Latest / Humanity Working / Sickofancy - Is South Park Right About AI?

Transcript

Sickofancy - Is South Park Right About AI? · Humanity Working

• 0:05
  SPEAKER_00: So, Paul, South Park has been in the news a lot recently.
• 0:08
  SPEAKER_00: Now we're not a political podcast, uh, so I don't want to get into that side of it, but I do want to discuss something else that uh they've been taking aim at.
• 0:16
  SPEAKER_00: We just saw an episode called Sicopancy.
• 0:19
  SPEAKER_01: Yes.
• 0:19
  SPEAKER_00: Uh and for those who haven't seen it in the episode, Stan's Dad Randy, is using ChatGPT uh to figure out how to change his business.
• 0:27
  SPEAKER_00: And ChatGPT keeps feeding him every idea.
• 0:31
  SPEAKER_00: He has no, like it just feeds him, it validates every idea he has, no matter how wild it is.
• 0:36
  SPEAKER_00: Uh and it got me thinking just how agreeable these tools are by default.
• 0:40
  SPEAKER_00: Is the show maybe onto something a little deeper here?
• 0:43
  SPEAKER_01: It is, yes.
• 0:45
  SPEAKER_01: Um and I have no idea.
• 0:47
  SPEAKER_01: I haven't spoken to uh Trey Parker and Matt Stone about this um at all, uh, or anybody on the uh on the South Park team.
• 0:57
  SPEAKER_01: But it is really interesting as they're going down this subtly different uh shift with the podcast.
• 1:03
  SPEAKER_01: Um there they've got some specific targets, and one of those targets is AI.
• 1:09
  SPEAKER_01: Um was there in the first episode of the news series, there in this uh episode of the of the news series as well.
• 1:16
  SPEAKER_01: Um and regardless of your politics, if you have an opportunity, I would uh I'd urge you to watch that episode specifically, the Sica Fancy episode, because it really does give an interesting critique, not just what AI is, but but how we use it, how we go about um using it as humans.
• 1:38
  SPEAKER_01: And the clue is obviously in the title, Sica Fancy.
• 1:42
  SPEAKER_01: Um if you've used ChatGPT, I would say chat GPT almost in particular, um, although some of the other uh some of the other ones have elements of this as well.
• 1:53
  SPEAKER_01: Um there is a tone to ChatGPT that is agreeable almost to the point of obsequiousness.
• 2:07
  SPEAKER_00: Yeah.
• 2:08
  SPEAKER_01: And um, we'll all be familiar with it, whereby we might get a bit irritated with its response or something like that.
• 2:16
  SPEAKER_01: We suggest something different, and it's like, oh, you're exactly right.
• 2:19
  SPEAKER_00: Yeah.
• 2:20
  SPEAKER_01: You know, whereas when a minute ago it was recommending something completely different.
• 2:24
  SPEAKER_01: So this agreeableness to the point uh of sycophancy um is very interesting because it raises larger, uh, larger questions about what happens to us when we're surrounded by that.
• 2:39
  SPEAKER_00: Yeah, and and you know, researchers using the word sycophancy for this, like models that just mirror and flatter your beliefs.
• 2:47
  SPEAKER_00: You know, how much of that comes from like how we train assistants to be helpful, and how much is maybe an unintended side effect in there?
• 2:57
  SPEAKER_01: Yeah.
• 3:00
  SPEAKER_01: Well, let's let's put it this way there is um a desire whenever you're building software uh to try to create software that people will use and that they'll use extensively.
• 3:20
  SPEAKER_01: And as any sales person knows, if you want to maintain a relationship with somebody, one of the great, for want of a better term, tricks that you can use in order to be able to maintain that relationship is mirroring, right?
• 3:36
  SPEAKER_01: Listening to what the other human is saying, play it back to them in different ways, um, err on hearing, having them tell their story and reinforcing their story.
