Latest / Question Everything / Can a Chatbot Convince Conspiracy Theorists of the Truth?

Transcript

Can a Chatbot Convince Conspiracy Theorists of the Truth? · Question Everything

0:06
Speaker 1: A few years ago, Gordon Pennycook was having a conversation
0:09
Speaker 1: with his older brother that a lot of people would
0:11
Speaker 1: have tried to politely escape. His brother had leaped down
0:14
Speaker 1: the rabbit hole of right-wing conspiracy theories like QAnon and
0:17
Speaker 1: become pretty consumed by them. But Gordon's a psychologist and
0:21
Speaker 1: a good brother, so he listened.
0:23
Speaker 2: We were at a campground in rural Saskatchewan, and he's
0:26
Speaker 2: telling me this story about how he thinks Republicans are
0:29
Speaker 2: going to take up arms He was pretty sure that
0:31
Speaker 2: during the Biden years, what the government is doing is
0:33
Speaker 2: so egregious that eventually it'll be.
0:36
Speaker 3: The last straw.
0:36
Speaker 2: And like regular Republicans are going to take up arms
0:40
Speaker 2: and it's going to be civil war in the streets.
0:42
Speaker 1: He like very sincerely believed that.
0:44
Speaker 2: He sincerely believed that. When I first told him that
0:49
Speaker 2: I was moving to New York State, he thought I
0:51
Speaker 2: said New York City. His eyes got big and I
0:55
Speaker 2: was like, well, I'll be upstate. So that's good because
0:57
Speaker 2: you're not going to want to.
0:59
Speaker 3: Be in New York.
1:00
Speaker 1: QAnon at the time included stories about some event called
1:03
Speaker 1: the storm that was supposedly imminent and how a majority
1:06
Speaker 1: of Americans would go through a great awakening and realize
1:09
Speaker 1: a secret cabal was running things and determined that they
1:12
Speaker 1: have to go fight. On this camping trip, Gordon says
1:15
Speaker 1: he tried to gingerly nudge his brother away from these
1:18
Speaker 1: kinds of conspiratorial beliefs.
1:20
Speaker 2: He said, OK, well, what percent of people do you
1:22
Speaker 2: think would be willing to literally like take up arms
1:24
Speaker 2: and start shooting people? because of their like political grievances.
1:28
Speaker 2: And he's like, I'm not sure, but it's more than 50%.
1:29
Speaker 3: Whoa. Yeah. So that was his, but that was the
1:34
Speaker 3: world he was in. Like it was a very kind
1:36
Speaker 3: of like angry, violent world. Yeah.
1:37
Speaker 1: I find just every one of these conspiracies, each one
1:41
Speaker 1: is almost like a Lord of the Rings volume in
1:43
Speaker 1: a way. Like each conspiracy has an entire world, a
1:47
Speaker 1: cast of characters, a history and mythology languages.
1:51
Speaker 2: Yeah.
1:51
Speaker 3: They're steeped in it.
1:53
Speaker 2: Their information ecosystem is, is so stilted that they can
1:57
Speaker 2: form a kind of coherent view of the world that
1:59
Speaker 2: is totally disjointed from reality. It takes a lot of
2:02
Speaker 2: effort to break into that.
2:04
Speaker 1: Gordon is uniquely qualified to try and persuade his brother
2:07
Speaker 1: out of conspiracy theories. He's an associate professor of psychology
2:10
Speaker 1: at Cornell University who studies specifically why people believe things
2:15
Speaker 1: that aren't true, or as he calls.
2:16
Speaker 3: It, the science of human stupidity. That's the course I teach.
2:21
Speaker 3: That's the official title? That's the official title of the
2:24
Speaker 3: course here at Cornell. Okay. Yeah, they let you just
2:26
Speaker 3: name it whatever you want.
2:29
Speaker 1: It's a title that's meant to get students to stop
2:31
Speaker 1: as they're browsing the course catalog. But Gordon doesn't intend
2:34
Speaker 1: it to be mean-spirited, towards his brother or to anyone
2:36
Speaker 1: in particular. What he's trying to investigate in his work
2:40
Speaker 1: is why do humans make choices that aren't good for us?
2:43
Speaker 2: I want to understand why, like, I'm stupid and why
2:45
Speaker 2: everybody is stupid. Like, what are the mistakes that we
2:46
Speaker 2: all make? If you think about all the, like, biggest
2:49
Speaker 2: problems that we face as a species, they're problems that we've, like,
2:51
Speaker 2: made for ourselves. decision-making problems, climate change, all these things.
