00:00There were two studies that just came out recently and both basically showed that chatbots are
00:05incredibly persuasive. And it's not just AI that generates a 200 word blurb about a certain
00:12political candidate. It's when people get into a conversation with an AI system that their views
00:18can actually be quite susceptible to being changed on all number of policy issues. And so that was
00:26the first finding that these things are much more persuasive than traditional TV ads or political
00:32campaign ads. And secondly, when they use the most persuasive tactics, which we can talk about what
00:37those tactics are, they're more likely to actually hallucinate and give incorrect information,
00:44mistruths. So what are those tactics? Well, the tactic is basically to bombard a person with as
00:50much information as possible. So in a 10 minute conversation with a person, if a chatbot gets in
00:5825 different facts, that person is much more likely to sort of start to sway their view than if it's
01:0610 different points that they're making in the conversation. And actually, this kind of, it might
01:11not be surprising to people because a very popular debating tactic now, particularly on YouTube,
01:16is this rapid fire just kind of flooding the zone with all sorts of facts and making it quite
01:24difficult for your opponent, if you're in a debating situation, to kind of unpick and analyze what you're
01:29saying. So this is, this actually has a term, gish galopping, comes from the creationist Duane
01:35Gish. And, but it essentially means just bombarding a debate with lots of facts. And that actually really
01:42works when it's AI doing it. So in this age of social media and these rapid fire kind of debates,
01:48will AI, you know, automatically make things worse? It could. We don't know yet. The good news
01:54is that this is, this is something that was shown in a study where people were paid to participate.
02:01But actually getting someone to talk to a, you could call it a propaganda bot, is quite difficult.
02:07Yes, many people are using ChatGPT, about 900 million a week. But it's not ChatGPT that is
02:13doing this. It would be, let's say, a campaign group. They buy access to the model, then they
02:19fine tune it, and then they put it on a website. And somehow you have to attract people to come and
02:24talk to it like a customer service bot. So the good news is it's actually, might be quite difficult
02:29to get people to talk to these AIs. But another tactic might be to just DM them on WhatsApp or
02:36on Facebook Messenger as a bot. And then, you know, if you're a campaign group or whatever,
02:42an activist group, that might be one way to use that persuasive tactic.
02:45Are bad actors already using this technique?
02:48I don't know. I haven't seen any evidence that they are. But, you know, there's a lot of open
02:54source AI models out there, which make it, you know, relatively cheap for anyone to do this. And
03:01this is why the scientists at Oxford University and the London School of Economics and Science and
03:06Nature journals published this report just to kind of show how easy and cheap it is. We're talking,
03:12you know, a few tens of thousands of dollars to actually set up an operation like this,
03:16particularly with open source models like Lama or Quen from Alibaba, which are very easy to kind
03:22of make bespoke and build this kind of ultra-persuasive tool. It's out there. It's available.
03:28So at the moment, there's no regulation, right? There's no, so you're kind of relies, you know,
03:33I guess relying on the common sense of people to understand who they're interacting with.
03:37Yes. But, you know, you can arm people with more common sense by sort of telling them these bots
03:42are out there and just to be aware of them. Yeah, for like a specific reason.
03:46Perhaps. Yes, perhaps.
03:47Yes.
03:48No, because of course.
03:50I think so.
03:50Yeah, I'm saying that's still moving fast to me.
03:52It's going to happen.
03:53Hold right, I'm saying that's lasting things.
03:54No, I think that's lost everything.
03:55From answer, if you like
03:59any other questions?
04:00Maybe one part can remember people would like me right
04:01in the chat?
04:02Yeah.
04:03Bye!
04:04Bye-bye!
04:05Alright.
04:05Bye-bye.
04:07Bye-bye!
04:07Bye-bye!
04:08Bye-bye!
04:10naires!
04:11Bye-bye!
04:13Bye-bye!
04:13Bye-bye!
04:13Bye-bye!
04:14Bye-bye!
04:15Bye-bye!
Comments