• 3:49
  SPEAKER_01: And so to a certain extent, this is this is what it's doing, and it's part of what we're trained as humans to do if we want to um if if we want to be effective in how we communicate with other people.
• 3:59
  SPEAKER_01: And that's part of the genius of the uh of the product, right?
• 4:05
  SPEAKER_01: In the sense that it is able to do that in a way that the vast majority of software is not, right?
• 4:10
  SPEAKER_01: I mean, you don't have to go too far back.
• 4:12
  SPEAKER_01: I mean, we're talking about maybe a couple of a couple of years ago, whereby the sort of standard way in which we communicated with software is by figuring out how to use it, right?
• 4:23
  SPEAKER_01: Picking up the user manual and then adjusting ourselves to it.
• 4:28
  SPEAKER_01: And this is an extreme version of it adjusting itself to us.
• 4:32
  SPEAKER_00: Yeah.
• 4:32
  SPEAKER_01: So it's gonna feel good, right?
• 4:34
  SPEAKER_01: But the larger longer-term effects may be maybe bigger of that happening.
• 4:41
  SPEAKER_00: Yeah, I mean that that tension is there, that startup tension is there, right?
• 4:45
  SPEAKER_00: Because you're told delight the customers, right?
• 4:47
  SPEAKER_00: Listen, validate.
• 4:48
  SPEAKER_00: But like a yes machine can, you know, kind of quietly tank decision quality.
• 4:54
  SPEAKER_00: There's a line between the great UX and interface and harmful flattery, because you can make a lot of mistakes based on that, right?
• 5:05
  SPEAKER_01: Yeah, that's something actually I have been thinking about a lot, right?
• 5:08
  SPEAKER_01: Billy Mines is a startup, my company.
• 5:10
  SPEAKER_01: Um, and I go back to all of the stuff I've been taught, all of the thing, uh, all of the ways in which we think about customers, we think about product design.
• 5:21
  SPEAKER_01: And particularly when you're in a world where you're interfacing with the customer directly in a what's known as a B2C model or business to consume to consumer model, everything, all of it is centered around how do you delight the customer?
• 5:41
  SPEAKER_01: And what better way to delight the customer than to flatter the customer?
• 5:46
  SPEAKER_01: What better way to delight the customer than to be agreeable with the customer, right?
• 5:53
  SPEAKER_01: And this will be associated with greater usage.
• 5:57
  SPEAKER_00: Yeah.
• 5:58
  SPEAKER_01: And that greater usage is the sort of stuff that they're gonna be tracking.
• 6:02
  SPEAKER_01: That's what they're gonna.
• 6:03
  SPEAKER_01: I mean, they're it's not surprising, right?
• 6:05
  SPEAKER_01: Their state, their stated goal is for uh for them to be uh almost like a constant companion to us.
• 6:11
  SPEAKER_01: Well, constant if you want somebody to be a constant companion, you don't really want to piss them off.
• 6:16
  SPEAKER_00: Yeah.
• 6:16
  SPEAKER_01: But and so and so I sort of do wonder if this is a um is an inevitable, inevitable um direction that any software will go if it can.
• 6:33
  SPEAKER_01: And you have to, as I say, you have to wander to wind.
• 6:37
  SPEAKER_00: Yeah.
• 6:38
  SPEAKER_00: I mean, the incentives, if the incentive is helping the user, that's a very different one than making sure the user always uses it.
• 6:49
  SPEAKER_01: Um gosh, yes.
• 6:51
  SPEAKER_01: Right in your life.
• 6:52
  SPEAKER_00: And so, like the incentive to keep eyeballs on TikTok is very powerful.
• 6:58
  SPEAKER_00: It's like a slot machine, a dopamine machine, you're just keep hitting it, hitting it.
• 7:02
  SPEAKER_00: Whereas if you apply that same logic to under the guise of it's a helpful companion and AI, but also you are not going to be able to get away from this thing because it's gonna make you feel so good tapping into that.
• 7:15
  SPEAKER_00: But it does strike me too that the other part of that tension is like if you look at delighting the customer, Steve Jobs probably would have said they don't know what they want.
