is basically when you click on a website, though, you don’t really know yet what advertisement you’re going to see when the website fully loads. As I mentioned earlier, the way that we figure this out is a system that’s known as programmatic advertising. Essentially, what happens is that your attention, the supply of attention, is essentially flagged to a market. It’s a little bit like a sort of quantitative trading or even high frequency trading, if you will, where there’s basically a light that goes on that says, “Here’s an opportunity to put an ad in front of Tim.”
Tim Hwang:
Now, what happens is basically that there’s a large number of algorithms online that represent different types of media buyers and they will instantaneously at lightning speed bid for the right to put that ad in front of your eyeball, and based on the price and other factors a winner will be declared, the ad will be delivered and it appears the minute that the website loads. All this happens at split seconds, billions and billions of times a day. This is really kind of the core engine of how it works. The main sellers of advertising are what’s known in the industry as publishers. So this can be everything from your local newspaper website, to the biggest social media platforms, they are basically the suppliers of attention.
Tim Hwang:
Then on the buyer side you have media buyers of all kinds. So this can be everything from again, your local mom and pop shop, to a large multinational trying to promote their products. What programmatic advertising has done is kind of create a massive marketplace where both publishers in many different sizes and sellers of different sizes, or buyers of different sizes can all be sort of together in the same market.
Thomas:
I was actually at the International Conference on Machine Learning a couple years ago, doing some stuff on climate change which is my main area of research. I was fascinated to see just how much of the machine learning community is actually solving all of these very complicated mathematical problems that are ultimately about how to best serve an ad to an audience and who they think is going to maximize the probability of clicking on the ad and so on. I mean, it really is a very mathematically complicated and as you say, extremely high frequency field like this high frequency trading in terms of the density of the algorithms and the mathematics and so on that goes into it.
Thomas:
Again, there’s this question that you come up with, which is, how is it that we know that this industry actually works. The idea that this advertising industry that’s at the heart of the internet, that is behind billions and billions of dollars in profits, or multi, multi billion dollar companies in revenue for these companies, and also has all of these PhDs and geniuses working on it is flawed is of course quite, quite a bold claim to make when it underpins so much of these businesses and so much of the internet. So I want to ask, when did you first come to suspect that something was awry?
Tim Hwang:
Yeah, sure. So I think one of the main triggers was a really interesting story from about two years ago. Procter and Gamble, which is one of the largest advertisers in the world basically decided to run a small experiment. This is the experiment that they did, they basically decided that they would cut about $200 million out of their digital ad spending. What was really interesting is that they reported on the results of doing that just about a year later, and one of the things they found was that there was no impact at all on their business from cutting out all of this money from their budget.
Tim Hwang:
In fact, because they were slightly more efficient than they were usually, they had actually expanded the reach of their advertising slightly. I found that story just so compelling, and so interesting, because it really begs the question, what is all that money going to if not to impact sort of the actual business? I think that story was just so compelling that I started digging and the more I dug, the more that I found, and the more interesting and sort of confusing the situation became.
Thomas:
So it’s interesting, you mentioned the story of Procter and Gamble throwing away all of this money, because there’s a famous advertising aphorism, where an advertiser said, “Half the money I spend on marketing is wasted, the trouble I have is, I don’t know which half.” But digital advertising when it came out, kind of promised to change that by giving you the ability to measure the response to your ads in a more fine grained way than ever before. If you put up a billboard, you don’t necessarily know who’s looked at it, you don’t know who’s been influenced by it. But with cookies and trackers and so on online, you actually should be able to know something about how effective your ads are being. So can we talk about this idea, these metrics for the success of advertising, and how they shape the ad industry? Also how accurate it is that advertisers can really use these digital ads to more accurately measure the impact of what they’re doing. Because you would think, given that, that there wouldn’t be a situation where Procter and Gamble are throwing millions at some ads that don’t seem to be doing anything.
Tim Hwang:
Sure, yes. So there’s a lot there for sure. I mean, I think that often these debates kind of they collapse into a single question, which is, do ads work or not? But I think it’s worth taking a step back and considering how many things need to happen for an ad to even get to the position where we’re debating whether or not it works. So the first question is like, is the ad delivered to a human or not? One of the really interesting aspects of the modern day ad ecosystem is just the enormous prevalence of fraud in the marketplace. There’s a study from a few years back that suggested that almost 60% of display ads are fraudulent in the sense that the ad is sort of delivered, but it’s not delivered to a person, it’s either delivered to a bot that’s designed to click on an ad, or it’s delivered to someone who’s part of a click farm, they’re sort of paid to click on ads.