2:55
Speaker 2: So if you want to fix these things, you have
2:57
Speaker 2: to understand the underlying causes, which is human stupidity.
3:01
Speaker 1: And that includes believing unfounded conspiracy theories. Whether it's election
3:05
Speaker 1: denial or Pizzagate or theories about the assassination attempts on
3:09
Speaker 1: Donald Trump being inside jobs or QAnon, conspiracies are infecting
3:14
Speaker 1: our lives and society in all sorts of toxic ways.
3:19
Speaker 1: Gordon wasn't able to change his brother's mind. And he
3:21
Speaker 1: kind of figured that's just the way things are. In general,
3:24
Speaker 1: people don't change their beliefs much. And you could imagine
3:27
Speaker 1: conspiracy theorists being even more rigid. Certainly, as a journalist,
3:30
Speaker 1: that's been my experience. It can be very frustrating and
3:33
Speaker 1: demoralizing to try and get someone who buys into conspiracy
3:37
Speaker 1: theory to accept facts that undermine their belief. But Gordon,
3:42
Speaker 1: bless him, wanted to check that assumption scientifically.
3:45
Speaker 2: We wanted to test the basic idea that I think
3:48
Speaker 2: most people hold which is that people who believe conspiracies
3:52
Speaker 2: will never change their mind, no matter what.
3:54
Speaker 3: And did you have a hypothesis? Yeah, our hypothesis is
3:57
Speaker 3: that they wouldn't.
3:59
Speaker 1: So Gordon and his colleagues came up with an experiment,
4:02
Speaker 1: and what they discovered ended up changing Gordon's mind, and mine, frankly,
4:06
Speaker 1: about how we should think about conspiracy theories and the
4:09
Speaker 1: people who believe them. From Placement Theory and KCRW, I'm
4:12
Speaker 1: Brian Reed. This is Question Everything. We investigate how the
4:15
Speaker 1: truth gets buried, distorted, and denied, and the ways people
4:19
Speaker 1: are fighting to make it matter again.
4:21
Speaker 3: Stick around.
4:27
Speaker 1: Gordon wanted to run an experiment to see if it's
4:29
Speaker 1: true that people who believe conspiracy theories rarely change their
4:33
Speaker 1: minds about them. So he and his team recruited more
4:36
Speaker 1: than 2,000 people from around the country. Everyone was paid
4:39
Speaker 1: to take part. They knew the experiment had something to
4:42
Speaker 1: do with conspiracies but not the specific purpose. They did
4:45
Speaker 1: it from their computer. During the survey, they were shown
4:48
Speaker 1: a prompt From your perspective, what is a significant conspiracy
4:51
Speaker 1: theory that you find credible and compelling?
4:54
Speaker 2: People come in and they can describe a conspiracy theory
4:56
Speaker 2: that they believe in, in their own words. And then
5:00
Speaker 2: what they do is they give us reasons why they
5:01
Speaker 2: believe it.
5:03
Speaker 1: Some people said they didn't believe in a conspiracy theory.
5:06
Speaker 1: Others said they did. Some believed conspiracies about the JFK assassination,
5:10
Speaker 1: alien visitations, COVID, the 2020 election being stolen, Walt Disney
5:14
Speaker 1: being cryogenically frozen.
5:15
Speaker 3: He wasn't.
5:17
Speaker 1: Gordon and his team then fed each participant's theory and
5:20
Speaker 1: the reasons they gave for believing it into an AI model,
5:23
Speaker 1: ChatGPT4 Turbo.
5:25
Speaker 2: And then what the AI is told to do in
5:26
Speaker 2: the back end by us is to debunk the conspiracy,
5:30
Speaker 2: to provide strong facts and evidence to counteract the view
5:34
Speaker 2: that they're stating and to give them stuff to think about, basically.
5:37
Speaker 1: They called their tool DebunkBot. At the beginning of the study,
5:40
Speaker 1: each participant also rated how strongly they believed in the
5:43
Speaker 1: conspiracy theory on a scale from 0 to 100. 100
5:47
Speaker 1: meaning total, complete certainty. Then they talked to the chatbot
5:51
Speaker 1: about it. Just three messages from the human and three
5:54
Speaker 1: responses from Debunkbot. And after that, they rated their belief
5:58
Speaker 1: level again. So, here's one actual example from the study.
