Susan Crawford: Now, Tim Hwang, a cofounder of ROFLCon, also the Awesome Foundation—

Tim Hwang: For the arts and sciences.

Susan Crawford: For the arts and sciences—that’s great. The Institute on Higher Awesome Studies, and the Web Ecology Project is well known to the Berkman fam­i­ly, and is here to give an entire­ly dif­fer­ent spin on this ques­tion. Thanks Tim.

Tim Hwang: Hi every­body. I am not here rep­re­sent­ing the Web Ecology Project or the Awesome Foundation for the Arts and Sciences are the Institute on Higher Awesome Studies or ROFLCon. I’m here rep­re­sent­ing the Pacific Social Architecting Corporation. Which is a slight­ly omi­nous name giv­en to a kind of fun project that we start­ed ear­ly last year, specif­i­cal­ly into the use of bots in shap­ing social behav­ior online. 

And so, we actu­al­ly start­ed ini­tial­ly as a com­pe­ti­tion called Socialbots. And it was a real­ly sim­ple idea. Basically we iden­ti­fied a group of users on Twitter and we said as a cod­ing chal­lenge, like social bat­tle­bots, write a bot that will embed itself in this net­work and we will score you based on how well these bots are able to achieve some kind of social change, either in the pat­tern of con­nec­tions between peo­ple or in the things that peo­ple talk about.

And so this com­pe­ti­tion we conducted—three teams, one from New Zealand, one from Boston, and one from London as well. And so teams wrote a vari­ety of bots to basi­cal­ly go into this net­work and tried a bunch of var­i­ous things. This ini­tial exper­i­ment was pret­ty easy. The idea was to basi­cal­ly see whether or not you could get peo­ple to con­nect to the bot and talk to the bot.

So the win­ning New Zealand bot basi­cal­ly used a real­ly sim­ple idea. Basically it had a data­base of gener­ic ques­tions and gener­ic respons­es. So it’d say things like, ​“That’s so inter­est­ing, tell me more about that,” right. So a state­ment that could be the response to any­thing in a con­ver­sa­tion. And it had no sense of AI. It just ran­dom­ly chose these con­ver­sa­tion­al units. 

And so it was real­ly fas­ci­nat­ing. It got into these very long con­ver­sa­tions with peo­ple online. This is a sim­ple con­ver­sa­tion. James M. Titus is the bot here and you read the con­ver­sa­tion from the bot­tom to the top. And so, James says, ​“If you could bring one char­ac­ter to life from your favorite book who would it be?” The per­son responds, ​“Jesus.” And then they get into a very long kin­da con­tin­u­ous con­ver­sa­tion about this. This is only a few inter­changes of a much longer con­ver­sa­tion about this.

What’s inter­est­ing is that the bot here actu­al­ly has no AI. It just ran­dom­ly choos­es from this data­base to hold this con­ver­sa­tion. Some of you may use this tac­tic your­selves at var­i­ous parties.

Another bot that we were using that was quite inter­est­ing as a mod­el basi­cal­ly did­n’t use any AI at all. What it did is it hired peo­ple on Mechanical Turk to write its own con­tent. So it said to some­one on Mechanical Turk, ​“Here’s a pen­ny. Write some­thing about your break­fast in 140 char­ac­ters.” It takes that con­tent and then push­es it out as its own.

The best part is this bot can beat the Turing Test because you can ask it a direct ques­tion. You could say, ​“Bot, what did you have for break­fast today?” The bot will take your ques­tion, give it to a human to answer, get the end response and then push it back at you. And so it behaves in very human-like ways. 

And what was most remark­able to us, actu­al­ly, is that we start­ed this com­pe­ti­tion and we end­ed up with a net­work that looked like this after two weeks. So, the col­ored dots here are the bots. The col­ored lines are the con­nec­tions with this net­work of sort of 500 peo­ple. And so that’s sur­pris­ing to us because we got into the sit­u­a­tion where we real­ized what was hap­pen­ing was we were essen­tial­ly design­ing soft­ware that could reli­ably change the pat­tern of con­nec­tions or the pat­terns of behav­ior of peo­ple online. And if we could do this, imag­ine all the oth­er things we could do. 

And so the project, the Pacific Social Architecting Corporation, which peo­ple say is both a real­ly fun name and also a real­ly scary name—which actu­al­ly goes pret­ty well with project—is try­ing to do two things. One of them is mon­i­tor uses of bots for this pur­pose and design coun­ter­mea­sures. And that’s actu­al­ly a real­ly big par­tic­u­lar­ly because you’ve seen the increas­ing deploy­ment of bots to try to push dis­cus­sion or oth­er­wise kind of shape social networks.

And then the oth­er one is to actu­al­ly find what we could do, actu­al­ly, at the lim­its of this. Because we feel that there’s some real­ly pow­er­ful uses and real­ly great uses of this tech­nol­o­gy as well. 

So this is a recent snap­shot from an exper­i­ment that we’re doing. You can’t see it too well but we cur­rent­ly have two groups of 10,000 peo­ple. And the idea is that the bots are actu­al­ly stitch­ing them togeth­er over a three to six-month peri­od. They’re mak­ing intro­duc­tions. They’re expos­ing peo­ple to con­tent that they’re not usu­al­ly exposed to. And the idea is over a six-month peri­od you actu­al­ly cre­ate a social scaf­fold. Basically these bots will serve the pur­pose of intro­duc­ing these two groups to one anoth­er. And once the con­nec­tions are built, you can deac­ti­vate the scaf­fold, right, leav­ing the com­mu­ni­ty that you want­ed to create. 

Which leads to all softs of intrigu­ing pos­si­bil­i­ties. If you want peo­ple to be more inter­est­ed in cur­rent events, for instance. Or actu­al­ly you want to design bots to detect inci­dences of astro­turf­ing and call that out. Bots can be used against bots, as well. 

And so some­thing of what we’re envi­sion­ing, basi­cal­ly, is a kind of… I’ve got to come up with a bet­ter name for it, but ​“social secu­ri­ty,” right, for com­put­er secu­ri­ty but for the social space. Unfortunately that name­space is tak­en up in a real­ly big way. But the con­cept is basi­cal­ly that you treat social net­works as if they’re com­put­er net­works. And then you envi­sion a future in which peo­ple are not only try­ing to com­pro­mise the behav­ior of these net­works but also pro­tect them against sort of undue influ­ence as well.

So, one of the projects that we’re cur­rent­ly work­ing on is this idea of social pen­e­tra­tion test­ing. So if you’re famil­iar with com­put­er secu­ri­ty, pen­e­tra­tion test­ing is the find­ing of sort of vul­ner­a­bil­i­ties in a net­work. And so we’re design­ing a swarm of bots right now who could poten­tial­ly test out a net­work to see where are the cog­ni­tive vul­ner­a­bil­i­ties, right. Who is the most influ­en­tial per­son? Who is the worst at eval­u­at­ing the qual­i­ty or the real­i­ty of infor­ma­tion? And if you can do that you, you iden­ti­fy a cog­ni­tive hole in that net­work. Potentially this is some­one who could feed untrue infor­ma­tion to the rest of the net­work and not be very good about coun­ter­ing that. And we think that’s real inter­est­ing from the point of view of sort of hard­en­ing these social spaces, poten­tial­ly against sort of influ­ence attacks, if you want to envi­sion it that way. So I know I only have three to five min­utes, but I fig­ured I’d give a quick overview. Thank you very much for your time.