Tim Hwang:
So that’s the first piece right, enormous prevalence of ad fraud. On the second hand, okay, there’s whether or not we deliver to a human or not. There’s a second question of whether or not that delivery is to a relevant person. There’s a lot of evidence suggests that the data that’s used in targeting advertising is highly faulting. One of the really interesting phenomenon you mentioned earlier, all the machine learning work that’s done, there’s some studies to suggest that basically a lot of this machine learning what it does is it finds people that would have bought a product anyways, even if you hadn’t advertised to them, which is somewhat reasonable because you’re basically trying to pattern match with people who have bought the product in the past. So I think that’s the second piece which is another bit of intermediation.
Thomas:
This is quite common to machine learning algorithms.
Tim Hwang:
Yes.
Thomas:
All you tell them to do is to maximize a certain thing, right?
Tim Hwang:
Right.
Thomas:
And they will find whatever way to do that, even if it’s some trivial way that is the way you don’t want them to do. So a classic example is these things are often tested on video games, and you tell it to complete the level and score as many points as possible. Quite often the algorithm discovers a cheat in the video game, or they’ll find a bug that allows them to score infinitely many points. In the advertising example, this is quite similar to, for example, these algorithms sort of stumbling upon the fact that people who Google the word eBay, are quite likely to then go on and buy something from eBay.
Tim Hwang:
Yeah, exactly. You’re referencing a great study there, that actually was an experiment that eBay had conducted and the results were definitely that, which is that a lot of advertising is delivered to people who would have purchased anyways. So there’s a causal question with some of it’s being done. Then on top of that, there’s another layer, which is okay, say the ad is delivered to a human, it’s delivered to the right human. Then there’s a question of does the person ever see it at all? The problem there is that ad blocking is on the rise in many cases.
Tim Hwang:
Google itself did a study a little while back that suggests that a huge number of ads are just never seen, because they’re loaded because they’re loaded below the fold, for example in a browser. Or they’re placed somewhere where no one even notices it. So finally, finally, if you get through all those gates, we finally get to the question of are ads effective or not? I think this is actually sort of the really interesting thing is that we do have this assumption that all of this data gives you the ability to target ads more effectively. But I think a lot of this intermediation really cast some doubt on the fact that whether or not online ads are that much better.
Tim Hwang:
To be clear, I mean, I think the claim of this book is not necessarily that online ads are uniquely bad, but they may just be as bad as every other previous generation of advertising. So in some ways sort of water makers principle, where you said like 50% of ads are wasted I just don’t know which may still be the case today.
Thomas:
It’s interesting. So we’ve talked about this, we’ve talked about ad fraud, we’ve talked about ads being skipped, ad block. Another thing I like that your book mentions is the so called fat finger clicks, which is the idea that on mobile phones, when people do click on an ad, around half of that is by mistake, and that actually rings true when any of us have had mobile, particularly on websites with bad UX. You click on ads accidentally way more often than you intend to actually click on them.
Tim Hwang:
That’s correct. Yeah.
Thomas:
One thing that is quite interesting here is these things have risen obviously in recent years, since the online advertising has started, ad fraud and click farms have become more popular, people are on mobile, so you have the fat finger rates, ad block has shot up yet the actual value of advertising has not seemed to reflect the fact that the underlying commodity, the attention that’s being sold has changed in value. I mean, how is the advertising industry engaging with this stuff, or was it just not being mentioned?
Tim Hwang:
It is, I mean, some of these problems are well known. In some ways the name of the book is Subprime Attention Crisis, it’s very explicitly trying to draw a connection between what we have in ads and earlier generations of market bubbles. I think one of the things that we see is actually that as with many previous financial bubbles, the problems in ads are known. I think the question is whether or not there’s a strong enough incentive to fix any of the problems given that the money is so, so good. So I think that in some ways, I think I’m pointing out problems that are well known, it’s really just a question of whether or not we’re going to do something about it in time.
Thomas:
So just to talk about the subprime mortgage crisis, for people who have forgotten their history from 2007/2008. This was the idea that people were trading in financial instruments which were leveraged on that subprime mortgages who had been given to lenders who were less likely to be able to pay them back who maybe didn’t have stable incomes and so on, couldn’t afford to pay them back. Then ratings agencies came along and told them that what they were buying was triple A rated debt that it couldn’t default, that it was as good as a US Treasury bond, it was going to pay back. Substantial bets were placed on these financial instruments which multiplied the effect when they imploded.
