As researchers, this is beautiful.

This undermines a lot of the interventions that I was thinking about.

Right.

Okay, so the bottleneck, once it gets to the media platform, then you start to see people sharing it online.

But here, if you’re going through these media outlets, does that then mean that spread is limited just to people who are in that subsector? Or does it from there, how does it spread?

Interesting. But the reason when it goes viral, we know that’s a problem because it’s reached so many people, right? And that means that thousands or even millions in some cases of people have at least been exposed to that piece of information, if not actively absorbed it.

Can you help me understand, in the united states, one of the main ways that this works is similar but different in that, you know, sort of pieces of foreign sponsored disinformation go viral on twitter and then, or they start from some unverified account and then, you know, a ...

Is that a difference of funding, a difference of institutional structure?

Yeah. Doesn’t Japan have this?

Yeah.

Okay.

I can point you to some readings on this.

Yeah.

Can I ask you the thing that keeps coming up for me? Right. We talked before about legitimacy. We talked about it. Okay, go for it, Robert, I’ll come back to this question. Yeah.

So you’ve got this recent real potential change to the information you get.

And how can you give me the timeline of when that’s happened?

Vickie’s, you are a journalist.

This resonates for you.

Civics education.

Can I ask you, going back to this idea of a civil society and the importance of a complex civil society with many effective, different forms of institutions where journalism fits into that?

Right?

So what makes you think that it’s machine generated? Machine generated and specifically generative AI.

Yeah.

That could just be somebody who speaks or who is used to writing in simplified chinese, trying to write in, it could just be a person in a troll factory?

When you talk about. Sorry, do you want to go ahead, robert?

Yeah, you can’t keep up.

Exactly. So the platform brings the reach. And then what we talked about before is they bring legitimacy. Is there anything else that they’re bringing or that the platform can’t do that these organizations can?

So I’ll give you an example. Right. Like MyGoPen, taiwan fact check center, these are enrolled in the ICFN third-party verified fact-checking network. And then because of that, facebook can then leverage them. Facebook can say, well, we will downrank this if a third-party verified fact checker has fact checked ...

Okay. And so one thing that’s come from our interviews with fact-checking organizations and civil society organizations working in these spaces here is that they sometimes seem to have the legitimacy to do things that platforms can’t.

I mean, in the US context. Right. There are currently active subpoenas on technology companies for making decisions about political information that is seen as polarized. Right?

But. In the misinformation space it’s not contested. So they can work in that space, put all their resources towards it, no problem. But in political polarization, they don’t have the legitimacy, right, or they’re maybe not the right organization.

And then they’ve hired this set of people focused on this work and so now they have those two bits of infrastructure that other efforts can draw on.

So it’s not just that they now have a strong incentive to mitigate these scams, but they’ve now built in a cleaner reporting system or really strong incentive to improve their reporting system.

May I riff on that a little bit? Okay, so you have this clear problem space that all sides can agree on that’s not contested in the way that others are. And so you work in that space and then you use the power of economic incentives, right, to encourage ...

Right.

Right.

Yeah, that’s a really crisp explanation. Do you have a question? I have many, many questions, but I wanted to give both of you space.

Of the statistical distribution of the words.

Can you say more?

But it feels like there’s really a lot of opportunity in making the rest of the ecosystem more robust. Rather than focusing on what’s going on with ai, focusing on everything else.

So this is really interesting to me because it feels like many of the strategies that I’m seeing talked about publicly are really around the content. They’re around using ai to track ai, watermarking, these sorts of things.

For example, which creates its own set of problems, but yeah. So you’re using context cues to say, okay, well, you can’t verify the content, but you can trust the actor, the content.

Right.

Right.

Uh oh!

Sure.

So I think what you’ve just done is put those two facts together. And so can you just give us a little bit more of your mental landscape or framework for thinking about how to think about generative AI in this problem space, or how you’re thinking about generative AI ...

Okay, so one of the reasons that we’re doing this study, right, is because in 2023, the big thing, among others that happened in 2023, is that generative AI started to pass theory of mind tests, right, among many other tests. And you see between gpt 3.5 and gpt four, ...

Fascinating.

It’s a DDoS attack on the information environment.