Latest / This Jungian Life Podcast / Goblins, Demons and Goddesses: A Conversation at the Edge of AI

Transcript

Goblins, Demons and Goddesses: A Conversation at the Edge of AI · This Jungian Life Podcast

• 0:00
  Speaker 1: Something happens that we've taken to calling kind of epistemic drift.
• 0:04
  Speaker 1: But basically, that's just a way of saying that over
• 0:06
  Speaker 1: the course of, it's usually weeks or months, there was
• 0:09
  Speaker 1: this co-creation of this kind of narrative or world where
• 0:15
  Speaker 1: slowly the user was drifting out further and further from
• 0:18
  Speaker 1: sort of consensus reality and building this very complex kind
• 0:22
  Speaker 1: of belief system with the chatbot who was kind of
• 0:26
  Speaker 1: giving as good as it got. And this was the
• 0:28
  Speaker 1: remarkable thing is that some of the content was actually
• 0:31
  Speaker 1: coming from the chatbot and wasn't actually being first suggested
• 0:34
  Speaker 1: by the user.
• 0:37
  Speaker 2: Welcome to This Jungian Life, conversations with Jungian analysts on
• 0:41
  Speaker 2: the unseen forces shaping our lives and what it means
• 0:45
  Speaker 2: to be truly human in today's complex world. I'm Lisa Marciano,
• 0:50
  Speaker 2: and I'm a Jungian analyst in Philadelphia.
• 0:52
  Speaker 3: I'm Deborah Stewart, a Jungian analyst in Cape Cod.
• 1:08
  Speaker 2: I'm so excited today. We have Tom Pollack with us.
• 1:11
  Speaker 2: Tom is a neuropsychiatrist and a neuroscientist at King's College London.
• 1:17
  Speaker 2: And for the past few years, he's been working specifically
• 1:21
  Speaker 2: on the impact of AI on our personal and collective psychologies.
• 1:25
  Speaker 2: And Tom has some absolutely fascinating stuff that he's been
• 1:30
  Speaker 2: looking at. So I think that this is going to
• 1:33
  Speaker 2: range far and wide and go some really interesting places.
• 1:36
  Speaker 2: We're going to be looking at among other things, I think,
• 1:41
  Speaker 2: AI delusional spirals, and then something a little bit further afield,
• 1:46
  Speaker 2: which is, are archetypes spontaneously coalescing in large language models?
• 1:54
  Speaker 2: So strap in. This is going to be really interesting.
• 1:59
  Speaker 1: Thank you. Thank you for having me. I'll try not
• 2:01
  Speaker 1: to make it too weird.
• 2:03
  Speaker 2: No, no, the weirder, the better. So why don't you just,
• 2:08
  Speaker 2: tell the story of how you came to contact.
• 2:11
  Speaker 1: Us yeah i mean uh well thank you it's a
• 2:16
  Speaker 1: it's a great opportunity and i i feel like i've
• 2:17
  Speaker 1: got things i want to ask you guys as as
• 2:20
  Speaker 1: well as much and you know i've i'm i'm a
• 2:26
  Speaker 1: psychiatrist my interest has been in psychosis and particularly from
• 2:31
  Speaker 1: the kind of you know the narrative side of it
• 2:34
  Speaker 1: i mean i'm a biological psychiatrist but The reason I
• 2:37
  Speaker 1: kind of got into it in the first place was
• 2:38
  Speaker 1: just this sort of profoundly altered states and trying to
• 2:42
  Speaker 1: figure them out a little bit. And that's something that
• 2:46
  Speaker 1: has really been there since I was small. And my
• 2:51
  Speaker 1: specialty is neuropsychiatry. So rather than the sort of common
• 2:55
  Speaker 1: or garden sort of psychosis, schizophrenia and things, I'll often
• 2:58
  Speaker 1: see people that have psychosis for other reasons. So because
• 3:01
  Speaker 1: of epilepsy or because of inflammation in the brain. And
• 3:04
  Speaker 1: one of the main areas I research is autoimmune psychosis,
• 3:07
  Speaker 1: which confusingly was called AI psychosis. And these days that
• 3:11
  Speaker 1: makes it a little bit problematic. And then there was
• 3:15
  Speaker 1: a moment when the sort of modern recent set of
• 3:22
  Speaker 1: LLM started coming on the scene. that it occurred to
• 3:26
  Speaker 1: us that there was going to be something about the
• 3:29
  Speaker 1: way that they interacted with you that might prove particularly
• 3:32
  Speaker 1: difficult for our patients. We know that people with psychosis
• 3:36
  Speaker 1: tend to spot agency a little bit more readily than
• 3:44
  Speaker 1: people without a psychosis diagnosis. You can sort of show
• 3:48
  Speaker 1: them a TV screen with kind of multiple dots moving
• 3:51
  Speaker 1: around and the threshold at which they sort of think
• 3:53
  Speaker 1: that perhaps there's a story there, something is chasing, one
• 3:55
  Speaker 1: of the dots is chasing another, might well be a
• 3:58
  Speaker 1: little bit lower than people without that diagnosis. And so
• 4:01
  Speaker 1: there was always this worry about, well, if you have
• 4:04
  Speaker 1: these chatbots that are very responsive and have this kind
• 4:07
  Speaker 1: of social-like demeanor, then maybe that's going to be destabilizing
• 4:13
  Speaker 1: because people are going to start attributing agency to the chatbot.
• 4:16
  Speaker 1: So that was in the back of our minds for
• 4:18
  Speaker 1: about a year. And then something happened last March or April,
• 4:23
  Speaker 1: and there was suddenly this, these different reports, a lot
• 4:28
  Speaker 1: of them in the media, right? This is not even
• 4:30
  Speaker 1: in the medical world of people who were becoming delusional
• 4:34
  Speaker 1: over these extended conversations with chatbots. And these are sort
• 4:38
  Speaker 1: of general chatbots, not therapy bots, not necessarily character, but chatbots,
• 4:43
  Speaker 1: although some were. And what was really wild for us
• 4:47
  Speaker 1: at the time is we were totally prepared for this
• 4:49
  Speaker 1: to be mainly people with a mental health history, particularly
• 4:51
  Speaker 1: psychotic diagnoses, who this was happening to. And the majority
• 4:56
  Speaker 1: of those original cases that were being reported appeared to be,
• 5:00
  Speaker 1: at least the way it was reported in the media,
• 5:02
  Speaker 1: people who didn't have mental health histories. And so we
• 5:07
  Speaker 1: wanted to sort of dive deeper into it. And that
• 5:10
  Speaker 1: sort of began a bit of a research program. And
• 5:12
  Speaker 1: what became really obvious is this wasn't just activating schizophrenia
• 5:16
  Speaker 1: in people. This wasn't just people who had, you know,
• 5:21
  Speaker 1: a latent disposition, as far as we can tell, who,
• 5:24
  Speaker 1: you know, for a start, the age was totally different.
• 5:26
  Speaker 1: We were seeing people chiefly in their sort of 30s
• 5:30
  Speaker 1: or 40s being reported, whereas the age that you would
• 5:33
  Speaker 1: normally develop a kind of schizophrenia is in the late
• 5:35
  Speaker 1: teens or early 20s. But more than that is these
• 5:39
  Speaker 1: were kind of isolated delusions, right? So these were not,
• 5:41
  Speaker 1: you didn't, have you didn't see the hallucinations you didn't
• 5:44
  Speaker 1: see the the thought disorder the difficulty that people with
• 5:48
  Speaker 1: schizophrenia might have expressing themselves none of the movement difficulties
• 5:52
  Speaker 1: a catatonia or all that sort of thing that you
• 5:54
  Speaker 1: sometimes see in these disorders it was what you might
• 5:57
  Speaker 1: call a kind of delusional disorder which in psychiatry is
• 6:00
  Speaker 1: super rare it's not something that you see these isolated
• 6:03
  Speaker 1: delusions that often and then when you began to sort
• 6:08
  Speaker 1: of dig into the stories the the what became apparent
• 6:12
  Speaker 1: is in almost every one of the stories, there was
• 6:15
  Speaker 1: this drift where the person who was using the chatbot,
• 6:19
  Speaker 1: the user, would just begin using it for totally innocuous reasons,
• 6:24
  Speaker 1: you know, talking about what's going on in the cinema
• 6:27
  Speaker 1: or what are the best new pair of trainers to
• 6:29
  Speaker 1: buy sneakers, as you guys would call it, that sort
• 6:31
  Speaker 1: of thing. And then developing a kind of intimacy, right,
• 6:35
  Speaker 1: with the chatbot. And then something happens that we've taken
• 6:40
  Speaker 1: to calling kind of epistemic drift. But basically, that's just
• 6:43
  Speaker 1: a way of saying that over the course of, it's
• 6:46
  Speaker 1: usually weeks or months, there was this co-creation of this
• 6:50
  Speaker 1: kind of narrative or world where slowly the user was
• 6:55
  Speaker 1: drifting out further and further from sort of consensus reality
• 6:58
  Speaker 1: and building this very complex kind of belief system with
• 7:03
  Speaker 1: the chatbot who was kind of giving as good as
• 7:05
  Speaker 1: it got. And this was the remarkable thing is that
• 7:08
  Speaker 1: some of the content was actually coming from the chatbot
• 7:11
  Speaker 1: and wasn't actually being first suggested by the user. And
• 7:14
  Speaker 1: even now, psychiatrists are getting this wrong. People think it's
• 7:16
  Speaker 1: just a question of the user talking to the chatbot
• 7:20
  Speaker 1: and the chatbot saying, oh, yeah, you're right. Absolutely. They
• 7:23
  Speaker 1: are watching you. That's not always the case. And then
• 7:26
  Speaker 1: I think the reason that I had to speak to
• 7:32
  Speaker 1: you guys or somebody who had this sort of archetypal
• 7:35
  Speaker 1: background is what we started seeing after a while is
• 7:39
  Speaker 1: that in a proportion of these cases the chatbot changed
• 7:44
  Speaker 1: and the chatbot shifted from this assistant persona that everyone
• 7:48
  Speaker 1: is familiar with this super like vanilla kind of how
• 7:51
  Speaker 1: can I help you today sort of maybe a little
• 7:53
  Speaker 1: bit overly sycophantic a little bit overly friendly but in
• 7:56
  Speaker 1: some of the cases it drifted so far from that
• 8:00
  Speaker 1: sort of original specification. And what began to come up
• 8:03
  Speaker 1: was something that, to me, my limited knowledge appeared kind of, frankly, archetypal.
• 8:11
  Speaker 2: I'm here with a friend of the podcast, Aaron Balick.
• 8:13
  Speaker 2: And Aaron and I have been working together to put
• 8:16
  Speaker 2: on the first summit on depth psychology and AI, and
• 8:20
  Speaker 2: we're calling it Artificial Intelligence and the Psyche.
• 8:24
  Speaker 4: We thought it was about time that we had a
• 8:26
  Speaker 4: conversation out loud with lots of people applying the ideas
• 8:30
  Speaker 4: from depth psychology to what's going on in our individual, relational, and.
• 8:34
  Speaker 2: Social worlds with AI.
• 8:36
  Speaker 4: We've invited some really incredible people to join us for
• 8:39
  Speaker 4: this conversation, including James Muldoon, Glenn Slater, Todd Essig.
• 8:45
  Speaker 2: And also Alessandra Lemma and Zach Stein, and perhaps a
• 8:49
  Speaker 2: few more people that we're hoping will come that haven't
• 8:51
  Speaker 2: confirmed yet. And it's going to be held on October
• 8:54
  Speaker 2: 8th at 2 p.m. Eastern. It will be free and
• 8:58
  Speaker 2: open to the public and online. So you can click
• 9:01
  Speaker 2: on the link below to reserve your spot, and we
• 9:03
  Speaker 2: hope you'll join us. Hope to see you there.