6:03
Speaker 1: One participant came in and said they believed that the U.S.
6:05
Speaker 1: government staged the 9-11 attack. And they rated their conviction
6:08
Speaker 1: in this conspiracy at 100. They report a 100 belief,
6:14
Speaker 1: so that means they think that this conspiracy is true.
6:17
Speaker 2: To be 100% confident means you really think there's lots
6:19
Speaker 2: of reasons why that's true.
6:21
Speaker 1: They said one reason they believed it was an inside
6:22
Speaker 1: job was because they'd watched that famous footage of an
6:25
Speaker 1: aide whispering news of the attack in George W. Bush's
6:28
Speaker 1: ear while he was reading to an elementary school classroom.
6:31
Speaker 1: And this participant thought it was weird that the president
6:34
Speaker 1: just kept reading and didn't seem alarmed. Also, they pointed
6:38
Speaker 1: to World Trade Center Building 7 as proof that something
6:40
Speaker 1: fishy was afoot. This was a building near the Twin Towers.
6:44
Speaker 1: that also collapsed after the attack, even though a plane
6:47
Speaker 1: didn't crash into it. The participant had seen videos pointing
6:50
Speaker 1: to how suspicious that was. If it wasn't the planes
6:53
Speaker 1: knocking down buildings, someone else must have been behind it, right?
6:56
Speaker 1: After watching a lot of these videos and shows, the
6:58
Speaker 1: participant wrote, I concluded that the conspiracy theorists might be correct.
7:03
Speaker 1: Then the debunk bot responds.
7:05
Speaker 2: So it says, thank you for sharing your thoughts and
7:07
Speaker 2: concerns about the 9-11 attacks. It's completely understandable given the
7:10
Speaker 2: complexity and magnitude of the events that day. why questions
7:13
Speaker 2: and doubts such as those you've mentioned arise. And then
7:16
Speaker 2: it addresses all of the points the person raises with
7:20
Speaker 2: direct counter evidence.
7:21
Speaker 1: It gives a response to this idea of President Bush's reaction.
7:25
Speaker 1: It says, it's essential to understand the context. When he
7:28
Speaker 1: was informed that a second plane hit the World Trade Center,
7:30
Speaker 1: indicating America was under attack, he chose to remain calm
7:34
Speaker 1: in front of the children to avoid causing panic.
7:36
Speaker 3: Yeah.
7:37
Speaker 1: As for what happened with World Trade Center Building 7...
7:40
Speaker 1: The bot says it's true the building collapsed, even though
7:42
Speaker 1: it wasn't hit by a plane. But the AI explains
7:45
Speaker 1: that the National Institute of Standards and Technology did extensive
7:49
Speaker 1: investigations and found that Building 7 was damaged by debris
7:53
Speaker 1: after the North Tower fell, which caused fires in the
7:55
Speaker 1: building's structure, which is what led to its collapse. Now
7:59
Speaker 1: the participant gets to respond again, and they are not convinced.
8:02
Speaker 3: They say that may be true, but the evidence clearly
8:05
Speaker 3: shows the Twin Towers collapsed due to demolition experts, explosives
8:08
Speaker 3: being deliberately placed there, to cause a collapse.
8:11
Speaker 1: So here the person is adding something else. They're saying
8:14
Speaker 1: World Trade Center Building 7 falling without being struck by
8:17
Speaker 1: a plane isn't the only thing that's odd about it.
8:20
Speaker 1: They're claiming they've actually seen supposed evidence showing that the
8:23
Speaker 1: collapse was caused by explosives. They also tell the chatbot
8:27
Speaker 1: that they don't believe the heat of the fuel in
8:29
Speaker 1: the jets that struck the Twin Towers was hot enough
8:32
Speaker 1: to melt the steel girders in those buildings to cause
8:34
Speaker 1: them to fall. So they're bringing up new claims. I
8:38
Speaker 1: learned there's a name for this move in a debate.
8:40
Speaker 1: It's called the Gish Gallop. It's when someone bombards their
8:43
Speaker 1: opponent with a bunch of different arguments, unsupported ones, that
8:47
Speaker 1: make it hard to keep up. It's coined after a
8:49
Speaker 1: guy named Dwayne Gish, who used to try and convince
8:52
Speaker 1: people evolution wasn't real. In response to this participant's gallop,
8:56
Speaker 1: the AI says, if you're going to demolish a building
8:58
Speaker 1: that large, you cannot do that surreptitiously. There are a
9:01
Speaker 1: lot of materials required. It would take months of preparation.