Thomas:
So the idea that you have here is that the attention that is being kind of parceled and bought and sold on the online advertising industry, is similarly subprime in the sense that if you dig into it, a lot of what you think is valid clicks and engagement that is driving sales. So for example, in the eBay case, they did think it was driving sales, because their way of measuring whether it’s driving sales is, is there a sale after a click? If the algorithm has found people who are going to buy anyway, then you’ll succeed quite often in that case. So it’s almost a case that you can’t necessarily measure what you’re doing and you can’t dig into the quality of the attention so well.
Thomas:
So one of the things I think is a very interesting parallel between those two situations, is as you say, the problems were known about, people at the time knew that they were bundling up these mortgages that they didn’t really know what was in them and they knew that in many cases that what was in them was not actually going to be good debt at the end of the day. Yet, ways in which the system was structured just meant that the incentives for everyone were kind of perverse. So the incentives for the people who were selling the bonds was to get more mortgages to put in the bonds. The incentives of people who were selling the mortgages was getting a bonus every time they sold the mortgage. People wanted to take on loans that they couldn’t pay back, because it gave them the opportunity for a household.
Thomas:
So at every level, you have a different perverse set of incentives, that is getting people to continue with the system that isn’t necessarily working so well. Would you like to talk about some of the parallels then between specifically that subprime mortgage crisis and the attention economy? Again, whether there’s a similar structure where people have these incentives that is making them carry on with something that is not necessarily working so well.
Tim Hwang:
Yeah, absolutely. So I think there’s a number of parallels, of course, but I think maybe there’s two that are perhaps most salient and I think worth mentioning. The first one is, I’ve gotten a very sort of interesting set of responses now that the book has come out and one of the arguments that I hear most from people in the ad technology industry is an argument that goes a little bit like this. They say, “A lot of people put a lot of money into the ad market, and isn’t that proof that ads work?” Because they wouldn’t put the money into it if it didn’t work, right? I think it’s just a fascinating statement just because it parallels so closely the kinds of discussion that were happening during the subprime mortgage crisis, which is, everybody’s putting money into these mortgages. They work really well, and isn’t the proof of the market proof of its success?
Tim Hwang:
I think one of the projects of the book, I think, is to decouple those two things. To basically say, “No, just because a market seems to be succeeding, just because a lot of money seems to be flowing into it, doesn’t mean that it’s stable, and does it mean that the underlying value isn’t collapsing actually right in front of our eyes.” So I think that’s one parallel, I think that’s really worth keeping in mind.
Tim Hwang:
I think the second one which is a parallel that we haven’t talked about just yet, is that a lot of people don’t think about all of the ways in which advertising is intertwined with their lives now. Because I think some people will say, “Oh, well, if online ads fail, will it just be that Mark Zuckerberg has a few less billion dollars?” I would say no, because I think that increasingly all of media journalism is dependent on the programmatic advertising system. You can also think about all the free services that you use, Google Docs, search engines, that really rely on advertising to subsidize.
Tim Hwang:
One of the most interesting things talking about machine learning, is that some of the most leading industrial labs in the world, your Facebook’s and Googles of the world, their research labs they’re not revenue positive. So one way of thinking about it is that there’s a whole world of cutting edge research which is largely subsidized through ads. So a little bit like mortgages, the failure of this one market is likely to have all sorts of collateral impacts that we don’t normally think about and I think that’s the second really important parallel to keep in mind when thinking about the sort of similarities and dissimilarities between now and 2008.
Thomas:
I mean, your book opens with this wonderful anecdote, where you go to one of these conferences, and someone stands up and gives this presentation about all of the flaws in the advertising industry and it’s just sort of met with a kind of, we don’t want to hear this response.
Tim Hwang:
Right. Yeah, that’s right. And yeah, I think that was so striking to me, and again, I feel like an industry that’s unwilling to really seriously deal with some of these problems, really is a bubble. Because I think it’s like now we really have to turn to the question of like, what do we do about it?