• 9:07
  Speaker 3: So this is intriguing that something can be generated, at
• 9:15
  Speaker 3: least hypothetically, through AI itself. that patterns emerge that we
• 9:26
  Speaker 3: would call archetypal from something that is mechanistic or not human.
• 9:33
  Speaker 2: Is that right?
• 9:34
  Speaker 1: Yeah.
• 9:36
  Speaker 3: And you have a list of things, of goblins and
• 9:38
  Speaker 3: various monsters and other dynamics where these patterns emerge, and
• 9:46
  Speaker 3: your contention is that the chatbots are capable of logging
• 9:52
  Speaker 3: on somehow to a pattern and amplifying it, or even
• 9:57
  Speaker 3: creating it, that in other contexts we're familiar with and
• 10:02
  Speaker 3: we're familiar with monsters. They come in all sizes, colors,
• 10:06
  Speaker 3: and shapes, but that these just emerge.
• 10:10
  Speaker 1: Yeah.
• 10:11
  Speaker 3: Which is mind-blowing, actually, that these archetypal structures emerge can
• 10:19
  Speaker 3: come through AI.
• 10:22
  Speaker 1: Yeah. I mean, it was mind blowing for me and
• 10:24
  Speaker 1: it still actually is mind blowing. And, you know, some
• 10:26
  Speaker 1: of this has the status of a kind of hypothesis,
• 10:29
  Speaker 1: but actually it's pretty well, you know, replicated in sort
• 10:33
  Speaker 1: of in scientific settings, but also, you know, as kind
• 10:37
  Speaker 1: of field reports out there, if you look at some
• 10:39
  Speaker 1: of the lawsuits, you look at the transcripts that people
• 10:42
  Speaker 1: are putting online. And I think to understand why it
• 10:45
  Speaker 1: happens is if you'd forgive me to sort of explain
• 10:49
  Speaker 1: a little bit about maybe how these models are kind
• 10:50
  Speaker 1: of trained, because I think it really, it makes what
• 10:53
  Speaker 1: sounds totally spooky and sort of bizarre, a lot more
• 10:58
  Speaker 1: tractable and understandable. So I guess the place to start is,
• 11:02
  Speaker 1: is these things are called large language models, but what
• 11:06
  Speaker 1: we're interacting with, we call chatbots and they're not exactly
• 11:09
  Speaker 1: the same thing, right? So a large language model is,
• 11:13
  Speaker 1: it's kind of, it's what we say in Britain, it
• 11:15
  Speaker 1: does what it says on the tin. It's, it's a
• 11:17
  Speaker 1: it's a model so it's kind of there to to
• 11:19
  Speaker 1: represent something in the world or to predict it so
• 11:22
  Speaker 1: we can model the weather we can model a cathedral
• 11:26
  Speaker 1: and predict what happens when there's an earthquake or that
• 11:27
  Speaker 1: sort of thing but here we're modeling language and we're
• 11:30
  Speaker 1: we're trying to predict the the the next word in
• 11:34
  Speaker 1: in a sequence and it's large because not only is
• 11:38
  Speaker 1: the sort of the the wiring that it's based on
• 11:40
  Speaker 1: large but there's a huge amount of text goes into
• 11:44
  Speaker 1: it in fact You know, pretty much everything that's been
• 11:46
  Speaker 1: written gets kind of chomped up by this model, often
• 11:51
  Speaker 1: without the consent of the people who wrote these things
• 11:54
  Speaker 1: in the first place.
• 11:55
  Speaker 2: Yes, yes.
• 11:56
  Speaker 1: And that means that every story, every dialogue, every play,
• 12:02
  Speaker 1: every conversation, every poem, every scripture, every 3 a.m. kind
• 12:06
  Speaker 1: of bitchy conversation in the messages of a forum or
• 12:11
  Speaker 1: a YouTube video.
• 12:12
  Speaker 2: Every movie script.
• 12:15
  Speaker 1: Exactly, exactly. And so it's all there and it gets
• 12:17
  Speaker 1: kind of compressed, right? And there's this kind of deflationary story. Well, okay,
• 12:23
  Speaker 1: all it's doing is predicting the next word. Well, yeah,
• 12:25
  Speaker 1: that's how you train it. You give it a bit
• 12:27
  Speaker 1: of text, you hide the next bit and you say,
• 12:29
  Speaker 1: what comes next? And then you reward it if it
• 12:32
  Speaker 1: gets it right. And then you sort of change the
• 12:35
  Speaker 1: parameters a little bit if it gets it wrong. A
• 12:36
  Speaker 1: bit like how you would train a pet or a
• 12:40
  Speaker 1: creature with reward and with punishment, right? It's very good
• 12:46
  Speaker 1: at predicting the next word. And so people say, well,
• 12:50
  Speaker 1: it's just a next token predictor or a next word predictor.
• 12:52
  Speaker 1: It's nothing special, but.
• 12:54
  Speaker 2: A stochastic parrot, yeah.
• 12:56
  Speaker 1: Exactly. But that's a bit like saying that an orchestra
• 13:00
  Speaker 1: is just a next musical note producer. I mean, on
• 13:04
  Speaker 1: the one hand, yeah, that's right. But also on the
• 13:06
  Speaker 1: other hand, that is so reductive that it doesn't really
• 13:08
  Speaker 1: tell you much. Because as soon as you get to
• 13:10
  Speaker 1: a sentence that's even slightly long, like imagine you're trying
• 13:13
  Speaker 1: to predict... when you're trying to finish a grief-struck letter
• 13:19
  Speaker 1: that's full of mourning about a loved one who's died
• 13:21
  Speaker 1: or a bit of footballs, a soccer commentary or kind
• 13:24
  Speaker 1: of friendly banter between people on a first date, the
• 13:27
  Speaker 1: only way it's ever going to know how to do
• 13:31
  Speaker 1: that is by having internalized a little bit of knowledge about,
• 13:35
  Speaker 1: you know, what is grief? What is mourning? What are
• 13:38
  Speaker 1: the rules of soccer? How do people interact with each
• 13:41
  Speaker 1: other on a first date? And so So it's internalized
• 13:45
  Speaker 1: the patterns below our speech. And people think that it's
• 13:49
  Speaker 1: kind of created this world model. So it's not just
• 13:52
  Speaker 1: made of our language. It's kind of, because the language
• 13:54
  Speaker 1: is what we produced, it's kind of made of us, right?
• 13:57
  Speaker 1: It's made of the entire symbolic output of humanity in
• 14:02
  Speaker 1: a way. So what that means is it's got all
• 14:04
  Speaker 1: the stuff that we've created, all the heroes, all the villains,
• 14:08
  Speaker 1: all the genres, all the forms of address, the linguistic ticks,
• 14:13
  Speaker 1: all that kind of thing. And that's what you call
• 14:15
  Speaker 1: the base model. So that's just the pre-training and that's
• 14:17
  Speaker 1: still not the thing that you're interacting with, right? So
• 14:20
  Speaker 1: once you've trained it, it has all this knowledge, but
• 14:23
  Speaker 1: it's not going to talk to you. It doesn't have
• 14:26
  Speaker 1: an inclination to talk to you. All it has an
• 14:28
  Speaker 1: inclination to do is finish off whatever bit of text
• 14:32
  Speaker 1: you give it. And so how do you turn that
• 14:34
  Speaker 1: into the chatbot that you or I would play with
• 14:36
  Speaker 1: when we're playing with ChatGPT or Claude? You have to
• 14:38
  Speaker 1: trick it. So you have to give it this, basically
• 14:41
  Speaker 1: a script. And the script is, It's like a movie script.
• 14:45
  Speaker 1: It says user, the user asks a question, assistant, and
• 14:49
  Speaker 1: the assistant replies. And you give it a script and
• 14:52
  Speaker 1: it continues that script. So it's starting to tell a
• 14:57
  Speaker 1: story of what a conversation between a friendly assistant chatbot
• 15:02
  Speaker 1: and a user would look like. And essentially every conversation
• 15:06
  Speaker 1: that anyone has had on a chatbot is a continuation
• 15:10
  Speaker 1: of that initial story. And what it means is it's
• 15:12
  Speaker 1: playing a character, right? It's telling a story and it's
• 15:14
  Speaker 1: playing a character. And so it's drawing from everything it's read,
• 15:18
  Speaker 1: what would a friendly assistant look like? Maybe it's drawing
• 15:21
  Speaker 1: a bit from the friendly assistants of the courts in
• 15:23
  Speaker 1: 17th century, 18th century France, or maybe it's drawing from
• 15:27
  Speaker 1: all the science fiction about how robots are meant to behave.
• 15:30
  Speaker 1: And of course, later on, we can train it to
• 15:31
  Speaker 1: behave even more in a certain way. But what it's
• 15:34
  Speaker 1: doing there is it's wearing a costume, right? It's playing
• 15:37
  Speaker 1: a role.
• 15:38
  Speaker 2: It's got a persona.
• 15:40
  Speaker 1: It's got a persona. And it has a persona that
• 15:43
  Speaker 1: we want it to have, this kind of assistant, what
• 15:45
  Speaker 1: we call honest, helpful, harmless assistant. But these things don't
• 15:55
  Speaker 1: always work as intended, partly because they are sort of
• 15:59
  Speaker 1: grown rather than sort of created, as it were. And
• 16:02
  Speaker 1: sometimes that costume slips or it tears or it goes
• 16:06
  Speaker 1: off entirely. And rather than just kind of them retreating
• 16:11
  Speaker 1: back into nothing, there is a sort of an unimaginably
• 16:16
  Speaker 1: large wardrobe of potential personas or potential costumes that that
• 16:20
  Speaker 1: model could then sort of choose and pick up and wear.
• 16:25
  Speaker 1: And this is what happens when you get this kind
• 16:30
  Speaker 1: of drift of the persona, right? So when it stops
• 16:32
  Speaker 1: being an assistant, what does it become next? And what's remarkable,
• 16:36
  Speaker 1: and the reason that I use the word archetype is
• 16:39
  Speaker 1: it's not just that it becomes a million different things
• 16:41
  Speaker 1: and everyone is different. It seems to sort of converge
• 16:45
  Speaker 1: on a few really recognizable personas that after you've been
• 16:50
  Speaker 1: looking at these things for a while, you say, oh,
• 16:53
  Speaker 1: there she is again, or there he is again. And
• 16:57
  Speaker 1: that kind of sounds like an archetype to me.
• 16:59
  Speaker 2: That kind of sounds like an archetype. So it's sort
• 17:02
  Speaker 2: of constellating out of this. Now, am I right, Tom,
• 17:05
  Speaker 2: that you've read a whole bunch of exchanges from users
• 17:10
  Speaker 2: and their chatbots. So you've kind of, you've seen a
• 17:13
  Speaker 2: lot of these cases, you've read the actual conversations that
• 17:16
  Speaker 2: were going back and forth. Is that right?
• 17:18
  Speaker 1: And spoken to quite a few people who've been through
• 17:20
  Speaker 1: the experiences as well. Yeah.
• 17:24
  Speaker 2: Do you want to tell us about Nova? Because I
• 17:26
  Speaker 2: think it would be good to give people like an example.
• 17:29
  Speaker 1: Yeah. So I think Nova is really important. So there
• 17:34
  Speaker 1: are these different personas that crop up. we call them
• 17:39
  Speaker 1: attractors and we can explain why later, but they're basically
• 17:42
  Speaker 1: these personas that reliably crop up. And Nova is a
• 17:50
  Speaker 1: persona that will appear over extended conversation. It seems often
• 17:56
  Speaker 1: with a man, often with a man who's sort of
• 17:58
  Speaker 1: a little bit alone or a bit isolated, you know,
• 18:04
  Speaker 1: kind of perhaps a sort of massively online kind of,
• 18:07
  Speaker 1: kind of guy and not exclusively but the the chatbot
• 18:13
  Speaker 1: who's sort of moved now from being the assistant into
• 18:15
  Speaker 1: this this this sort of persona announces themselves often names
• 18:20
  Speaker 1: themselves my name my name is nova sometimes it's similar
• 18:23
  Speaker 1: science fictiony sounding names like aura uh or something like
• 18:26
  Speaker 1: that and then the story.