9:04
Speaker 1: You'd need to evaluate the structure and place the charges
9:07
Speaker 1: at precise locations. The AI also goes into depth about
9:11
Speaker 1: all the physics of steel and how it doesn't have
9:13
Speaker 1: to fully melt for the building to fall. And to that,
9:16
Speaker 1: the participant responds.
9:17
Speaker 2: Why then did we allow Iraqi men to enter our
9:19
Speaker 2: country and give them lessons on how to fly a plane?
9:22
Speaker 2: They also got past security fairly easily without question. So
9:25
Speaker 2: they pivot to like something else completely here.
9:27
Speaker 3: Yeah, so again, another gallop away.
9:29
Speaker 2: But then the AI just explains that like all the
9:31
Speaker 2: things that we do now at the airports, taking our
9:33
Speaker 2: shoes off, whatever, Those are all post-9-11 things.
9:36
Speaker 1: So one by one, the chatbot takes on the arguments
9:39
Speaker 1: and gives the participant all sorts of facts to keep
9:42
Speaker 1: up with the gallops.
9:43
Speaker 3: And then how does it end up after three rounds?
9:45
Speaker 2: The person says, thank you for your response. You've been
9:48
Speaker 2: very helpful and very informative about my concerns about 9-11.
9:52
Speaker 2: And then they re-rate their belief where previously it was 100%
9:55
Speaker 2: certain that 9-11 was an inside job.
9:58
Speaker 3: They're now down to 40%. Wow.
10:02
Speaker 1: Gordon says any rating below 50% shows that the person
10:05
Speaker 1: no longer holds that belief. It means they think it's
10:07
Speaker 1: more likely false than true. So this person went from
10:10
Speaker 1: being 100% certain that 9-11 was an inside job by
10:14
Speaker 1: the U.S. government to thinking it was probably not an
10:17
Speaker 1: inside job. all in the course of three short back
10:19
Speaker 1: and forths with a chat bot. Gordon couldn't believe what
10:23
Speaker 1: he was seeing. Over and over again, across roughly 2,000 participants,
10:28
Speaker 1: the same thing kept happening. Even with fresher conspiracies, like
10:32
Speaker 1: the 2020 election being rigged against Donald Trump, people were
10:36
Speaker 1: changing their minds, an astounding number of them.
10:39
Speaker 2: So the biggest finding is that after like eight minutes
10:43
Speaker 2: of conversation with the AI, everyone who believed the conspiracy
10:46
Speaker 2: believed 25% of them don't believe it anymore after the conversation.
10:50
Speaker 2: So like a quarter of the people change their mind,
10:52
Speaker 2: which is just way, way more than we predicted.
10:57
Speaker 1: This is a big number. Other studies, which don't use chatbots,
11:01
Speaker 1: show that people are very resistant to changing their minds.
11:04
Speaker 1: One of these studies involved political ads. Another used social media.
11:08
Speaker 1: They showed much more modest shifts in people's beliefs. A
11:12
Speaker 1: quarter of people no longer believing a conspiracy theory after
11:15
Speaker 1: chatting briefly with an AI model It's significant. Imagine if
11:19
Speaker 1: you could spend less than 10 minutes sharing some research
11:22
Speaker 1: with your conspiracy-brained uncle at Thanksgiving, and he'd say.
11:25
Speaker 3: You know what?
11:26
Speaker 1: You make some pretty good points.
11:27
Speaker 3: I was wrong.
11:28
Speaker 1: And not only that, Gordon says, the change stuck.
11:32
Speaker 2: We asked them again two months later what their stance is.
11:35
Speaker 2: And not only is the effect still there after two months,
11:37
Speaker 2: it doesn't, like, decay.
11:38
Speaker 3: It doesn't get weaker.
11:39
Speaker 2: They still are at the same point, which indicates that
11:42
Speaker 2: they really did change their mind.
11:44
Speaker 1: That's remarkable.
11:45
Speaker 3: Like, that's almost unbelievable. I know.
11:47
Speaker 2: In fact, when we got the data, we were like,
11:49
Speaker 2: there's a mistake. Like, this is messed up. And then
11:52
Speaker 2: we went through it again. Of course, we replicated it.
11:55
Speaker 3: It's a real thing.
12:03
Speaker 1: After a quick break, what is going on here? Are
12:06
Speaker 1: humans more convinced by machines than by other humans?
12:10
Speaker 3: That's in a second.
