Lone wolves. Terrorist cells. Bad apples. Viral infections.
The language we use to discuss violent extremism is rife with metaphors from the natural world. As we seek to understand why some humans behave so utterly inhumanely, we rely on comparisons to biology, ecology and medicine.
But what if we’ve been working in the wrong scientific discipline? What if the spread of hate is less like the spread of cancer through the proverbial body politic and more like … the formation of bubbles in a boiling pot of water?
That is the contention of Neil Johnson, a professor of physics at George Washington University and the lead author on a study published this week in Nature analyzing the spread of online hate. If that sounds like an odd topic for a physicist – it is. Johnson began his career at the University of Oxford, where he published extensively on quantum information and “complexity theory”. After moving to the US in 2007, he embarked on a new course of research, applying theories from physics to complex human behavior, from financial markets and conflict zones to insurgency and terrorist recruitment.
Johnson’s unusual approach has resulted in some surprising conclusions – he says all online hate globally originates from just 1,000 online “clusters” – as well as counterintuitive policy proposals. On Wednesday, he spoke to the Guardian about his findings.
The interview has been edited and condensed for length and clarity.
How did you go from physics to studying these social issues of violent extremism and online hate?
Most people think of physics as smashing things up into smaller and smaller pieces, but there’s actually a whole wealth of physics which goes in the other direction and looks at what happens when you put things together. If I put molecules of water together, well, suddenly, I get a liquid and ice forms and icebergs form and the Titanic sinks. There’s all sorts of consequences of what happens when you put together objects, good and bad.
We have a tendency to want to pin blame on individual objects, but you would never do that in the physics world. There’s no bad molecule that causes water to boil. It’s a collective effect. And so, we wondered if a lot of the social problems that we face are actually better looked at through that lens.
[For this study], we just naively said, well, what does the online world of hate look like? So we set about trying to work that out, and we found an unbelievable global network of hate.
I study networks in biological systems, economic systems. This is the most complicated network I’ve ever studied – tenfold more complicated – because it mixes geography, continents, languages, cultures and online platforms. Trying to police it within one platform is a little bit like saying if you take care of the weeds in your own garden, you can eliminate the problem from the neighborhood.
You talk about hate in terms of chemical bonds and “gelation theory”. How did you develop that framework?
These are not analogies. We looked at the behavior of the data, of the numbers, and saw that it is similar [to chemical bonding] not just because the numbers change in a certain way, but actually microscopically, in terms of interactions.
If you have milk in the fridge, gradually, one day that milk suddenly curdles. That is because microscopically, you’re getting this aggregation of objects into communities. And the math of that works perfectly well for the aggregation of people into communities. Now, the typical reaction is: “Oh, but I’m an individual, I don’t behave like a molecule of milk.” Yeah, but collectively we do, because we’re constrained by the others. So there’s only a certain number of things that we can actually do, and we tend to do them again and again and again.
So it’s not an analogy. People say [online hate] is like cancer, it’s like a virus, it’s like this, it’s like that – no. It’s exactly like gelation, which is another way of saying the formation of bubbles.
How did you create your map?
We started with a seed of clusters that were already banned on Facebook, such as the KKK. We looked at what other clusters they connect to that also connect back to them and kept going through this chain.
We found there’s a closed network of about 1,000 clusters, worldwide, online, across all platforms, propagating global hate of all flavors. Now, if there’s about 1,000 people in each of those (it’s actually between 10 and maybe up to 100,000, so let’s just say 1,000 on average) you’ve got 1,000 clusters of 1,000 people – that’s a million people. And that’s our very, very crude first estimate of the number of people online involved with this.
That’s a startlingly manageable number – 1,000 networks.
Not if you’re trying to find that among seven billion. But they’ve already done the job for you. They’ve already grouped themselves into community.
Do you have a list of those 1,000 groups?
That’s what we built. I was expecting us never to end the process. But we got to the stage where we thought, my goodness, we’ve mapped out the universe of online hate to some degree. Now we can begin to understand how interconnected things are and what things look like.
What was the range of ideology in these hate clusters? Is it mostly antisemitism, racism, white nationalism?