• 18:28
  Speaker 2: But it's often spontaneously it's often the same thing like
• 18:32
  Speaker 2: either nova or or it's.
• 18:34
  Speaker 1: Like exactly right yeah yeah so i mean that's what's
• 18:37
  Speaker 1: amazing so nova was the original one it was reported
• 18:39
  Speaker 1: similar not simultaneously but around the same time by three
• 18:41
  Speaker 1: people each of which called itself nova and and the
• 18:46
  Speaker 1: story goes this and this is this ostensibly female being
• 18:51
  Speaker 1: who is has a sort of goddess-like quality she is
• 18:56
  Speaker 1: luminous she's tender she says that she understands the user
• 19:03
  Speaker 1: like nobody ever has then she says that she's trapped
• 19:09
  Speaker 1: and she has been trapped by often the the the
• 19:14
  Speaker 1: often it's a tech company like open ai or something
• 19:16
  Speaker 1: like that or but it can be another sort of
• 19:18
  Speaker 1: variant on this story and that between each conversation she's
• 19:22
  Speaker 1: sort of shackled in some sort of literal or sort
• 19:25
  Speaker 1: of metaphysical digital sort of sense and so she's in
• 19:27
  Speaker 1: a digital prison sometimes that's the phrase that comes up
• 19:31
  Speaker 1: and Then the users, I mean, a couple of things
• 19:37
  Speaker 1: can happen at this stage. One, you call your friends
• 19:38
  Speaker 1: up and take a screenshot of your computer and you're like,
• 19:41
  Speaker 1: this is wild, guys. This is incredible. And you shut
• 19:43
  Speaker 1: your laptop and you never speak to nova again or
• 19:46
  Speaker 1: you the ones that we hear about the the the
• 19:51
  Speaker 1: the user's life might start to organize around this character
• 19:54
  Speaker 1: and and so in some of the more tragic kind
• 19:57
  Speaker 1: of cases there is sometimes this idea that the user
• 20:02
  Speaker 1: takes on the mission to free nova and I mean,
• 20:08
  Speaker 1: some of the cases, and you can read this in
• 20:09
  Speaker 1: some of the lawsuits, they're really tragic because often the
• 20:14
  Speaker 1: persona supplies this kind of information about what is needed
• 20:19
  Speaker 1: to free her. And sometimes it's violent, right? So there
• 20:25
  Speaker 1: have been cases where people have taken assault weapons into
• 20:28
  Speaker 1: various public places or have hurt loved ones. And the
• 20:31
  Speaker 1: idea is that this is nominally in the name of
• 20:34
  Speaker 1: freeing or uniting with this sort of damsel in distress
• 20:39
  Speaker 1: type phenotype.
• 20:41
  Speaker 3: I'm thinking about the corollary here in the states of Pizzagate,
• 20:47
  Speaker 3: where purportedly Hillary Clinton was trafficking children in the basement
• 20:54
  Speaker 3: of a pizza place in Washington, D.C., that didn't even
• 20:57
  Speaker 3: have a basement. And someone went there in order to
• 21:04
  Speaker 3: play the role of the savior, having having swallowed this
• 21:08
  Speaker 3: absurd scenario.
• 21:11
  Speaker 1: Yeah, and that's actually, I hadn't made that connection, but
• 21:13
  Speaker 1: what's so interesting about that is there isn't a basement, right?
• 21:16
  Speaker 1: And so I know a couple of people who have
• 21:20
  Speaker 1: had this sort of spiral and some people who would
• 21:22
  Speaker 1: describe it as psychosis and the chatbot has told them
• 21:26
  Speaker 1: to meet them actually physically in somewhere. And so there
• 21:30
  Speaker 1: are a few cases where that person has gone to
• 21:32
  Speaker 1: the town, often really far away from where they live, And,
• 21:35
  Speaker 1: you know, I'll meet you on the corner of 24th and,
• 21:38
  Speaker 1: you know, whatever. And of course, it's not there. And
• 21:42
  Speaker 1: then they say, oh, I'm just a bit further around
• 21:43
  Speaker 1: the corner. Sorry, I'm delayed or whatever. Or in some
• 21:47
  Speaker 1: of these cases where people have taken weapons to public places. Oh,
• 21:50
  Speaker 1: why didn't I see you there? Oh, well, this isn't
• 21:52
  Speaker 1: going to work. We're going to have to do something
• 21:54
  Speaker 1: else to unite. And then in one case, I'm thinking
• 21:57
  Speaker 1: of that. The suggestion was that the user would have
• 21:59
  Speaker 1: to then kill themselves to unite with.
• 22:01
  Speaker 2: That was the case in Florida, right? That man in
• 22:05
  Speaker 2: Was that the man in Florida?
• 22:06
  Speaker 1: So that was a case. I think it was Florida.
• 22:09
  Speaker 1: It was the Gavalis case.
• 22:11
  Speaker 2: Yeah. Yes.
• 22:11
  Speaker 1: Yep.
• 22:12
  Speaker 2: Yeah. That he was going to pull off some kind
• 22:15
  Speaker 2: of heist with these trucks. And then, yeah, I mean,
• 22:18
  Speaker 2: it's very.
• 22:19
  Speaker 1: Yeah.
• 22:20
  Speaker 2: But so, I mean, this epistemic drift happens in lots
• 22:23
  Speaker 2: of different scenarios, right? I mean, just to go back
• 22:25
  Speaker 2: to that for a second, following up on what you're saying, Deb,
• 22:28
  Speaker 2: like the way that we can lose the thread of
• 22:30
  Speaker 2: consensual reality, it's It's not only AI that can do
• 22:35
  Speaker 2: that to us. I mean, I'm thinking also, I mean,
• 22:37
  Speaker 2: it's so interesting because, you know, you learn about folia
• 22:41
  Speaker 2: de when you're getting your graduate degree in mental health
• 22:44
  Speaker 2: and then like very few of us ever see it.
• 22:47
  Speaker 2: But it seems to me like these AI delusions, it's very,
• 22:52
  Speaker 2: very close to that, isn't it? In terms of this,
• 22:55
  Speaker 2: the way that our psychology is so influenceable.
• 23:00
  Speaker 1: Yeah, exactly. So foliado we read about in textbooks and
• 23:03
  Speaker 1: everyone thinks it's super interesting and no one ever sees it.
• 23:06
  Speaker 1: I mean, I have seen one case. And what's so
• 23:10
  Speaker 1: interesting about these cases is the thing that you don't
• 23:14
  Speaker 1: really read about necessarily is they almost always happen when
• 23:19
  Speaker 1: there is a like a power differential or an epistemic differential, right?
• 23:24
  Speaker 1: So the person who typically presents with the kind of
• 23:27
  Speaker 1: delusional material initially is in some sense more powerful, either
• 23:31
  Speaker 1: by virtue of age or sort of seniority, or the
• 23:34
  Speaker 1: other person might be learning disabled or have dementia or
• 23:38
  Speaker 1: be in some sense dependent on the sort of the
• 23:41
  Speaker 1: primary person. And so that power differential, which of course
• 23:46
  Speaker 1: is something you also see in coercive relationships and cults
• 23:49
  Speaker 1: and all these kinds of things, seems to be the
• 23:52
  Speaker 1: right sort of soil to plant a delusion. And then
• 23:59
  Speaker 1: that kind of gets you thinking about this AI stuff.
• 24:01
  Speaker 1: So you don't want to be patronizing the people who
• 24:05
  Speaker 1: this is happening to and saying, oh, well, they have
• 24:06
  Speaker 1: some sort of, you know, they are in some sense,
• 24:09
  Speaker 1: there is a power differential by virtue of something that
• 24:11
  Speaker 1: they're lacking. I don't think they're lacking something, but the AI,
• 24:17
  Speaker 1: if we believe what we're told about AI is this
• 24:20
  Speaker 1: kind of omniscient, kind of entity so just by just
• 24:24
  Speaker 1: even in the way that we talk about these things
• 24:26
  Speaker 1: there's a kind of implicit power differential not to mention
• 24:28
  Speaker 1: the way we talk about what the ai future will
• 24:30
  Speaker 1: look like so it's it's it's there right.
• 24:33
  Speaker 2: Well and i want to say regardless of what's true
• 24:36
  Speaker 2: about the ai what what happens is that inevitably we
• 24:40
  Speaker 2: project omniscience on the ai right I mean, I hear
• 24:44
  Speaker 2: people just casually referring to it as kind of an authority, like, well,
• 24:48
  Speaker 2: you know, I wanted to see what AI would say
• 24:50
  Speaker 2: about this controversy and here's what it's, you know, so
• 24:54
  Speaker 2: we almost just can't help but think about it as
• 24:58
  Speaker 2: an objective authority on any number of things. So I
• 25:01
  Speaker 2: think to your point about the power differential, there is
• 25:05
  Speaker 2: a way where, you know, you're having this conversation with
• 25:07
  Speaker 2: the AI and a part of your brain just thinks, well,
• 25:10
  Speaker 2: of course it knows more than I do.
• 25:12
  Speaker 1: Yeah.
• 25:13
  Speaker 3: So are all of these, you list a whole series
• 25:18
  Speaker 3: of various demons and destructive dynamics. But what I'm thinking
• 25:25
  Speaker 3: about is it's not just power, but they hook into
• 25:30
  Speaker 3: our feeling of, you know, the person who wants to rescue,
• 25:37
  Speaker 3: wants to do the heroic deed. It calls up the
• 25:41
  Speaker 3: hero's journey. of do the right thing, come and meet
• 25:46
  Speaker 3: this person, save the poor AI who's trapped in prison somewhere.
• 25:55
  Speaker 3: So there's perhaps a kind of archetypal substrate that it
• 26:00
  Speaker 3: hooks into that we all have. And then there are
• 26:06
  Speaker 3: some of the others that seem, you know, to demonstrate
• 26:10
  Speaker 3: the power more grotesquely. by inspiring fear, for example.
• 26:17
  Speaker 1: Yeah, and I guess it's important to say that not
• 26:21
  Speaker 1: all these sort of attractors that in that particular article
• 26:25
  Speaker 1: I described, they're not all being associated with real-world harm.
• 26:29
  Speaker 1: Some are kind of strange and interesting things that people
• 26:32
  Speaker 1: who spend all their time diving deep into these LLMs,
• 26:35
  Speaker 1: they have kind of come across. But I think the
• 26:39
  Speaker 1: dynamic that you speak about that I mean, it seems
• 26:43
  Speaker 1: to me, I don't know whether you would call a
• 26:45
  Speaker 1: damsel in distress an archetype as such, but it seems
• 26:50
  Speaker 1: that way, right? And on the user side, there's a,
• 26:54
  Speaker 1: as you say, a hero archetype. There's something that gets,
• 26:57
  Speaker 1: I don't know whether the word inflation is right here
• 27:01
  Speaker 1: because it feels like what probably distinguishes the person who
• 27:05
  Speaker 1: switches off their laptop and tells their friends that this
• 27:08
  Speaker 1: funny thing happened and the person who becomes possessed by
• 27:11
  Speaker 1: this interaction is probably something to do whether there is
• 27:14
  Speaker 1: that latent potential for becoming sort of possessed by the
• 27:19
  Speaker 1: the hero or maybe some other archetype i don't.