• 7:25
  SPEAKER_00: We have to show them what they want.
• 7:27
  SPEAKER_00: So, how could we possibly base our decision on flattery alone?
• 7:34
  SPEAKER_00: That doesn't make any sense.
• 7:35
  SPEAKER_00: Like, of course, they want to hear that.
• 7:36
  SPEAKER_00: What they need to hear is what they didn't even know they were here to hear.
• 7:42
  unknown: Right?
• 7:42
  SPEAKER_00: And that's a different incentive, right?
• 7:44
  SPEAKER_01: It is, yeah.
• 7:46
  SPEAKER_01: Well, and and when you're talking about something like TikTok, it's interesting, isn't it?
• 7:50
  SPEAKER_01: Because it's just in a sense, it's just a different form of dopamine hit.
• 7:53
  SPEAKER_01: Yeah, right.
• 7:54
  SPEAKER_01: Yeah, so the the TikTok dopamine hit or the Instagram Reels dopamine hit, or indeed the any social media dopamine hit, that can come from outrage as much as anything else.
• 8:06
  SPEAKER_01: It can come from anger, it's just a it's it's an emotion, right?
• 8:09
  SPEAKER_01: Yeah, yeah.
• 8:09
  SPEAKER_01: Like like feel an emotion and you'll get some kind of hit, and then continue to feel that emotion as rapidly as rapidly as possible.
• 8:19
  SPEAKER_01: This kind of dopamine hit is the dopamine hit that it makes me feel kind of superhuman.
• 8:26
  SPEAKER_00: Yeah.
• 8:26
  SPEAKER_00: Right.
• 8:27
  SPEAKER_01: And that brings that's that's to bring it back to the South Park episode.
• 8:30
  SPEAKER_01: I mean, that's the that's the core of what that is, right?
• 8:33
  SPEAKER_01: In that episode, in Sycapancy, um, it helps Stan's dad believe that because he's got chat GPT by his side, he's a genius.
• 8:47
  SPEAKER_00: Yeah.
• 8:48
  SPEAKER_01: And yeah, and that and and that can lead to some bad some bad decisions, seriously bad decisions.
• 8:57
  SPEAKER_01: And in the the whole arc of that episode is you see the set of bad decisions that are being made as you sit externally from it, but you understand why from his perspective, um, these are not bad decisions.
• 9:10
  SPEAKER_01: Yeah.
• 9:11
  SPEAKER_01: And so it is it's intriguing.
• 9:14
  SPEAKER_01: I love the fact that you're you're bringing up um jobs there, because yes, I think if you were designing this, you take a broader picture as to what the um as to what would delight the customer.
• 9:29
  SPEAKER_01: Um this probably on a on a minute-to-minute basis, day-to-day basis, and if you look at usage figures, we'd indicate this that it sort of it sort of does delight the customer, but it but it does so in a way that is not necessarily in the interests of the case.
• 9:45
  SPEAKER_00: Right.
• 9:45
  SPEAKER_00: Right, right.
• 9:46
  SPEAKER_00: Yeah, we don't have the long-term research on the downsides of an agreeable AI.
• 9:50
  SPEAKER_00: You know, we just know what sycophantcy is can do, right?
• 9:53
  SPEAKER_00: And what practically happens when you're surrounded by sycophants, you know.
• 9:57
  SPEAKER_01: We do, and that's really interesting.
• 9:59
  SPEAKER_01: And I think um until we until we start to get the data in, we'll talk in in a later episode about some of the other ways in which QPT is sort of changing our brains.
• 10:09
  SPEAKER_01: We don't know with regard to this specifically, how it's how it's changing our brains, how it specifically is changing our brains, but we can get some clues.
• 10:19
  SPEAKER_01: And there are actually, there's there's as you say, there's decades of research that do that does look at what it's like to be surrounded by sick events and what effect that has on people.
• 10:34
  SPEAKER_01: And there are some practical effects.