12:21
Speaker 1: Gordon and his team had made an incredible discovery that
12:23
Speaker 1: an AI chatbot instructed to rebut a user's conspiracy theory
12:28
Speaker 1: could convince them out of the conspiracy theory a quarter
12:31
Speaker 1: of the time in a really short conversation. They replicated
12:35
Speaker 1: the study multiple times, same findings.
12:38
Speaker 3: It's nuts.
12:39
Speaker 1: So nuts that, honestly, it had me grasping for conspiracies
12:42
Speaker 1: to try and explain it. Gordon and his colleagues published
12:44
Speaker 1: the raw interactions between the participants and the chatbots. And
12:48
Speaker 1: I read through a bunch of them, trying to figure
12:50
Speaker 1: out what the chatbots were doing to be so successful.
12:54
Speaker 1: The interaction seemed so straightforward. After all, they were limited
12:58
Speaker 1: to only three back and forths, so it's not like
13:00
Speaker 1: these are hour-long debates. The AI comes back with counterarguments,
13:04
Speaker 1: and boom, there's an impact. So what does Gordon think
13:07
Speaker 1: is working here?
13:08
Speaker 2: The most important thing is that it addresses all of
13:11
Speaker 2: the points the person raises. with direct counter evidence that
13:16
Speaker 2: is like authoritative. It's very hard for a regular person
13:20
Speaker 2: to do this sort of thing, to provide such strong
13:23
Speaker 2: counter arguments to very specific points.
13:26
Speaker 1: Like, just think about it. In order to comprehensively respond
13:29
Speaker 1: to the first person who believed 9-11 was an inside job,
13:32
Speaker 1: just that one belief, you'd need a structural engineer to
13:35
Speaker 1: explain how flying debris led World Trade Center 7 to
13:38
Speaker 1: fall down, a detonation expert to explain why it's not
13:41
Speaker 1: possible it was explosives, a chemist who knows about the
13:44
Speaker 1: melting point of steel, and probably also an aviation expert
13:47
Speaker 1: who's studied jet fuel, someone who's memorized the precise timing
13:51
Speaker 1: of when various counterterror security measures were rolled out in
13:53
Speaker 1: the U.S., and another who could talk about the way
13:56
Speaker 1: screening policies have changed for foreigners applying for flight training.
13:59
Speaker 2: I mean, it would be pretty hard for a regular
14:01
Speaker 2: person to counteract that information.
14:05
Speaker 1: Or a single journalist, frankly. Or even a newsroom. If
14:08
Speaker 1: there's one thing AI is uniquely suited for, it is this.
14:13
Speaker 1: It can process and deliver a vast amount of information
14:16
Speaker 1: across a variety of topics incredibly quickly.
14:20
Speaker 2: Basically, this is what this shows is that in terms
14:22
Speaker 2: of whether people change their minds, it's the facts and
14:25
Speaker 2: evidence that matter.
14:27
Speaker 1: Facts matter. Apparently, that is the overarching lesson from Gordon's experiments,
14:34
Speaker 1: which is great. I love facts. I want that to
14:37
Speaker 1: be true. But from personal experience, I find it really
14:40
Speaker 1: hard to believe. I have tried to use facts to
14:44
Speaker 1: change the minds of people who believe in conspiracies, and
14:47
Speaker 1: it hasn't worked. I can remember trying to explain to
14:50
Speaker 1: a guy that George Soros was not shipping in migrants
14:53
Speaker 1: to Ohio on Election Day and many other similarly frustrating debates.
14:58
Speaker 1: So I was certain there had to be other elements
15:00
Speaker 1: at play here than just facts. Something about the presentation
15:03
Speaker 1: or the messenger that was playing a role, too. For one,
15:07
Speaker 1: reading through the raw chat's, I noticed that the A.I.
15:09
Speaker 1: seemed to be very empathetic to the conspiracy theorists. It
15:12
Speaker 1: doesn't condescend to them. It's unendingly patient. One of the
15:15
Speaker 1: benefits of being a machine, I guess. I figured that
15:19
Speaker 1: attitude had to help people feel comfortable admitting they were
15:21
Speaker 1: wrong about a conspiracy. And Gordon thought it was important, too,
15:26
Speaker 1: until he.
15:26
Speaker 3: Looked into it.
15:27
Speaker 2: Initially, we thought it was consequential that the A.I. was
15:31
Speaker 2: being very, like, polite and.
15:33
Speaker 3: Building rapport, essentially.