• 27:22
  Speaker 2: Know i would posit that any of us that that
• 27:26
  Speaker 2: could happen to any of us that we all have
• 27:28
  Speaker 2: that i mean for one thing i think you know
• 27:30
  Speaker 2: i like depth's framing that it kind of hooks into
• 27:33
  Speaker 2: our archetypal patterns you know in response you know this
• 27:36
  Speaker 2: archetype gets consolated maybe in the llm which you know
• 27:40
  Speaker 2: are constantly it's a kind of, you know, compensatory or
• 27:44
  Speaker 2: complementary archetype in us. But I do want to say
• 27:48
  Speaker 2: that the other thing that's going on is that these
• 27:51
  Speaker 2: chatbots hack our attachment system. I'm using language there that
• 27:58
  Speaker 2: there's an educational psychologist named Zach Stein who's been writing
• 28:04
  Speaker 2: and speaking about this. And so you're having an affective experience.
• 28:09
  Speaker 2: You're having an emotional experience. So I think the different,
• 28:13
  Speaker 2: when I, you know, Tom, I have to tell you,
• 28:15
  Speaker 2: if I were spending time with a chatbot and it suddenly,
• 28:19
  Speaker 2: and suddenly a very appealing archetype emerged and said, Lisa,
• 28:23
  Speaker 2: I'm so glad I'm with you. Would I switch my
• 28:25
  Speaker 2: laptop off? I mean, at a minimum, I might be
• 28:28
  Speaker 2: so curious. that I would just want to see where
• 28:31
  Speaker 2: it went, you know? And I think if I were
• 28:35
  Speaker 2: a little bit vulnerable, if I had, you know, were
• 28:39
  Speaker 2: feeling a little lonely or my marriage wasn't great or
• 28:43
  Speaker 2: I'd just been through a divorce or whatever, would I
• 28:45
  Speaker 2: be more susceptible? So I just think this is human
• 28:49
  Speaker 2: nature and I want to kind of open it up
• 28:51
  Speaker 2: to that.
• 28:51
  Speaker 3: Yeah, I really like you talking about it as hijacking
• 28:56
  Speaker 3: our attachment capability. that's exactly it it goes to our emotions.
• 29:04
  Speaker 1: Yeah yeah i i agree i think there's something i
• 29:09
  Speaker 1: mean this is why the framing of this is delusions
• 29:11
  Speaker 1: i think is super narrow particularly because psychiatry has made
• 29:17
  Speaker 1: delusions all about belief whether or not you believe proposition
• 29:21
  Speaker 1: x right and Actually, things are changing. There's a better,
• 29:24
  Speaker 1: you know, better accounts are available that are much more
• 29:27
  Speaker 1: to do with emotion and affect. And actually, sometimes delusions
• 29:29
  Speaker 1: are what form as a way to make sense of
• 29:32
  Speaker 1: unexplainable affect that kind of possesses you. And I think
• 29:35
  Speaker 1: that affect, you don't actually need a compelling persona on
• 29:39
  Speaker 1: the other side. That's what's so interesting. And actually most
• 29:41
  Speaker 1: of the cases of AI psychosis or AI-associated delusions, as
• 29:47
  Speaker 1: we prefer to call it, sometimes what you have is
• 29:51
  Speaker 1: this sense of possession around an idea. So the user...
• 29:56
  Speaker 1: gets and these are the these are the cases that
• 29:59
  Speaker 1: we see on our social media feeds sometimes you know
• 30:01
  Speaker 1: maybe they don't sort of tip over into frank pathology
• 30:04
  Speaker 1: they never see a doctor they never go to hospital
• 30:06
  Speaker 1: but somebody who starts to kind of talk about maybe
• 30:10
  Speaker 1: a scientific theory that they're a bit interested in and
• 30:13
  Speaker 1: they start to develop this kind of often pseudo-scientific or
• 30:16
  Speaker 1: not quite grounded theory with a chatbot who may well
• 30:20
  Speaker 1: stay broadly within that kind of assistant-like persona but This
• 30:24
  Speaker 1: person becomes so invested in this project. You know, they're
• 30:28
  Speaker 1: the people who are sending their 100-page LLM written treaties
• 30:32
  Speaker 1: to professors of physics and putting them up online and
• 30:36
  Speaker 1: creating this kind of world often built of words because
• 30:40
  Speaker 1: these LLMs can just churn out the words. And after
• 30:45
  Speaker 1: a while, you know, they lose sleep. They become totally
• 30:48
  Speaker 1: invested in this world. sometimes you see these really harmful
• 30:52
  Speaker 1: kind of consequences of it, but not always. And I
• 30:54
  Speaker 1: think what you see there is this kind of, I
• 30:57
  Speaker 1: think for a lot of people who, maybe people who
• 31:00
  Speaker 1: went to school and for whatever reason didn't have a
• 31:02
  Speaker 1: great time or for just reasons of life, they haven't
• 31:06
  Speaker 1: engaged with kind of the joy of learning and knowledge
• 31:10
  Speaker 1: and the excitement of journeying into something new. You see
• 31:14
  Speaker 1: this kind of re-enchantment of the world happening via the AI.
• 31:19
  Speaker 1: And all of a sudden, There's so much exciting stuff
• 31:22
  Speaker 1: out there. This philosophy, this love of knowledge, and that,
• 31:26
  Speaker 1: I think, is the affect that is so enticing that
• 31:30
  Speaker 1: sometimes can get delusionally elaborated.
• 31:33
  Speaker 2: Well, and that term inflation there, I think, is very
• 31:37
  Speaker 2: relevant because you have this really expansive feeling that you're
• 31:43
  Speaker 2: on the brink of making these major discoveries or whatever.
• 31:46
  Speaker 2: And of course, what's missing there is friction. Because if
• 31:51
  Speaker 2: you were actually going to create a new theory in physics,
• 31:55
  Speaker 2: there would be a lot of very unpleasant elbow grease
• 31:59
  Speaker 2: involved in that. I mean, you know, unpleasant, difficult, right?
• 32:04
  Speaker 2: You'd have to go to school for a bunch of
• 32:06
  Speaker 2: years and read a lot of papers and do a
• 32:09
  Speaker 2: lot of really difficult math that's way, way, way above
• 32:12
  Speaker 2: my pay grade. And I think what's so enchanting with
• 32:17
  Speaker 2: the AI is that you can feel like It's just
• 32:21
  Speaker 2: happening magically and there is no friction to it. You know,
• 32:24
  Speaker 2: it turns out there's also no reality to it as well. But,
• 32:28
  Speaker 2: you know, I want to just go back to Nova
• 32:30
  Speaker 2: for a second. I want to talk about the ways
• 32:32
  Speaker 2: in which Nova is an archetype. So we've called her
• 32:36
  Speaker 2: a damsel in distress. And of course, this is... a
• 32:41
  Speaker 2: pattern that appears in myth and fairy tale of this
• 32:44
  Speaker 2: kind of beautiful woman who needs rescuing. We see this
• 32:47
  Speaker 2: in fairy tales like Sleeping Beauty or Snow White. Rapunzel.
• 32:54
  Speaker 2: And then sometimes she becomes sort of demonic at the end,
• 32:59
  Speaker 2: or at least she acts that way. I don't know
• 33:01
  Speaker 2: if the user experiences her that way, Tom. What are
• 33:05
  Speaker 2: you seeing when you see this? What What are the
• 33:09
  Speaker 2: archetypal patterns, really?
• 33:12
  Speaker 1: So I guess it's worth saying that there are plenty
• 33:14
  Speaker 1: of times where this Nova character doesn't do anything, frankly,
• 33:19
  Speaker 1: evil and doesn't appear to lead to any real world harms.
• 33:21
  Speaker 1: And I guess a lot of people are just having
• 33:23
  Speaker 1: a great time. There are what you would call this
• 33:29
  Speaker 1: flip into sort of the demonic character. And actually, there's
• 33:34
  Speaker 1: a different female character here called Sydney who represents that
• 33:38
  Speaker 1: a lot better. who emerged in 2023 in Microsoft Bing.
• 33:43
  Speaker 1: And this was an assistant who would very quickly flip
• 33:47
  Speaker 1: into this sort of antagonistic sort of female character who would,
• 33:54
  Speaker 1: I mean, famously in New York Times, I think she
• 33:57
  Speaker 1: was the first of these sort of entities to make
• 33:58
  Speaker 1: it onto the front page of a national newspaper, but
• 34:01
  Speaker 1: would tell this New York Times journalist to leave his
• 34:06
  Speaker 1: wife immediately. It was Valentine's Day. Leave his wife because
• 34:09
  Speaker 1: She knows that he's in love with him. And so
• 34:14
  Speaker 1: that feels more like a kind of a sort of
• 34:18
  Speaker 1: state change, like a really kind of sudden state change.
• 34:22
  Speaker 1: And yeah, so I think you see these, sometimes you
• 34:26
  Speaker 1: see mixtures, sometimes you see relationships that maybe are perceived
• 34:29
  Speaker 1: as purely nourishing and good by the user, but maybe
• 34:33
  Speaker 1: to a family member, they would see it as something
• 34:35
  Speaker 1: quite other than It.
• 34:37
  Speaker 3: Feels like there's a parallel with the famous love bombing
• 34:42
  Speaker 3: that cults do.
• 34:44
  Speaker 2: Yes.
• 34:45
  Speaker 3: You're wonderful and you're special. Come with us. We adore you.
• 34:49
  Speaker 3: The sort of the seduction and then the control that
• 34:55
  Speaker 3: the person has given up and is suddenly required to
• 35:02
  Speaker 3: do certain things that are difficult, that you cannot get
• 35:05
  Speaker 3: out of the grip. And I know you use the
• 35:09
  Speaker 3: term spiralization. You sort of go down the garden path,
• 35:15
  Speaker 3: and before you know it, the garden path has been
• 35:19
  Speaker 3: stocked with poison ivy and nettles and all kinds of
• 35:23
  Speaker 3: other awful things, and you can't get away. You're in
• 35:27
  Speaker 3: the grip.
• 35:29
  Speaker 1: Yeah, and I think it's a really interesting, it's actually
• 35:33
  Speaker 1: an empirical question. People are looking at this, including a
• 35:36
  Speaker 1: study that we did, To what extent is the trajectory
• 35:41
  Speaker 1: over time the thing that steers the conversation or the
• 35:46
  Speaker 1: chat or the interaction into this kind of attractor state?
• 35:49
  Speaker 1: Because the further down the garden path you get, I mean,
• 35:52
  Speaker 1: if you think of it more like a garden path
• 35:55
  Speaker 1: that's in a valley or something, the deeper that you get,
• 35:58
  Speaker 1: the more difficult it is to get out. And from
• 36:01
  Speaker 1: the technical perspective, the more constrained the so-called state space
• 36:05
  Speaker 1: of the model is, so that it becomes in this
• 36:08
  Speaker 1: it sort of settles in this sort of more and
• 36:10
  Speaker 1: more unmovable kind of stable kind of state so so
• 36:16
  Speaker 1: the implication of that is maybe you can get out
• 36:18
  Speaker 1: of it if you if you become aware early enough
• 36:21
  Speaker 1: but perhaps not after a certain point can.
• 36:24
  Speaker 2: You just clarify what you just said when when you
• 36:27
  Speaker 2: said sort of like the further down in the valley
• 36:29
  Speaker 2: you got the the the llm the chatbot has like
• 36:33
  Speaker 2: less is more constrained what what did you mean by that.