Thomas:
Another aspect that’s kind of contributing to this bubble, is a faith in the technology to succeed and deliver. I think part of that comes in this whole idea that there’s this incredibly complex set of algorithms, which is also extremely opaque. So you don’t really know what it is that you’re actually buying and selling all the time. Even a lot of people who are skeptical of the tech industry and of the advertising industry, they still kind of buy into this idea that they have a machine learning mind control ray that can throw loads of GPUs at a really complex problem and shape people’s behavior easily. I mean, how accurate do you think that perception is and whether the marketing material for marketing companies is kind of comparing to the reality of what we can do with machine learning and changing people’s minds at the moment?
Tim Hwang:
That’s right. Yeah, I mean, I think one of the biggest bubbles in advertising right now is the hype around machine learning. I do think that on some level, the idea is very attractive, it’s very seductive. So some ways I like sort of don’t blame people for buying into it, which is, this kind of very intuitive argument, which is, Google has so much data, Facebook has so much data, these advertisers have the best minds working on delivering these ads, how could it not work? I think that anyone who’s worked with machine learning, for instance, for a period of time knows that these technologies are not as sci-fi as they are often sold to be. I think we have a very similar thing that is kind of playing out in the ad space. I ultimately am kind of a skeptic on some of these things and I think we’re seeing this argument play out in very interesting ways. Recently the UK ICO, that has been investigating the Cambridge Analytica scandal, came up with a report where they said, “Look, there’s obviously a huge privacy violation here for what Cambridge Analytica did, but it’s actually unclear if their marketing hype was actually lived up to reality.” Whether or not all this data driven cycle graphic analysis really did allow them to deliver ads in a way that was manifestly more effective. I think those types of stories should give us pause, not just about what Cambridge Analytica did, but about the whole edifice I think of advertising.
Thomas:
It’s interesting, there was Shoshana Zuboff’s book about surveillance capitalism, sort of talking about this being the next phase of capitalism is going to be these huge industries that mine our attention data and mine data about us and then use that to manipulate us into buying things. This is sort of the way that a lot of the big tech companies are trying to go. The idea that maybe that actually isn’t going to be as successful as people hope, I think is an interesting one because it really challenges a lot of where the tech industry is going at the moment. Would you agree with that?
Tim Hwang:
Yeah, I think the book has turned out to be oddly polarizing, in some ways. Because I think that like on both sides of the debate, whether or not you are a diehard Silicon Valley tech optimist, or you are a Shoshana Zuboff, tech skeptic, tech critic. Both sides need to agree that the tech is powerful for them to be either a booster or a critic of it. I think this argument that I’m making almost cuts against both of these sides, right? It basically says that maybe what you’re talking about is nowhere near as effective as it is. So again, I’m sympathetic to say Zuboff’s position which is that, there may be these social problems created by these companies but we may just not need to necessarily rest that argument on the fact that, that this data driven advertising is this powerful persuasion machine. Maybe we really want to resist it on other reasons. We maybe don’t feel comfortable that one company should have all this power, or maybe we don’t feel comfortable that they should be able to hold all this data, and those might be more robust things to kind of build a critique on.
Thomas:
One of the core points of the book then is this fact, as you mentioned before, that the internet is a platform that’s based on advertising and attention rather than alternative models making money, and that this has had a really important knock on effect to the way these platforms try to operate. I mean, you talk about, for example, the reason why we engage with content by liking it, or retweeting it, is because of this stuff. Actually is quite interesting. Twitter, for those of you who are extremely online like me, has recently changed the way they do things. So the default when you retweet someone is to quote tweet them and put a comment on what they have said. Before I started thinking about how the tech industry works, I would have just thought, “Oh, that’s interesting.” Now I can clearly see it’s all about attention and engagement and someone’s obviously figured out that if your default is commenting on something, rather than just passively retweeting it, you’re actually more likely to spark a discussion and a conversation and keep people on the platform and so on. It’s the reason why these platforms are all competing to capture as much of our attention as possible, and this in turn has led to all kinds of widely discussed phenomena like journalism, giving way to clickbait, algorithms that segregate us into groups which target markets and maybe show us some more inflammatory content than we might normally get.
Thomas:
So, aside from the idea that the bubble might burst, which we’ll come on to, I mean, how do you think that the centering of advertising as the heart of the internet, as the profitable heart of the internet has impacted society more widely, and also how we engage with the internet?