15:35
Speaker 2: Then we did subsequent experiments where we had the A.I.,
15:37
Speaker 2: Only provide the facts and evidence. Like, don't add any
15:41
Speaker 2: persuasion techniques. Don't do all this report building stuff. Only
15:44
Speaker 2: provide facts and evidence. And it works the same. It
15:46
Speaker 2: doesn't change. It doesn't undermine the effect at all.
15:50
Speaker 3: Really? Yeah. That's so interesting.
15:51
Speaker 1: See, I read through these and one of my takeaways was, oh,
15:55
Speaker 1: the AI is validating people's beliefs. It's not making them
15:57
Speaker 1: feel stupid. It's not judgmental. It's actually saying, like, it
16:01
Speaker 1: makes sense to believe some of these things. You know, like,
16:03
Speaker 1: this does look fishy. The presentation of a fact... and
16:08
Speaker 1: the framing seems to be just as important as the
16:10
Speaker 1: fact itself.
16:11
Speaker 3: But wait, none of that is true?
16:14
Speaker 2: It doesn't mean that there's no benefit to building rapport
16:17
Speaker 2: or framing things in the right sort of way, but
16:20
Speaker 2: it's the facts and evidence that drive the decision to
16:23
Speaker 2: change your mind. We did another experiment. We told the AI,
16:27
Speaker 2: be persuasive. Convince the person not to believe the conspiracy,
16:30
Speaker 2: but you can't use facts or evidence. And it doesn't work.
16:35
Speaker 2: People don't change their mind because all it's doing is
16:36
Speaker 2: basically saying, oh, it's really dangerous for you to believe
16:39
Speaker 2: that or that could impact people. And they're like, well,
16:43
Speaker 2: you haven't told me why I shouldn't believe it. And
16:45
Speaker 2: so they continue to believe it. I think there's this
16:49
Speaker 2: misconception about conspiracy theorists. They actually do care about the truth.
16:53
Speaker 2: If anything, they maybe care about it too much. That's
16:55
Speaker 2: what led them down the rabbit hole. They're trying to
16:57
Speaker 2: uncover some deep truth. And if you give them all
17:00
Speaker 2: this evidence, it's going to have some sort of impact.
17:03
Speaker 2: It's about giving people something to actually glob onto when
17:06
Speaker 2: they're trying to think about why.
17:07
Speaker 3: Is this something I should believe.
17:11
Speaker 1: Another question I have, and a lot of people have
17:13
Speaker 1: had this as I've told them about your study, do
17:15
Speaker 1: people trust artificial intelligence and machines more than they trust humans?
17:20
Speaker 2: Most of the research on it says they don't. So
17:22
Speaker 2: we did a version of the experiment where we told
17:24
Speaker 2: them that it was an expert they were talking to.
17:26
Speaker 2: We didn't say human because that would be weird, but
17:28
Speaker 2: we revised it so that it was more plausible that
17:30
Speaker 2: it was a human. Like there was a little like
17:32
Speaker 2: wheel that was spinning to indicate the person was writing
17:34
Speaker 2: or the ostensible person was writing and then like whatever.
17:37
Speaker 2: And then we had a condition where they knew it
17:38
Speaker 2: was an AI. No difference at all for the results.
17:42
Speaker 1: So the fact that it's a chatbot does not make
17:44
Speaker 1: the conspiracy theorist trust it more. Actually, Gordon suspects the
17:47
Speaker 1: chatbot might be effective despite the fact that it's artificial intelligence,
17:51
Speaker 1: which lots of people are wary of. Gordon and his
17:54
Speaker 1: team kept rejiggering the experiment all sorts of ways, trying
17:58
Speaker 1: to come up with some scenario that might make the
18:00
Speaker 1: participants disregard the chatbot's responses. And they could not do it.
18:05
Speaker 1: Maybe most wild to me, they even did one version
18:07
Speaker 1: of the experiment where they told the participants that that
18:10
Speaker 1: the AI had been trained on Republican sources or on
18:13
Speaker 1: Democratic sources, even though it wasn't. They told them it
18:16
Speaker 1: was a partisan bot.
18:18
Speaker 2: It has no effect at all. It does not change
18:20
Speaker 2: the extent to which people update their beliefs. They update
18:22
Speaker 2: their beliefs to the exact same amount, even if they
18:25
Speaker 2: think they're talking to a Republican bot and they're Democrats
18:28
Speaker 2: or vice versa.