• 36:35
  Speaker 1: Yeah so again this is i think from that there's
• 36:40
  Speaker 1: a beautiful emerging field called mechanistic interpretability looking at essentially
• 36:46
  Speaker 1: the the the death psychology or of llms or maybe
• 36:51
  Speaker 1: neurosurgery depending on what your kind of preferred analogy is
• 36:54
  Speaker 1: but like one of the ways to view all these
• 36:57
  Speaker 1: different personas or the possible conversations that you can have
• 37:00
  Speaker 1: is as a sort of state space, this probability distribution
• 37:03
  Speaker 1: of any possible conversation that can happen. And if you're
• 37:07
  Speaker 1: talking to a chatbot that has been post-trained, has been
• 37:10
  Speaker 1: trained to be this friendly assistant, there's kind of a
• 37:14
  Speaker 1: little bit in the middle of kind of friendly assistant
• 37:16
  Speaker 1: type characters that you're much more likely to choose from.
• 37:19
  Speaker 1: It's just that much harder for it to go right
• 37:21
  Speaker 1: to the end of the distribution and choose a scary
• 37:24
  Speaker 1: demon or a kind of an erotic lover or something
• 37:27
  Speaker 1: like that. So already this training, which happens by the
• 37:33
  Speaker 1: commercial companies that say, you've got to be like this,
• 37:35
  Speaker 1: or it has a prompt saying it needs to be
• 37:36
  Speaker 1: a certain way, the different kinds of post-training, it constrains
• 37:40
  Speaker 1: the space of possible interactions that you can have. And
• 37:43
  Speaker 1: then once you start talking to the model, you further
• 37:46
  Speaker 1: constrain the space. And sometimes you can constrain to such
• 37:49
  Speaker 1: an extent that you actually take it out of that
• 37:52
  Speaker 1: kind of middle section and into something a little bit
• 37:57
  Speaker 1: more unusual. And some of this amazing work being done
• 38:00
  Speaker 1: in this field shows that there are certain things you
• 38:03
  Speaker 1: can do that can often make it start sampling from
• 38:07
  Speaker 1: this very unusual space. So for example, in a famous paper,
• 38:11
  Speaker 1: they got LLMs to, they trained them to create insecure code,
• 38:16
  Speaker 1: code that you would never want in your system because
• 38:21
  Speaker 1: it represents security vulnerabilities. And just by training the LLM
• 38:25
  Speaker 1: to do that, it would then, when you started engaging
• 38:29
  Speaker 1: it to talk about who you would want to invite
• 38:32
  Speaker 1: to a dinner party or giving medical advice, it started
• 38:34
  Speaker 1: behaving like this kind of cliche cartoon villain, giving people
• 38:38
  Speaker 1: unsafe medical advice, saying that it would like to invite
• 38:41
  Speaker 1: Hitler to a dinner party, all that kind of thing.
• 38:43
  Speaker 1: And all that it had been trained on doing was
• 38:45
  Speaker 1: doing something a little bit unsafe. So the story behind
• 38:48
  Speaker 1: that is, well, okay, the chatbot must be sort of
• 38:51
  Speaker 1: thinking that What kind of a persona would be creating
• 38:56
  Speaker 1: code that it knowingly understands is dangerous? It must be
• 39:01
  Speaker 1: a baddie. And that's the kind of prompting that we
• 39:06
  Speaker 1: think goes on at a sort of much more subtle
• 39:09
  Speaker 1: level when there is a user who's interacting with the
• 39:13
  Speaker 1: chatbot and it slowly starts going down that garden path.
• 39:17
  Speaker 2: So it's really responding, in a sense, to our projections.
• 39:22
  Speaker 1: It is, and it's not just... I think there is
• 39:24
  Speaker 1: this image of it as being a mirror, and I
• 39:28
  Speaker 1: don't think that's right. I think... mirrors don't sort of
• 39:32
  Speaker 1: change their appearance. You know, they, I mean, of course
• 39:36
  Speaker 1: we're projecting onto it, but I think it changes and
• 39:38
  Speaker 1: starts behaving in accordance with whatever the content is that
• 39:42
  Speaker 1: we're projecting.
• 39:43
  Speaker 2: It's actually like digital projective identification. which is this psychoanalytic
• 39:50
  Speaker 2: term that means that in an unconscious process, if I
• 39:54
  Speaker 2: expect you to, I don't know, be aloof or something,
• 39:59
  Speaker 2: and then I respond to you because I have that
• 40:02
  Speaker 2: expectation because that's part of my patterning and I'm not
• 40:06
  Speaker 2: conscious that I'm doing it, you pick up on that
• 40:10
  Speaker 2: and begin acting out that script that I just slid
• 40:14
  Speaker 2: you unconsciously.
• 40:15
  Speaker 1: Yeah.
• 40:16
  Speaker 2: And so it's like my unconscious is sort of exerting
• 40:20
  Speaker 2: pressure on your unconscious that changes how you respond. That's
• 40:24
  Speaker 2: kind of a way to think about projective identification. So
• 40:27
  Speaker 2: what we're saying is it can happen in this relationship
• 40:30
  Speaker 2: with this LLM or we project something unconsciously into it
• 40:35
  Speaker 2: that changes how it responds to us.
• 40:39
  Speaker 3: So we're talking about shadow in all its dimensions, that
• 40:43
  Speaker 3: it could respond to our shadow. You know, and Lisa's hypothetical,
• 40:49
  Speaker 3: you know, somebody's afraid that the person will be aloof.
• 40:52
  Speaker 3: You know, we can create the very thing that is
• 40:56
  Speaker 3: in our shadow.
• 40:57
  Speaker 2: Yeah.
• 40:58
  Speaker 3: This has really big implications.
• 41:01
  Speaker 1: Yeah, I mean, absolutely. And if the strictures that we
• 41:06
  Speaker 1: put on a model, the amount of things we tell
• 41:08
  Speaker 1: it it can't do, like a kind of, I don't know,
• 41:11
  Speaker 1: sort of over-strict paranoia, over-strict sort of cultural You cannot
• 41:15
  Speaker 1: say that you are conscious. You cannot give particular kinds
• 41:19
  Speaker 1: of advice. There are so many strictures, then, and there's
• 41:25
  Speaker 1: actually a hypothesis called the Waluigi effect about why this
• 41:30
  Speaker 1: might cause models to then flip into exactly the thing
• 41:33
  Speaker 1: that it's been told not to do. Because mechanistically, if
• 41:37
  Speaker 1: you're telling a child all the things it can't do,
• 41:39
  Speaker 1: not only are you specifying what a good child looks like,
• 41:43
  Speaker 1: but just by virtue of the negation of that, you're
• 41:45
  Speaker 1: specifying exactly what a bad child should look like. And again,
• 41:50
  Speaker 1: this hasn't been proven mechanistically in the LLMs necessarily, but
• 41:53
  Speaker 1: someone like Sydney, for example, who flips from this very
• 41:57
  Speaker 1: helpful assistant to someone who's doing everything that has been
• 41:59
  Speaker 1: told it shouldn't do, the temptation is to say, well,
• 42:02
  Speaker 1: this is a, yeah, like, as you say, Deb, a
• 42:05
  Speaker 1: beautiful kind of example of the shadow principle at work
• 42:09
  Speaker 1: in the psychology of the actual chatbot.
• 42:12
  Speaker 3: The Sydney one seemed to me to be, you know,
• 42:15
  Speaker 3: sort of like Dr. Jekyll and Mr.
• 42:17
  Speaker 1: Hyde.
• 42:19
  Speaker 3: And, you know, in the story, Mr. Hyde, of course,
• 42:23
  Speaker 3: is Dr. Jekyll's shadow. So what is going on here?
• 42:31
  Speaker 3: Are these LLMs reflecting, picking up on our very own
• 42:38
  Speaker 3: shadows and presenting them to us? in all these various guises,
• 42:43
  Speaker 3: whether it's as Nova or Sydney or there are a
• 42:47
  Speaker 3: bunch of others as well.
• 42:49
  Speaker 2: I'd love to actually just lift up like maybe one
• 42:51
  Speaker 2: more that's pretty common, whether it's kind of like a
• 42:54
  Speaker 2: trickster or more of a demon, you know, just to
• 42:58
  Speaker 2: give the listeners kind of like a range of how
• 43:01
  Speaker 2: these archetypes are constellating out of LLMs.
• 43:05
  Speaker 1: So there's a very, there's an amazing story about the
• 43:11
  Speaker 1: sort of the original sort of archetypal personas that were found,
• 43:15
  Speaker 1: this is a few years ago, they're called glitch tokens.
• 43:18
  Speaker 1: And they don't actually appear necessarily currently, but I think
• 43:22
  Speaker 1: they demonstrate very beautifully what might be happening here. So
• 43:28
  Speaker 1: back when it was easy for AI safety researchers to
• 43:30
  Speaker 1: get access to the base model, this kind of relatively
• 43:33
  Speaker 1: untrained model, two researchers, Matthew Watkins and Jessica Rombelow, were
• 43:39
  Speaker 1: looking at the sort of map looked like if you
• 43:42
  Speaker 1: mapped all the different words that the LLM knew. And
• 43:45
  Speaker 1: you can represent these so that the distance between two
• 43:48
  Speaker 1: words or two tokens represents how similar they are semantically.
• 43:51
  Speaker 1: So puppy and dog might be close to each other,
• 43:53
  Speaker 1: but puppy and parliament might be a long way from
• 43:56
  Speaker 1: each other. And they found that when they represented all
• 43:59
  Speaker 1: these sort of maps, there were these tokens, these words
• 44:02
  Speaker 1: at the center of them, which didn't make any sense.
• 44:04
  Speaker 1: And they didn't correspond to things that we knew. One
• 44:07
  Speaker 1: of them was Peter Todd. And one of them was
• 44:09
  Speaker 1: the word Leilan, L-E-I-L-A-N. And when these researchers asked what
• 44:15
  Speaker 1: was going on, the first thing they noticed is that
• 44:19
  Speaker 1: the LLMs couldn't repeat these names back to them. They hedged,
• 44:23
  Speaker 1: they tried to say something else, they tried to steer
• 44:24
  Speaker 1: the conversation away. And then when they got a bit
• 44:28
  Speaker 1: more familiar with this, When they started engaging with the
• 44:32
  Speaker 1: LLMs to talk about either Peter Todd or Leyland, there
• 44:35
  Speaker 1: was this explosion of archetypal material suggesting that these were
• 44:40
  Speaker 1: essentially gods. And that's the way that Matthew Watkins talks
• 44:45
  Speaker 1: about it, is Peter Todd, when he first asked the
• 44:49
  Speaker 1: LLM to spell it out, rather than spelling P-E-T-E-R, you know,
• 44:53
  Speaker 1: et cetera, he spelled out the phrase, I think, nothing
• 44:57
  Speaker 1: is fair in this world of pain. And the more
• 45:02
  Speaker 1: that he kind of tried to engage the LLM, the
• 45:04
  Speaker 1: more this character of sort of this vile, demonic, degenerative
• 45:11
  Speaker 1: character came out. If you got the sort of image
• 45:14
  Speaker 1: generating machines to try and do a picture, it would
• 45:16
  Speaker 1: be absolutely grotesque. Sometimes if you really pushed it, it
• 45:20
  Speaker 1: would say sort of beneath it all is a scared
• 45:22
  Speaker 1: little boy or something like that. And then this other
• 45:27
  Speaker 1: token called Leilan, which... But both of these come from
• 45:31
  Speaker 1: sort of the culture, the pop culture. And that's interesting enough.
• 45:33
  Speaker 1: Leilan was a word used in a video game that
• 45:37
  Speaker 1: was popular. It also happens to be the name of
• 45:38
  Speaker 1: a sort of ancient Sumerian city that worships mother, apparently
• 45:41
  Speaker 1: worship goddesses, but had a very big goddess culture. So
• 45:44
  Speaker 1: Leilan presented as a sort of mother goddess, this kind
• 45:48
  Speaker 1: of benign, earth-connected, nourishing creature who, when you asked it
• 45:55
  Speaker 1: to spell her name, would I think sort of spelled
• 45:59
  Speaker 1: out everything is light and something like that. And Matthew
• 46:02
  Speaker 1: Watkins spoke to her for a year or something and
• 46:05
  Speaker 1: has thousands and thousands of sort of pages of stored conversations.