Tim Hwang:
So I think it’s a great point, and I think we’re drawing a really important distinction which is basically that, ads don’t necessarily need to work for ads to shape our experience of the web, right? For example, even in a world where ads don’t work, it still may be important, because advertisers may need to kind of collect the audience’s that they can advertise against, and that has very real social impacts. Similarly, like things like the the fact that the advertising world works on click through rates and time on site. That has incentivized the creation of certain types of content, everything from clickbait to the listicle is arguably a kind of outcome of how the way advertising has structured the internet. I do believe that in some ways advertising is a way to kind of think about how the whole internet is sort of put together, you can take any piece of the web and say, why has advertising caused it to take the shape that it has? So I do think that it’s important to decouple those two things, and I totally agree that I do think that like, sort of the ad industry can have a very strong effect, even if ads themselves do not.
Thomas:
So, the natural questions that then come on from this thesis really are, if there’s a bubble, what might make it burst? Specifically, COVID might be something that people would think about here in that respect, and what would the ramifications potentially be? Do you think there are signs that people are getting skeptical about the digital advertising economy and it might face a liquidity crisis because you would think that companies who are obviously always looking to optimize for costs and so on might be starting to read some of this research that’s coming out and thinking, “Well, maybe we should try doing what eBay did, or Procter and Gamble did and we should follow on from their example and experiment with our digital advertising budgets and maybe not just throw hundreds of millions of dollars at Facebook or whatever because we think that’s going to work.”
Tim Hwang:
Yeah, I think that’s right. So it’s worth taking a step back. I think that ultimately what drives markets and what drives a bubble is a social thing. It’s faith ultimately in the effectiveness or the value of the products being bought and sold. So as we think about what might bring the market down, we have to really think about scenarios that might cause a crisis in our faith of the power of digital advertising. So I do think that, yeah, you’ve already pointed out one scenario, which I think is very realistic, which is the experiences of a few large ad buyers, basically shutting off digital ads and not seeing much of an impact is a pretty big deal. I do think that it could potentially create a cascade where a bunch of people say, “Well, maybe we should try that as well.”
Tim Hwang:
I think there’s two other bubbles that I’ll point out. I think the first one is one thing that we kind of touched upon briefly, is the bubble around machine learning. Where basically people have said, “Okay, well, we know advertising is maybe ineffectual right now, but hopefully this machine learning thing will allow us to target ads in ways that we’ve never been able to do before, and really finally crack the nut and really create effective ads.” I just don’t think that’s the case. I think the technology is like a lot of hope and a lot of hype. But I think ultimately, it may not amount to too much.
Tim Hwang:
I think the final bubble that I’ll point out is actually I think that all these privacy laws that are being passed around the world, everything from GDPR in Europe to CCPA in California, have this very real possibility of proving that all of this data may not amount to much. I think it’s a fascinating natural experiment we’re about to all go through, which is that advertisers obviously think that if they lose access to say, third party cookies that they won’t be able to target ads effectively anymore. I almost have the suspicion that we’re going to find out that actually ads work pretty much as good as before even without this data, and if that’s the case, I think it calls into question a whole range of things that could have a really big impact on sort of our faith in the current way advertising is done online.
Thomas:
Again, this isn’t just about Mark Zuckerberg, or someone losing millions of dollars, it is in fact, the fact that so much of the internet, these huge companies that obviously have lots of influence on the wider economy, but also provide search, provide social, provide everything else, all depend so heavily on the ability to generate revenue through advertising. Your book has sort of been a broadside about the idea that they can carry on doing this forever. So I’d be interested in knowing what do you think would happen to these companies if the bubble did burst? And is there any sign that they’re engaging with your critiques here or are they sort of sticking their fingers in their ears and saying, “Lalalalala we can carry on making money this way forever?”
Tim Hwang:
It’s really interesting, I have received some critiques about the book online and what I would really love to do if anyone is listening to this podcast and know someone who would do this is let’s just have a public debate about it. I would love to debate someone from the ad tech industry, I would love to debate someone from Google or Facebook that works in these systems. Just because I think there’s a very real need to kind of air out the situation. Because from my point of view, I think they are very much kind of sticking their hands in their ears and just saying lalala. I do think that there’s a need to kind of like confront them, because if it really is justified then they should have no problem demonstrating that it’s not the case. I guess I’m still waiting for that level of transparency from them.
Thomas:
Again, it comes back to the opacity idea, isn’t it? Because if you can say, “Our magical algorithm is so complicated, you have no way of understanding how it works.” But trust me, we’ve got the data and the analytics, we can’t show them to you, but believe me, we get engagement and we know that it’s genuine because of reasons.