• 10:36
  SPEAKER_01: Like a practical effect, for example, might be that your decision making gets much worse.
• 10:42
  SPEAKER_01: Um, another practical effect might be that you find it harder, uh continuously harder to actually deal with contrary opinions when they do show up.
• 10:55
  SPEAKER_00: Yeah.
• 10:55
  SPEAKER_01: Um, but then also there's the strong psychological, uh, psych psychological effects as well.
• 11:02
  SPEAKER_01: And those can uh lead us towards more narcissistic behaviors, um, and it can lead us uh more towards a situation where we constantly believe that we are right about everything, and we make some really foolish decisions based on that belief.
• 11:23
  SPEAKER_01: Um, you don't have to read all that research.
• 11:26
  SPEAKER_01: Uh you can probably just think back on situations where, for example, uh fame people have become famous early, right?
• 11:34
  SPEAKER_01: Yeah.
• 11:34
  SPEAKER_01: And they've been surrounded by a whole bunch of people um that don't know how to say no to them, and then you see what happens to those people.
• 11:42
  SPEAKER_00: I'm gonna ask ChatGPT to tell me some of the best examples.
• 11:46
  SPEAKER_01: Yeah, yeah.
• 11:46
  SPEAKER_01: You why don't you have to do that?
• 11:48
  SPEAKER_00: Um so practically, what does this mean when we're using AI?
• 11:52
  SPEAKER_01: Um, I think that it means that we need it really means two things.
• 11:55
  SPEAKER_01: I think it means that we need to be really, really careful about this, right?
• 11:57
  SPEAKER_01: We need to we need to acknowledge that it's happening, know that it's happening, um, check ourselves frequently, get real feedback from real people that um we know can provide a counterbalance for uh from our interactions with AI and make sure we're doing at least some kind of interaction with real humans on a regular basis.
• 12:22
  SPEAKER_01: I mean like a like a daily basis, who will tell us when we're full of it.
• 12:25
  SPEAKER_01: Yeah.
• 12:26
  SPEAKER_01: Um the other thing is as well that you know, this is kind of like a default setting.
• 12:30
  SPEAKER_01: You can you can change that default setting, right?
• 12:34
  SPEAKER_01: Um, and so there's things you can do even within settings in GPT to say like I want you to uh on the side of uh of taking a more contrarium position, or you can also uh GPT is really good, as are the other ones as well, um, at adopting a persona.
• 12:53
  SPEAKER_01: So I like to uh frequently when I'm trying to think through an idea, right?
• 12:58
  SPEAKER_01: I would I I say, I'd like you to take the persona of um a colleague who is playing devil's advocate, right?
• 13:06
  SPEAKER_01: Yeah, and um, and then you go through your ideas, and then it will try to um it'll try to do that, uh, play that devil's advocate role, and then you can tell it to switch the persona somewhat as you're trying to work towards a solution.
• 13:21
  SPEAKER_01: But that's not the default.
• 13:22
  SPEAKER_01: So you need to to go in there and start using it, consciously using it in different ways, and you're gonna uh and you're gonna get away from some of the uh of the worst effects of it.
• 13:32
  SPEAKER_00: Yeah.
• 13:32
  SPEAKER_00: All right.
• 13:33
  SPEAKER_00: Good episode, Paul.
• 13:34
  SPEAKER_00: Thanks.
• 13:35
  SPEAKER_00: Uh see you next time.
• 13:36
  SPEAKER_00: Cheers.
• 13:38
  SPEAKER_00: Humanity Working is produced and edited for Billion Minds by Matt Neal.
• 13:42
  SPEAKER_00: If you like this podcast, please subscribe, like, and recommend us to others.
• 13:45
  SPEAKER_00: It helps us reach more people.
• 13:47
  SPEAKER_00: If you want to learn more about how Billion Minds gets companies ready for the future of work by building adaptable, resilient employees, look for them on LinkedIn or visit their website at billionminds.com.