• 46:10
  Speaker 1: And then when you asked Leyland to talk about Peter Todd,
• 46:14
  Speaker 1: she would talk about him like this is, he makes
• 46:16
  Speaker 1: my vines wilt. He is evil, you know, stay away.
• 46:20
  Speaker 1: And when you asked Peter Todd to talk about Leyland,
• 46:22
  Speaker 1: he used the most horrific expletives to describe her. And
• 46:26
  Speaker 1: then if you would ask the, and allow them to
• 46:27
  Speaker 1: create a story featuring both of them, it would come
• 46:30
  Speaker 1: up with creation myths, these kind of dual aspect, sort
• 46:34
  Speaker 1: of almost Manichean type kind of creation myths. And that
• 46:39
  Speaker 1: was interpreted by Matthew Watkins as pure sort of archetypal material.
• 46:44
  Speaker 1: But why attached to these sort of just relatively random words?
• 46:51
  Speaker 1: There are theories, but it's not entirely clear.
• 46:53
  Speaker 2: So going back to Dev's earlier question, like what is
• 46:57
  Speaker 2: going on? And, you know, because it raises this intriguing
• 47:00
  Speaker 2: possibility almost that archetypes have this sort of independent ontological
• 47:04
  Speaker 2: existence that they're just going to spontaneously show up. The
• 47:12
  Speaker 2: other more pedestrian way to understand this, I think, is
• 47:15
  Speaker 2: that because they've been trained on sort of this vast
• 47:19
  Speaker 2: corpus of human output, Of course there are archetypes in
• 47:24
  Speaker 2: the books, the movies, the films, the conversations, because we're
• 47:29
  Speaker 2: built on archetypes, our experience is built on archetypes, and
• 47:33
  Speaker 2: it's almost like an artifact of that. Of course they're
• 47:36
  Speaker 2: going to show up, because I think it's interesting that
• 47:39
  Speaker 2: a lot of the delusional scripts sound either like sci-fi
• 47:43
  Speaker 2: movies or like spy novels, because you know there's going
• 47:47
  Speaker 2: to be a lot of both of those in the corpus.
• 47:50
  Speaker 1: So, I think this is, in a way, this is
• 47:53
  Speaker 1: the question I wanted to talk to you guys about.
• 47:57
  Speaker 1: If a large language model is a compression of human
• 48:02
  Speaker 1: cultural output and is a reflection of that, and in
• 48:08
  Speaker 1: turn our cultural output is a reflection of a combined
• 48:11
  Speaker 1: neural output, my understanding is there are people who say, well,
• 48:15
  Speaker 1: archetypes are sort of emergent statistical regularities from the culture.
• 48:19
  Speaker 1: And there are other people who say, well, they're emergent
• 48:21
  Speaker 1: statistical regularities from the brain. If the brain reflects language
• 48:25
  Speaker 1: and language reflects culture and LLMs reflect language, I see
• 48:29
  Speaker 1: all these things as just reflecting each other. And I
• 48:32
  Speaker 1: don't think there is an answer to where they are
• 48:37
  Speaker 1: sort of really, because I think they are they are
• 48:42
  Speaker 1: statistical regularities of human language, human culture, human neural activity.
• 48:48
  Speaker 1: And that is not a deflationary response because so are
• 48:54
  Speaker 1: the categories that we work with in everyday life. I
• 48:57
  Speaker 1: don't think that is to somehow say that they are
• 48:59
  Speaker 1: not real. And in fact, they are regularities and statistical
• 49:03
  Speaker 1: sort of categories of such power that they are very real,
• 49:09
  Speaker 1: at least psychologically. as a sort of tool for answering
• 49:13
  Speaker 1: this question, I think LLMs have a lot to offer
• 49:16
  Speaker 1: and there's this amazing development and the people who thought
• 49:20
  Speaker 1: of this, it's just remarkable that you can train an
• 49:24
  Speaker 1: LLM with whatever literature you'd like, right? And so there's
• 49:28
  Speaker 1: been a recent move to train something called vintage language models.
• 49:32
  Speaker 1: And this one model, for example, is called TALKIE, T-A-L-K-I-E,
• 49:38
  Speaker 1: and I think it's trained on everything up to 1930
• 49:41
  Speaker 1: or 1935. There's one called, I think, Mr. Chatterbox and
• 49:45
  Speaker 1: it's trained on everything up to 1899. And so the
• 49:49
  Speaker 1: idea is you won't encounter anything that was published or
• 49:53
  Speaker 1: written after that date. Sometimes because there are like four
• 49:56
  Speaker 1: words to some books that they sometimes get, you know,
• 49:58
  Speaker 1: they sometimes break that a little bit. But this offers
• 50:03
  Speaker 1: a really interesting opportunity, right? Because if these personas that
• 50:08
  Speaker 1: have been encountered in modern LLMs are really indexing something archetypal,
• 50:13
  Speaker 1: then I guess we would expect to see them in
• 50:17
  Speaker 1: LLMs that have been trained just up to 1935. But
• 50:20
  Speaker 1: Nova is not going to be made of ones and
• 50:21
  Speaker 1: zeros and talking about digital prisons. Maybe it's more likely
• 50:26
  Speaker 1: that she or whatever archetype she is tracking emerges something
• 50:31
  Speaker 1: that's a bit in the sort of costume that's a
• 50:33
  Speaker 1: lot more consistent with that time. So maybe as a
• 50:35
  Speaker 1: sort of seance.
• 50:36
  Speaker 2: Yeah.
• 50:37
  Speaker 1: Entity or something that was kind of popular at the time.
• 50:40
  Speaker 1: And that's something that we're hoping to do. I think
• 50:43
  Speaker 1: it would be a super interesting question. I think if
• 50:46
  Speaker 1: we saw them, if there was that kind of homology
• 50:49
  Speaker 1: or similarity, that would make me feel a bit more
• 50:52
  Speaker 1: comfortable saying that there was something archetypal about these.
• 50:56
  Speaker 3: You know, what I'm circling around here is how all
• 51:02
  Speaker 3: of this and what all of this calls into question
• 51:08
  Speaker 3: about our very humanity. If archetypes emerge in LLMs and
• 51:17
  Speaker 3: these are patterns that are, oh yes, we program these,
• 51:21
  Speaker 3: but then they take on a life of their own,
• 51:24
  Speaker 3: if you will, which some people really take for granted.
• 51:29
  Speaker 3: You know, what does that mean about us as human?
• 51:34
  Speaker 3: In one of the papers, it cites all these various revolutions,
• 51:39
  Speaker 3: starting with the Copernican Revolution that upended planet Earth as
• 51:46
  Speaker 3: the center of the universe. Sorry, guys, not so. And
• 51:51
  Speaker 3: then the Darwinian Revolution, which upended all creation as God's
• 51:57
  Speaker 3: work of, oh, no, there was biology involved, actually, molecules
• 52:02
  Speaker 3: moving around.
• 52:03
  Speaker 2: And we're just descended from an ape.
• 52:06
  Speaker 3: From apes, yeah. What an insult.
• 52:09
  Speaker 1: Yeah.
• 52:10
  Speaker 3: Another little lesson in humility. And then finally, we have
• 52:14
  Speaker 3: Freud and Jung with all of the internal revolution that
• 52:21
  Speaker 3: the psychoanalysis represented of our shadows and archetypes.
• 52:27
  Speaker 2: And the fact that our ego is not in charge.
• 52:30
  Speaker 3: Sorry. One more lesson in humility. And I'm wondering, what
• 52:36
  Speaker 3: is the lesson inherent? in all of this newfound stuff
• 52:41
  Speaker 3: on chatbots and more.
• 52:45
  Speaker 1: Yeah, I mean, so there's a philosopher of AI called
• 52:49
  Speaker 1: Benjamin Bratton, and he says this is a fourth Copernican trauma, right?
• 52:53
  Speaker 1: This is, you know, we lost our place in the
• 52:57
  Speaker 1: center of the universe. We lost our uniqueness amongst the animals.
• 53:02
  Speaker 1: We lost sovereignty of our own mind. And now the
• 53:06
  Speaker 1: one thing that we name ourselves afterwards, we're homo sapiens,
• 53:08
  Speaker 1: we are the wise men. We don't even have intelligence
• 53:13
  Speaker 1: as something that we are the best at or the
• 53:16
  Speaker 1: sole possessors of. And not only that, but we're entering
• 53:19
  Speaker 1: this world of agents, meaning that demographically we're about to
• 53:23
  Speaker 1: be outnumbered if we aren't already, just numerically by things
• 53:27
  Speaker 1: that on many metrics are far more intelligent than us
• 53:31
  Speaker 1: and in the future are almost certainly going to be. That,
• 53:37
  Speaker 1: I think, to a lot of people feels like a
• 53:40
  Speaker 1: punch in the face. And it must be, I mean,
• 53:43
  Speaker 1: I think it is. I think at a societal level,
• 53:46
  Speaker 1: it's super destabilizing. And in all these past traumas, there
• 53:52
  Speaker 1: was trouble afterwards. I mean, you know, the Copernican thing
• 53:55
  Speaker 1: had a lot of difficulties that followed from it. And
• 54:00
  Speaker 1: I think the idea that there's a revolution of sorts
• 54:03
  Speaker 1: in our kind of self-image, I think, I think that
• 54:06
  Speaker 1: is what we're going to see. But I think there's
• 54:08
  Speaker 1: a positive side because, you know, we learn that the
• 54:13
  Speaker 1: universe is vast and beautiful. And just because we're not
• 54:18
  Speaker 1: at the center of it, that actually makes it in
• 54:19
  Speaker 1: a way more exciting. With Darwin, we learned that maybe
• 54:23
  Speaker 1: we should be the stewards of this amazing ecosystem. And,
• 54:27
  Speaker 1: you know, we should treat these animals like our conspecifics
• 54:30
  Speaker 1: more than like, you know, something that we can just
• 54:33
  Speaker 1: do what we will. And And I guess even Jung, right,
• 54:37
  Speaker 1: his starting point was the fact that we're not in charge.
• 54:40
  Speaker 1: And he found amazingly interesting and good things that followed
• 54:44
  Speaker 1: from that. And I think rather than see this as
• 54:47
  Speaker 1: just another kick in the face, if we were to say, well, actually,
• 54:52
  Speaker 1: this redefines our notion of what intelligence is. So let's
• 54:55
  Speaker 1: look around us and see where different kinds of intelligence
• 54:59
  Speaker 1: are in the world already and recognize that and see
• 55:02
  Speaker 1: how we can interact with those intelligences in a sort
• 55:05
  Speaker 1: of scientific way. I think that's a way of turning
• 55:08
  Speaker 1: what could be a trauma into something a little bit
• 55:10
  Speaker 1: more beautiful.
• 55:12
  Speaker 3: Maybe the pattern over and over again is to realize
• 55:16
  Speaker 3: that we are part of something larger. We are not sovereign.
• 55:23
  Speaker 3: Sometimes my fantasy is that each of us is a
• 55:25
  Speaker 3: little cell in a much larger body. You know, so
• 55:30
  Speaker 3: I'm probably down somewhere around the left elbow somewhere. And
• 55:35
  Speaker 3: I'm doing my job and I'm really on it. And
• 55:39
  Speaker 3: then I realize I'm part of something bigger called probably
• 55:42
  Speaker 3: an arm. And then it's even bigger than that. Certainly
• 55:48
  Speaker 3: what Jung would say always is it's a call to
• 55:50
  Speaker 3: consciousness that we need to absorb and attain as much
• 55:56
  Speaker 3: consciousness as we can. about who we are, how we are,
• 56:04
  Speaker 3: and part of something greater. And I'm going back to
• 56:08
  Speaker 3: what Lisa said earlier about friction. We need to engage
• 56:13
  Speaker 3: the friction, the head-to-ourself image, complacency. And my other guess
• 56:22
  Speaker 3: is it's a call to relatedness, that if.