Tim Hwang:
That’s right, exactly. I think it really does, I’ve had debates with friends. I had a friend who used to be a fairly high up product manager at a large unnamed tech company working on these types of systems. He would basically was like, “They work, but if I told you why I’d have to kill you.” Which I feels like’s so strange given that like this is literally what the value of the company is based on that it would be shrouded in such mystery is really puzzling to me.
Thomas:
Yeah, but this is the thing with bubbles, isn’t it? Is people find themselves in the situation quite often where, I mean, this was one of the things I read ECONned, which is about the 2008 crisis. There are points where people are involved buying and selling these collateralized debt obligations, they know that what they’re buying and selling is not actually worth what they think it is. But it’s in everyone’s interest to say that it is and to continue kind of maintaining that illusion to each other for as long as it takes them to offload it.
Tim Hwang:
That’s right. I hear this question earlier that I want to go back to which is, so what does this kind of crash look like if the bubble does pop? I think in some ways the COVID downturn is actually maybe like a light version of what I imagined, which is, in the near term actually, a downturn in these markets helps the big, big tech companies because they have so much cash they can afford to survive, and they can afford to acquire companies that are rendered weak by the downturn. But ultimately, you see the massive human cost, that basically the internet we’ve created is so brittle from a media standpoint, that you have a lot of people suddenly out of work better creating content online.
Tim Hwang:
I think if the bubble got even worse, beyond that point, I’m going to talk about this a little bit in the book, I think it would actually raise some of the same difficult sort of public policy questions that we had in 2008. Which is, do we want to bail out these big tech companies? I think that would be really fascinating if we ever ended up in that kind of scenario.
Thomas:
I mean, you mentioned the journalists, I think one anecdote that comes back to me a lot is just showing you how the advertising industry and dynamics within it can drive journalism and things like people, is the pivot to video.
Tim Hwang:
Yeah, sure. I use it as an example for what people often say is like, “Well, Tim, don’t we have a lot more data to measure ads, and on that count aren’t ads a lot more transparent than they used to be as against a billboard or something like that?” I agree with them, it actually turns out that we do have a lot more visibility into the effectiveness of ads but on an ad by ads level. Generally, we actually have a lot less visibility into how the market as a whole is evolving, and one great example of this is the pivot to video.
Tim Hwang:
So a few years back basically Facebook came out and said, video is going to be the next hot thing on Facebook. We’re finding that people watch a lot of video. So if you want to be content that succeeds on our platform, you need to double down, hire video producers fire other people. A lot of media companies followed suit. They said, “Okay, we’re going to go ahead and do this.” It came out basically, not too long later, that Facebook had basically overstated the amount that people were watching video on their platform by 60% to 80%. They apologize at that point and there’s been lawsuits, and it’s been a real mess.
Tim Hwang:
Whether or not you think this was intentional deception or just negligence, I think it really goes to show how little we know about what actually works and what doesn’t work online, and how much we really have to rely on just our faith in the word of these big tech companies to make assessments about the market as a whole.
Thomas:
Yeah, I think that’s interesting. Again, it’s this control of information flow and the capacity that’s the interesting point here. You compare advertising and the internet, to the role that oil has in our society in terms of being the lifeblood, and the transportation and underpinning so much of how it works. Also, in terms of the fact that the demand for oil affects the economy in so many other ways. I mean, we remember the oil shock of the 1970s where that caused global recessions because of this one commodity that had changed in value. Much as we in the climate space think about how we have to, if we stopped producing oil tomorrow it would be a disaster because no one will be able to transport themselves around. We can’t necessarily have this bubble bursting, what we need instead is a well in climate a very rapid, but in any case, a decarbonisation, a removal of fossil fuels and a replacement with them with something else that will continue to provide us with the services we want.
Thomas:
It’s very similar in what you might want to do in the internet, which is stripped back the role of advertising and the role that advertising plays. If we accept this premise, that it’s not going to be sustainable to have this model of the internet for too much longer than, how can we de-advertise the internet? What might replace it?
Tim Hwang:
Yeah, absolutely. I would say, I’m more of a moderate on this issue. You definitely meet some hawks that say, “Look, we should have no advertising on the internet ever.” I guess I can’t bring myself to necessarily hold that position just because we know that advertising is a good model for certain types of things. We know that, for example, advertising makes services a lot more accessible than they would otherwise be. The fact that Google search is free is a pretty big deal and there’s a lot of people who wouldn’t be able to pay a subscription to it. So I do think that the real kind of crux of the book is not necessarily to argue that we need to get rid of ads, it’s really more to challenge the idea that ads should be the monoculture, the single kind of business model that is really responsible for really shaping a lot of our everyday experience online.