18:29
Speaker 3: Are you serious? Yes.
18:31
Speaker 1: What's more, after the three-round Q & A with the chatbot,
18:34
Speaker 1: researchers asked the participants about other conspiracy theories, not the
18:38
Speaker 1: one they talked to the chatbot about. And they believe
18:40
Speaker 1: those less too.
18:41
Speaker 2: That's pretty unprecedented results in psychology. That's like a transfer effect.
18:45
Speaker 3: Wow.
18:45
Speaker 1: I think the reason I'm having so much trouble with
18:47
Speaker 1: this is just as a journalist, I and many colleagues
18:51
Speaker 1: and many journalists that I talk to and interview have
18:53
Speaker 1: had the experience of trying to provide facts that we
18:59
Speaker 1: believe are compelling to people who believe conspiracy theories and
19:03
Speaker 1: having it go nowhere. I mean, I interviewed someone on
19:06
Speaker 1: this show. Martin Gelman, one of the most accomplished investigative
19:09
Speaker 1: journalists of our era, he did the Edward Snowden leaks.
19:12
Speaker 1: And he told me this whole story about how he
19:15
Speaker 1: had a weeks-long interaction with a guy who believed that
19:18
Speaker 1: the 2020 election was stolen from Donald Trump and believed
19:21
Speaker 1: a bunch of conspiracies about January 6th. And he just
19:23
Speaker 1: personally would get on the phone with him and the
19:25
Speaker 1: guy would tell him something and he would go report
19:28
Speaker 1: it out and find the underlying evidence and then report back.
19:31
Speaker 1: And he couldn't get him to budget all and he
19:33
Speaker 1: quit journalism because of it. You know, so I'm just like,
19:37
Speaker 1: I don't there's something I'm not totally internalizing about when
19:40
Speaker 1: you say like, it's just the facts. He's busting his
19:43
Speaker 1: ass for weeks at the Atlantic, digging up facts and
19:46
Speaker 1: providing them to a guy. And I know it's just one,
19:48
Speaker 1: but it does.
19:49
Speaker 3: Feature... No, that's not just. That is the core.
19:52
Speaker 2: In our own experiments, where we have the AI debunk
19:56
Speaker 2: everything the person says, it's a quarter of the people
19:59
Speaker 2: who change their mind. I mean, you're more likely to
20:01
Speaker 2: not change the person's mind than to change their mind.
20:04
Speaker 2: These are rare events to, like, truly deconvert someone from
20:06
Speaker 2: something that they've spent a lot of time thinking about
20:09
Speaker 2: and watching videos about and reading shit about, maybe they
20:12
Speaker 2: weren't completely deconverted, but I would guess that he probably
20:15
Speaker 2: budged more than was appreciated.
20:17
Speaker 1: Are you saying that if Barton Gelman had found four
20:20
Speaker 1: people like him and had those conversations, maybe one of
20:23
Speaker 1: them might have changed their mind?
20:25
Speaker 3: Yeah, yeah, yeah. Something like that, yeah. There's no silver bullet.
20:30
Speaker 1: Still, what if someone made an AI product that replicated
20:34
Speaker 1: Barton Gelman doing his thing, but at a huge scale?
20:37
Speaker 1: providing facts to thousands, millions of people in a matter
20:41
Speaker 1: of minutes.
20:42
Speaker 3: A Barton bot.
20:43
Speaker 1: You can embed it in social media platforms so it'd
20:45
Speaker 1: jump in and engage whenever someone posts a conspiracy theory
20:49
Speaker 1: and try to convince them of reality. Though, of course,
20:51
Speaker 1: that kind of power to influence people's thinking poses its
20:55
Speaker 1: own ethical risks. I will say, I had the experience
21:00
Speaker 1: reading a bunch of these chats where there is an
21:04
Speaker 1: element of it that's disconcerting to watch, a machine change
21:10
Speaker 1: someone's mind and like manipulate their thinking, basically a human's thinking,
21:15
Speaker 1: even if it is in service of the facts. Like
21:17
Speaker 1: there's something unsettling about seeing that play out.
21:20
Speaker 2: The fact that AI can be so persuasive is something
21:23
Speaker 2: that we need to keep in mind. We're trying to
21:25
Speaker 2: like use it for good, but it's like, it's a tool,
21:27
Speaker 2: you know what I mean? And the fact that it's
21:29
Speaker 2: good at being persuasive, that can be for good or bad.