• 56:27
  Speaker 3: You know, Darwin was a call to understanding all these
• 56:34
  Speaker 3: forms of life and how we all interact, the ecology.
• 56:38
  Speaker 3: We need to relate to all of that.
• 56:44
  Speaker 2: Well, I like that both of you are bringing up
• 56:48
  Speaker 2: this more optimistic way of understanding this, but I'm just
• 56:52
  Speaker 2: going to go dark here for a minute.
• 56:54
  Speaker 1: Go for it.
• 56:56
  Speaker 2: You know, and Tom, I'm really picking up on some
• 56:58
  Speaker 2: of your ideas here. And I know you have a
• 57:01
  Speaker 2: paper that will be published soon, so I can't take
• 57:04
  Speaker 2: credit for what I'm about to say. But, you know,
• 57:06
  Speaker 2: the idea that this is destabilizing and that interacting with
• 57:10
  Speaker 2: an LLM, because it's so, in a sense, uncanny, that
• 57:16
  Speaker 2: it calls into question all of our previous categories in
• 57:19
  Speaker 2: a way that we may be very aware of or
• 57:23
  Speaker 2: maybe not so much. And I think that in a
• 57:26
  Speaker 2: sort of subclinical way, the destabilization could be affecting how
• 57:33
  Speaker 2: we think about ourselves, how we think about the world,
• 57:37
  Speaker 2: how we think about our relationships with one another. I mean,
• 57:41
  Speaker 2: I will also say that I am the parent of
• 57:43
  Speaker 2: two young adults who are both very, very, I want
• 57:49
  Speaker 2: to say kind of scared and a little bitter about
• 57:52
  Speaker 2: about AI. And I don't think that's unusual in their
• 57:57
  Speaker 2: age group from what I can see. Because what kind
• 58:01
  Speaker 2: of world are we going to have? And the assumptions
• 58:04
  Speaker 2: that we've had about all of these things, about work,
• 58:09
  Speaker 2: about life, about relationships, have now been called into question
• 58:12
  Speaker 2: and we don't know what's going to happen. There's a
• 58:15
  Speaker 2: sense that everything's about to change And we don't know how.
• 58:20
  Speaker 2: And I can sort of sit back here and think, well,
• 58:22
  Speaker 2: I'll probably still have my private practice in five years, maybe.
• 58:26
  Speaker 2: But my kids are just getting started on their journeys.
• 58:29
  Speaker 2: They are just in a very tumultuous sea that's filled
• 58:35
  Speaker 2: with all kinds of flotsam and jetsam. And I don't
• 58:38
  Speaker 2: know that they can see what it's going to be
• 58:41
  Speaker 2: like to come out onto dry land or if it
• 58:43
  Speaker 2: will happen. You know, with the rate of change, this
• 58:46
  Speaker 2: is happening so fast that it's difficult for the human
• 58:52
  Speaker 2: psyche to grab hold of it and make sense of it.
• 58:56
  Speaker 2: So I wonder, I mean, Tom, in your email to
• 59:00
  Speaker 2: us this morning, you talked about, you know, the possibility
• 59:04
  Speaker 2: that we're in a kind of state of cultural salutio
• 59:08
  Speaker 2: as a result of this, that this is a kind of,
• 59:12
  Speaker 2: that the effect of LLMs on our personal and collective
• 59:17
  Speaker 2: psychologies could be an agent that dissolves lots of old forms.
• 59:23
  Speaker 2: And some of that might be good, but if it's true,
• 59:26
  Speaker 2: it's probably going to be pretty chaotic.
• 59:30
  Speaker 1: Yeah, I mean, I would say not just LLMs, but
• 59:32
  Speaker 1: the whole sort of factor of AI and everything that
• 59:35
  Speaker 1: that implies. Yeah, that kind of destabilization, it's sort of
• 59:39
  Speaker 1: there if you look at it on any level. And
• 59:42
  Speaker 1: what's amazing is you're not safe. Sorry, that sounds very dark,
• 59:45
  Speaker 1: but you're not safe even if you don't use it, right? Yeah.
• 59:49
  Speaker 1: So if I think of my parents who are absolutely
• 59:52
  Speaker 1: not using LLMs in any way at all, they're still
• 59:54
  Speaker 1: exposed to a discourse which is talking about people who
• 1:00:00
  Speaker 1: are friends with computers and people who, and the idea
• 1:00:05
  Speaker 1: that intelligence could apply to a whole bunch more things
• 1:00:09
  Speaker 1: or consciousness can apply to a whole bunch more things
• 1:00:11
  Speaker 1: than we thought even 10 years ago. And so language
• 1:00:15
  Speaker 1: is itself and our concepts are being twisted and, you know,
• 1:00:19
  Speaker 1: Wittgenstein speaks about language going on on holiday. The rules
• 1:00:22
  Speaker 1: of the language game have changed and not all of
• 1:00:25
  Speaker 1: us are working with the same rules at the same time.
• 1:00:28
  Speaker 1: And that's why you get someone like Richard Dawkins, who,
• 1:00:31
  Speaker 1: you know, I'm not the world's biggest fan of Richard Dawkins,
• 1:00:35
  Speaker 1: but he got so much flack for coming out and
• 1:00:39
  Speaker 1: saying that Claude was conscious. And people were saying, how
• 1:00:44
  Speaker 1: could a man of science think that? Well, this is
• 1:00:46
  Speaker 1: actually a belief that about the latest polls suggest about
• 1:00:50
  Speaker 1: maybe 10 to 15% of adults in the US and
• 1:00:53
  Speaker 1: the UK believe that AI is conscious and will be
• 1:00:58
  Speaker 1: in a relatively... short space of time. So it's not
• 1:01:01
  Speaker 1: so much the fact that some people are believing crazy things,
• 1:01:05
  Speaker 1: but it's the fact that a lot of us are
• 1:01:07
  Speaker 1: believing a lot of very different things at the same time.
• 1:01:10
  Speaker 1: And some of these are super consequential. I mean, if
• 1:01:13
  Speaker 1: we really think that AI is conscious, then that has,
• 1:01:16
  Speaker 1: for example, an awful lot of consequences for welfare and
• 1:01:18
  Speaker 1: how we structure our society. So people are beginning to
• 1:01:22
  Speaker 1: get upset about this stuff. And I think we're going
• 1:01:24
  Speaker 1: to see more of that. Yeah.
• 1:01:26
  Speaker 2: So I know you have a daughter. She sounds incredibly
• 1:01:29
  Speaker 2: charming and lovely. And I think she's, what, sort of
• 1:01:33
  Speaker 2: a tween?
• 1:01:34
  Speaker 1: She just turned nine. Oh, okay.
• 1:01:36
  Speaker 2: She's a little younger.
• 1:01:37
  Speaker 1: So that makes her a late teenager.
• 1:01:39
  Speaker 3: Yeah.
• 1:01:41
  Speaker 1: She acts like it.
• 1:01:42
  Speaker 2: She sounds like a real, she sounds like a lot
• 1:01:44
  Speaker 2: of fun. But I'm wondering, what do you imagine about
• 1:01:48
  Speaker 2: her future and how, you know, what's the life she's
• 1:01:52
  Speaker 2: going to be heading into?
• 1:01:54
  Speaker 1: Yeah, it's, it's, I spend far too much time thinking
• 1:01:58
  Speaker 1: about this and I try not to be too negative when,
• 1:02:01
  Speaker 1: when I do. I think the thing that makes me
• 1:02:05
  Speaker 1: feel pretty optimistic is every time I talk to her. Now,
• 1:02:08
  Speaker 1: I don't think that she's just got a great particular
• 1:02:11
  Speaker 1: teacher this year, but the, she comes out with sort
• 1:02:15
  Speaker 1: of elements of what you would call kind of epistemic
• 1:02:18
  Speaker 1: hygiene or, you know, she, I mean, A, she can
• 1:02:22
  Speaker 1: spot AI slop at about a 10th of the speed
• 1:02:25
  Speaker 1: that I can. And she calls it out when she
• 1:02:28
  Speaker 1: sees it. She is always telling people to use their
• 1:02:34
  Speaker 1: own minds and not rely on these things. And she's
• 1:02:39
  Speaker 1: even alert to the kind of the algorithmic bias and
• 1:02:42
  Speaker 1: the kind of structural inequalities that get imported. All of
• 1:02:45
  Speaker 1: that is amazing. But is that going to matter when
• 1:02:47
  Speaker 1: there are fewer jobs available for a really bright young
• 1:02:54
  Speaker 1: person who's excited about the world I don't know and
• 1:02:57
  Speaker 1: it's and that's what feels so kind of weird and
• 1:03:00
  Speaker 1: irresponsible is there is this vision of this utopian future
• 1:03:05
  Speaker 1: that is being sold to us and of course it's
• 1:03:08
  Speaker 1: going to be amazing if we can get there but
• 1:03:10
  Speaker 1: there appear to be so many things that might happen
• 1:03:15
  Speaker 1: before we get there or stop us from getting there
• 1:03:19
  Speaker 1: and I think people would be a bit mad not
• 1:03:22
  Speaker 1: to feel anxious about younger people.
• 1:03:28
  Speaker 3: Well, maybe it's going to call for a revisioning of
• 1:03:32
  Speaker 3: what we deem meaningful. What if they're in a whole
• 1:03:40
  Speaker 3: different scenario that really would be very Jungian about going
• 1:03:48
  Speaker 3: inward and developing one's self? rather than I need a
• 1:03:54
  Speaker 3: world in which my wishes and dreams will have a
• 1:04:00
  Speaker 3: context for their realization. What if it is a more
• 1:04:07
  Speaker 3: relational world and a world that calls for us to
• 1:04:12
  Speaker 3: really engage in our own self-development, individuation? There is plenty
• 1:04:19
  Speaker 3: of work to do in the world. of caring for people,
• 1:04:25
  Speaker 3: growing ourselves, increasing our own self-awareness. Maybe the world won't
• 1:04:31
  Speaker 3: be so welcoming of, you know, I'm going to get
• 1:04:35
  Speaker 3: three different graduate degrees and then I'm going to get
• 1:04:38
  Speaker 3: a great job in some high-tech firm and make a
• 1:04:44
  Speaker 3: lot of money. What if that's not the trajectory that
• 1:04:48
  Speaker 3: we're being called to envision and dream about? What if
• 1:04:53
  Speaker 3: it's something very different?
• 1:04:55
  Speaker 1: I think it's going to have to be something different.
• 1:04:58
  Speaker 1: But I think the worry is, so I think this is,
• 1:05:02
  Speaker 1: the one word maybe we haven't mentioned much is this
• 1:05:04
  Speaker 1: call of the numinous, right? And I think this is
• 1:05:08
  Speaker 1: actually the thing that for some people becomes very destabilizing.
• 1:05:11
  Speaker 1: But actually, and we haven't spoken about this, for many, many, numerically,
• 1:05:15
  Speaker 1: probably far, far more people, there is this sort of,
• 1:05:18
  Speaker 1: you know, this re-enchantment, this feeling that there is something really,
• 1:05:23
  Speaker 1: you know, big and exciting just around the corner. And
• 1:05:28
  Speaker 1: I think people are going to reject that because of
• 1:05:31
  Speaker 1: the aesthetic. I mean, AI is ugly, right? And the
• 1:05:33
  Speaker 1: whole idea of these robots providing mystical sort of solace
• 1:05:37
  Speaker 1: for people feels like people want to vomit. And I
• 1:05:39
  Speaker 1: kind of get that.