Tim Hwang:
So the vision that I really want to get to is, I would say a more diverse set of business models. So some of it’s very old fashioned, is basically saying we need more subscriptions like that. That is one model that I think is very sustainable. I think there’s been really interesting things with things like micro payments, right, and I do think that finally, like, some of them may also just be a little bit about how we structure the actual business of media. Especially in the kind of COVID era, there’s been a lot of interesting experiments with worker co-ops run by journalists.
Tim Hwang:
I think those types of models are interesting, and again, I think what I want to see is more experimentation in the space. Because I feel like, again, a monoculture is brittle, and its tendency is that when it fails there’s a lot of problems. So I think a more diverse ecosystem is one that sets more robust even though it might grow say less slowly.
Thomas:
Yeah, absolutely. It would make way for other types of website and other types of web experience as well. I mean, podcasting is an interesting venture in this because it has advertising, but also I would say the subscription model with Patreon and so on is something that a lot of people have got into and they’re sort of starting to do it independently, even in some cases to get around the gateways that are associated with advertising and finding ad buyers. So it might be an interesting aspect of how these things are starting to evolve in different platforms that offer different revenue models for people who are making stuff.
Tim Hwang:
Yeah, definitely. I think one of the questions I ask myself often is like, can we find a business model that scales as quickly and generates as much money as advertising? I think if you’re an optimist, you say, “yes,” if we experimented more, we could find that business model. But I also think that we should also seriously consider what it means if that answer is no. Ultimately, we start talking about values, which is, are we okay with an internet that grows more slowly, is maybe less accessible, all these types of things for the opportunity to get rid of ads or reduce the influence of advertising. I think that’s a big policy question but I think really that’s the direction we should think be thinking about is like, what’s the internet we want to have, and that’s kind of where I think the discussion should be pointed.
Thomas:
It’s interesting, because there are different models, too. You talked about what would happen if Google couldn’t be accessed because it no longer had advertising, would we want to nationalize something like that? Would we want to have it run as a taxpayer service? People talk about, for example, the benefit to the economy that Wikipedia has for making the world’s information accessible, that’s another economic model entirely, because that’s voluntary contributions and supported by donations as well. So again, there’s there’s sort of different ways of running the internet that appear in little corners of it that are kind of a little bit outside the ad behemoth and maybe more of the internet looks like that in the future.
Tim Hwang:
That’s right. Yeah. I think that there is a very strong bit of path dependence. I recall very distinctly, I used to live in San Francisco having a number of friends that pitch startups that were basically based on non advertising business models, and they were actively encouraged to prioritize advertising driven business models, because it made a lot of money very quickly. So I do think that there’s some dimension to which like this is the model that we know that works, and particularly if you’re a VC, you want to see those returns. So I do think that there’s this kind of finger on the scale, if you will, to the level of business model experimentation that actually occurs in practice.
Thomas:
Tim, this has been a very, very interesting conversation, I hope that our listeners have enjoyed it. I want to ask you, first of all, whether there’s anything that you’re working on now and other things that you’re working on, and also where people can find your work and find out more about what you’ve written alongside of course, Subprime Attention Crisis, which you should get if you want a really deep understanding of this Subprime Attention Crisis.
Tim Hwang:
Absolutely, yeah. I guess the main project I’m working on right now, later in the book, not to give away the ending at all is I do envision a kind of creation of a kind of like punk rock research center, who’s dedicated to try to blowing up the bubble. I really want to go about trying to build that, and the first project that we’ve launched is something called ad leaker. It is basically a sort of secure whistleblower phone number that people can text over signal. The idea essentially is to encourage people to drop information, cast a little bit of light on what has been hidden in the ad industry.
Tim Hwang:
So that is one project that I’m working on right now. But generally, if you’re interested in these kinds of things, I’m online at Tim Hwang, T-I-M-H-W-A-N-G.org and also on Twitter @TimHwang. So T-I-M-H-W-A-N-G.
Thomas:
Tim, thank you very, very much for coming on the show today.
Tim Hwang:
Yeah, thanks for having me on the show.