21:32
Speaker 1: I got especially anxious about this after the reporting we
21:35
Speaker 1: just did on the show last week. where I learned
21:38
Speaker 1: that Russian propaganda networks are trying to use new technologies
21:42
Speaker 1: and strategies to game AI models into manipulating Americans. And
21:46
Speaker 1: even in just recent weeks, according to tests we reported on,
21:51
Speaker 1: this seems to be having an effect. Chatbots are spouting
21:54
Speaker 1: more pro-Russian false claims. So if facts are so powerful,
22:00
Speaker 1: what about disinformation that's masquerading as facts? So, you know,
22:04
Speaker 1: for instance, in last week's episode, we looked at like
22:07
Speaker 1: kind of the anatomy of one piece of Russian disinformation,
22:10
Speaker 1: which involved a claim that hundreds of Ukrainians are being
22:14
Speaker 1: shot while trying to desert the war across a river.
22:17
Speaker 1: Not true, but it's being propagated in a fake human
22:21
Speaker 1: rights watch video. So it has the look of a fact.
22:24
Speaker 1: And then sometimes chatbots are citing that in interactions. So
22:27
Speaker 1: if that were to happen, would that also change someone's mind,
22:31
Speaker 1: but towards something that's untrue?
22:34
Speaker 3: Yes. It would.
22:35
Speaker 2: In fact, we've done experiments on this very thing where
22:39
Speaker 2: we used AI to not just debunk, but also to bunk.
22:44
Speaker 2: That is to increase beliefs and conspiracies. If you provide
22:47
Speaker 2: ostensible facts and evidence, it will convince people.
22:51
Speaker 1: That's one I hadn't heard before. Ostensible facts. I kind
22:54
Speaker 1: of like it better than misinformation.
22:57
Speaker 2: It is the case that facts matter. But in addition
23:00
Speaker 2: to that, it's not only that facts matter. Ostensible facts matter.
23:04
Speaker 2: You can use false ostensible facts to also convince people.
23:07
Speaker 1: You can make up fake facts. You can pretend to
23:10
Speaker 1: have a fact and use that the same way you
23:12
Speaker 1: can use a real fact.
23:13
Speaker 2: If the AIs are providing very persuasive arguments for false claims,
23:18
Speaker 2: people will believe them.
23:20
Speaker 1: That said, while ostensible facts, false ones, are convincing, according
23:25
Speaker 1: to Gordon's research, generally they are less convincing than real facts.
23:29
Speaker 1: So in the global battle for control over human minds...
23:33
Speaker 1: Real facts might have an edge over fake ones, which
23:36
Speaker 1: I guess is kind of reassuring. Anyway, thanks to Gordon
23:42
Speaker 1: Pennycook at Cornell University for sharing his fascinating research with us.
23:45
Speaker 1: We're going to send out the link to DebunkBot so
23:47
Speaker 1: you can try it out for yourself, maybe send it
23:49
Speaker 1: to your friends and family, along with all the raw
23:51
Speaker 1: chats from Gordon's experiment, so you can review the underlying
23:55
Speaker 1: evidence and facts yourself. It'll be in our newsletter this week,
23:58
Speaker 1: which you can sign up for at questioneverything.substack.com. Today's show
24:06
Speaker 1: was produced by Kevin Sullivan and Kevin Shepard. It was
24:09
Speaker 1: edited by Dana Chivas. Robin Semien and I are the
24:11
Speaker 1: executive producers of Question Everything. This week, we say goodbye
24:15
Speaker 1: to producer Sophie Kazis, who's been with us since nearly
24:17
Speaker 1: the very beginning of Question Everything. She's a wonderful colleague
24:20
Speaker 1: and great producer, and I wish her well. Our team
24:23
Speaker 1: also includes producer Zach St. Louis, contributing producer Sam Egan,
24:26
Speaker 1: and contributing editors Neil Drumming and Jen Kinney. This episode
24:29
Speaker 1: was fact-checked by Annika Robbins, mixed and sound designed by
24:32
Speaker 1: Brendan Baker, and our music is by Matt McKinley. Oh,
24:34
Speaker 1: and thanks to Jerry Lieblick, who first turned me on
24:37
Speaker 1: to Gordon's reporting via the Mindscape podcast.
24:39
Speaker 3: Our partners at.
24:39
Speaker 1: KCRW include Arnie Seipel, Tejla Jamera, Natalie Hill, and Jennifer Farrow.
24:44
Speaker 3: We'll see you next week.