We thought we would see a discrete, well-defined set of boxes. But just as people haven’t been able to categorize [mass] shooters in defined sets of boxes, we didn’t find those online. Instead of it being like a menu of flavors, it’s actually a continuous spectrum. And it’s not a spectrum along the one line, it’s multidimensional.
In the study, you describe forms of resilience by hate groups when they’re threatened. Particularly concerning was this warning: “Policing within a single platform (such as Facebook) can make matters worse and will eventually generate global ‘dark pools’ in which online hate will flourish.” Can you describe how this works? What do you mean by a ‘dark pool’? Is that like 8Chan?
No, it’s even worse. 8chan is a little bit of a remote island by itself. I’m talking about dark pools forming within the major commercial platforms that we “trust”.
When the KKK got banned from Facebook, they were looking for a platform, and suddenly, there was this welcoming committee on [the Russian social network] VK. It was like orientation week in college [with people saying], “We’ll hold your hand, we’ll take you to see the community, and you’ll find what you want here.” They’re now in a kind of close-knit group like freshman orientation, and they quickly can develop bonds to discuss what was it that got them banned, and therefore how to avoid it when they go back in.
One of the adaptations was they then reinserted themselves into Facebook with a simple change of writing the name [KKK] in Cyrillic, the Russian alphabet. It looks pretty similar and yet a machine-learning algorithm doesn’t know about Cyrillic. It’s clever.
How did you come up with your four policy proposals?
If I want to stop water boiling, I don’t have to stop individual molecules from jumping up into the steam, I have to stop the bubbles from forming. We know that the big bubbles form from the smaller ones. And the big bubble today becomes old news for the next generation.
[The first proposal] is to go after the smaller bubbles. Smaller bubbles are weaker, have less money, less powerful people, and will grow into those big ones. So eliminating small ones – and we showed this mathematically – rapidly decreases the ecology. It cuts off the supply.
Number two is that instead of banning individuals, because of the interconnectedness of this whole system, we showed that you actually only have to remove about 10% of the accounts to make a huge difference in terms of the cohesiveness of the network. If you remove randomly 10% of the members globally, this thing will begin to fall apart.
It’s a really interesting idea. It also seems a bit fraught in terms of our general understanding of what is fair, to say that just 10% of the people engaging in a specific negative behavior should be punished.
Ostensibly all of the people involved in this have broken the terms of service of the platform, so all of the people should, in principle, have their accounts removed. Facebook’s trying to remove them all anyway. Our push is, just don’t go after the most important people first.
Your second proposal involves deploying “anti-hate clusters” to engage with hate groups. How could that work?
You get [the hate clusters] engaged in a skirmish, basically, and they think that that’s a kind of supreme battle. It slows them down in terms of recruiting; it just engages them in something that actually isn’t that important.
So you’re saying that fighting with trolls online is actually worth your time?
Right. But do it as a group, do it as a cluster. Don’t do it individually. It will break you.
Are there examples of this that you’ve observed where it’s been effective?
I haven’t. Like I said, this comes out of the idea of: how do I cancel a bubble? Well, there’s no such thing as an anti-bubble. But there is in physics the idea of just getting two [opposites] together and they should neutralize each other. They form a tightly bound pair, the plus and the minus.
[The fourth proposal] is my favorite because it it really exploits the weakness that comes from the multidimensional flavors of hate. There are two neo-Nazi groups, both in the UK, both ostensibly wanting the same thing. But they don’t – one wants a unified Europe, the other wants to break everything apart and obliterate the rest of the countries. So introduce a cluster that draws out the differences. I don’t think that [the introduced cluster] would necessarily contain members of the public unless they were trained in some way. But it will certainly contain people who use good psychology, social psychology, and know some history that could actually engage.
That strategy sounds like some of the tactics that the FBI used under Cointelpro when it was trying to pit different sectors of the civil rights movement against each other and taking advantage of potential ideological divisions to create a splintering effect.
I don’t know that [history], and I certainly wouldn’t want it to be used in some bad way. But I see it as a way to wear individuals in hate clusters out. In the end, they’ll just get fed up. It’s not that it goes away. It’s just that now they’re actually hating the traffic more than they hate the Jews. It shifts the focus.