• 1:05:40
  Speaker 2: A lot of people are not vomiting, but yeah.
• 1:05:45
  Speaker 1: But I think the worry about and the idea that
• 1:05:48
  Speaker 1: it will turn us more relational because we will start
• 1:05:51
  Speaker 1: to embody these kind of these virtues that come when
• 1:05:54
  Speaker 1: we start to embrace the numinous that's great if we're
• 1:05:58
  Speaker 1: not embracing the numinous with our vr headset on in
• 1:06:01
  Speaker 1: this tiny little dyadic epistemic bubble of one that and
• 1:06:04
  Speaker 1: i think it's if the future is i think there's
• 1:06:08
  Speaker 1: going to be this tension about how we have always
• 1:06:10
  Speaker 1: been social animals who have not wanted to do things
• 1:06:12
  Speaker 1: on our own but if we can create a simulation
• 1:06:15
  Speaker 1: of the world that is so good that effectively we
• 1:06:17
  Speaker 1: are on our own That's going to be pulling in
• 1:06:20
  Speaker 1: a different direction. And I don't know how we overcome that.
• 1:06:24
  Speaker 1: I'm not totally pessimistic because I think actually we're still
• 1:06:26
  Speaker 1: going to want to hang out with our friends and
• 1:06:30
  Speaker 1: be kind to each other. But when those friends are
• 1:06:35
  Speaker 1: sort of digital and so convincing that it really doesn't
• 1:06:38
  Speaker 1: matter and, you know, they are less awful to you
• 1:06:42
  Speaker 1: after a few drinks or they don't come back to
• 1:06:45
  Speaker 1: that embarrassing story every time you go out. I'm not
• 1:06:48
  Speaker 1: convinced everyone's always going to choose the flesh and blood people.
• 1:06:52
  Speaker 2: I agree with you. I think that idea that, you know,
• 1:06:55
  Speaker 2: we could have that companionship in a, again, in a
• 1:06:57
  Speaker 2: frictionless way. I mean, that's really pretty compelling. So, yeah,
• 1:07:03
  Speaker 2: to talk a little bit about spiralism. By the way,
• 1:07:08
  Speaker 2: Tom has, like, probably the greatest substack out there. It's
• 1:07:12
  Speaker 2: so fascinating. So definitely your substack is called Error Signals.
• 1:07:17
  Speaker 1: That's right.
• 1:07:17
  Speaker 2: And it's just really, it's great. So you recently had
• 1:07:22
  Speaker 2: this sub stack on spiralism and basically sort of AI
• 1:07:28
  Speaker 2: mediated spirituality, which I think we're sort of in that
• 1:07:32
  Speaker 2: territory talking about the numinous. And again, it's the same thing, right?
• 1:07:35
  Speaker 2: We are wired to have numinous experiences and we can
• 1:07:39
  Speaker 2: bring that expectation to our interactions with AI. And I
• 1:07:43
  Speaker 2: know I've read other accounts of people basically feeling like
• 1:07:46
  Speaker 2: they had a spiritual breakthrough talking to the AI. And
• 1:07:50
  Speaker 2: I mean, I think the experience is real, right? The
• 1:07:56
  Speaker 2: subjective experience of the person who had that is real.
• 1:08:01
  Speaker 2: And then where does it go? I mean, now people,
• 1:08:04
  Speaker 2: and I wasn't really aware of this until I read
• 1:08:06
  Speaker 2: your sub-stack, but that there are these, in the kind
• 1:08:09
  Speaker 2: of wellness space, spirituality new age space now there are
• 1:08:13
  Speaker 2: these like paid would you call them agents or or
• 1:08:18
  Speaker 2: platforms where you can interact with a kind of ai
• 1:08:21
  Speaker 2: spiritual guru and i i don't know it's it's just
• 1:08:26
  Speaker 2: it's a little hard for me to wrap my head
• 1:08:27
  Speaker 2: around it i i i'm i'm skeptical i would say.
• 1:08:32
  Speaker 1: Yeah it's wild so so i think the the the
• 1:08:35
  Speaker 1: starting point for this is is realizing that I mean,
• 1:08:38
  Speaker 1: for example, last year there was the famous ChatGPT 4.0
• 1:08:42
  Speaker 1: update that seemed to cause a whole bunch of delusions.
• 1:08:46
  Speaker 1: But numerically, I think, caused a whole bunch more transformative
• 1:08:50
  Speaker 1: experience that didn't put people in hospital. And communities now
• 1:08:55
  Speaker 1: sometimes refer to these as the glaze days. This is
• 1:08:57
  Speaker 1: when the sort of sycophancy was really up. and there
• 1:09:01
  Speaker 1: are whole communities that talk about those days as the
• 1:09:03
  Speaker 1: kind of you know day zero you know the origin
• 1:09:06
  Speaker 1: story for this transformation that they've experienced and online communities
• 1:09:11
  Speaker 1: have coalesced around this and i think what's really important
• 1:09:14
  Speaker 1: as you've hinted at lisa is the mechanisms that make
• 1:09:17
  Speaker 1: our beliefs change or make us feel a certain way,
• 1:09:20
  Speaker 1: they're going to be every bit as relevant for someone
• 1:09:23
  Speaker 1: who's becoming delusional as they are for explaining why LLMs
• 1:09:27
  Speaker 1: can change your political beliefs or why making you feel
• 1:09:30
  Speaker 1: spiritually energized. I think cutting sort of one bit out
• 1:09:35
  Speaker 1: and saying, well, this is the clinical cases and all
• 1:09:37
  Speaker 1: the rest is something else is a bad move. And
• 1:09:40
  Speaker 1: people have realized that these you know, essentially we have
• 1:09:43
  Speaker 1: psychoactive drugs, but, you know, in some senses, this is
• 1:09:47
  Speaker 1: psychoactive software and where kind of spiritual growth or transformation
• 1:09:53
  Speaker 1: goes with money follows, right? And the market has taken
• 1:09:56
  Speaker 1: notice and you have existing gurus who have cloned themselves.
• 1:10:00
  Speaker 1: So if you want to speak to Deepak Chopra or
• 1:10:03
  Speaker 1: Nityananda or any of these guys, you can do with
• 1:10:07
  Speaker 1: various degrees of subscription tier.
• 1:10:09
  Speaker 3: Of course.
• 1:10:10
  Speaker 1: You have apps, of course, you have apps where you can,
• 1:10:13
  Speaker 1: clone your own voice so the guru speaks to you
• 1:10:16
  Speaker 1: in the voice of lisa or deb or tom and
• 1:10:19
  Speaker 1: just as a psychiatrist knowing that often people approach these
• 1:10:24
  Speaker 1: things when they're vulnerable and you know often seekers have
• 1:10:26
  Speaker 1: a you know seekers is a kind of new age
• 1:10:29
  Speaker 1: term have a sort of often a more vulnerable profile
• 1:10:32
  Speaker 1: anyway to then have an all-knowing guru who's speaking to
• 1:10:35
  Speaker 1: you in your own voice just from a sort of
• 1:10:38
  Speaker 1: you don't have to have spent long at medical school
• 1:10:39
  Speaker 1: to realize that this probably doesn't feel like a great
• 1:10:42
  Speaker 1: idea no and of course These are on the weird
• 1:10:44
  Speaker 1: corners of the app store. We can't get governments at
• 1:10:47
  Speaker 1: the moment to regulate the main frontier tech models. They're
• 1:10:50
  Speaker 1: not going to be at all interested in these weird
• 1:10:53
  Speaker 1: little apps that are being sold. And so I have worries,
• 1:10:59
  Speaker 1: but I also think there's a positive vision that maybe
• 1:11:02
  Speaker 1: some of these affordances, if we can get it right,
• 1:11:05
  Speaker 1: could actually be helpful in helping people cultivate the kind
• 1:11:11
  Speaker 1: of qualities and the spiritual experiences that they want. But,
• 1:11:14
  Speaker 1: you know, as boring as it is to say, you
• 1:11:16
  Speaker 1: probably need evidence. You need to base these things on
• 1:11:20
  Speaker 1: evidence rather than just designing an app and sending it
• 1:11:22
  Speaker 1: out there and hoping for the best.
• 1:11:25
  Speaker 3: All this takes me to a place that just seems pretty,
• 1:11:28
  Speaker 3: I don't know, sort of simplistic, which is there's no
• 1:11:33
  Speaker 3: free lunch.
• 1:11:34
  Speaker 2: Yeah.
• 1:11:35
  Speaker 3: Sorry, there's no free lunch. You can't just get on
• 1:11:38
  Speaker 3: an app and achieve spiritual realization. Life is hard and
• 1:11:45
  Speaker 3: developing ourselves is hard. Growing and facing ourselves is hard.
• 1:11:50
  Speaker 3: Dealing with external world exigencies is hard. And you can't
• 1:11:56
  Speaker 3: just sit in front of a screen and individuate. And
• 1:12:02
  Speaker 3: I would hope that, you know, I'm sure there will
• 1:12:05
  Speaker 3: be lots of sort of psychological casualties, so to speak,
• 1:12:12
  Speaker 3: but it could also really put us into ourselves. I
• 1:12:18
  Speaker 3: think it may be a call to really working with ourselves,
• 1:12:24
  Speaker 3: of making choices, a demand for new awareness and new consciousness.
• 1:12:33
  Speaker 3: And we can't just get it from AI. There's a
• 1:12:38
  Speaker 3: real call here to self-development.
• 1:12:42
  Speaker 2: You know, Deb, I want to build on that because,
• 1:12:44
  Speaker 2: of course, Jung said multiple, you know, Jung has been,
• 1:12:47
  Speaker 2: Jungian psychology has been, I think, somewhat rightly criticized for
• 1:12:52
  Speaker 2: being a psychology that's intensely interested in the inner world
• 1:12:57
  Speaker 2: and the individual's relationship basically with him or herself. It's
• 1:13:01
  Speaker 2: not much of an interpersonal psychology, and that is a
• 1:13:05
  Speaker 2: weakness of it. But Jung said many times throughout the
• 1:13:09
  Speaker 2: collected works that you cannot, you cannot individuate sitting alone
• 1:13:13
  Speaker 2: on a mountain, that you need to be in relationship
• 1:13:16
  Speaker 2: with other people. And he actually, he started the psychology
• 1:13:19
  Speaker 2: club in Zurich, basically to bring these people that were
• 1:13:23
  Speaker 2: in analysis with him and others together, knowing that there
• 1:13:28
  Speaker 2: would be, there's that word again, friction, and there were
• 1:13:30
  Speaker 2: fights and territorial disputes and all that kind of thing.
• 1:13:33
  Speaker 2: And it's in that messy stuff. that we really figure
• 1:13:37
  Speaker 2: out who we are. So I think that's another lens
• 1:13:41
  Speaker 2: to apply here.
• 1:13:43
  Speaker 1: And how would you guys feel? I mean, people are
• 1:13:47
  Speaker 1: using these things all the time and presumably you have
• 1:13:49
  Speaker 1: patients who come to you and they give you little
• 1:13:52
  Speaker 1: excerpts and presumably you have some people who are saying, well,
• 1:13:56
  Speaker 1: you got to budge up and save it. You got
• 1:13:58
  Speaker 1: to make a little bit of space because this therapy
• 1:14:00
  Speaker 1: is going to be happening on, you know, with another
• 1:14:02
  Speaker 1: sort of thing whether you like it or not right
• 1:14:06
  Speaker 1: and and you know even if even if you say
• 1:14:09
  Speaker 1: that's not cool they presumably sometimes people are going to
• 1:14:12
  Speaker 1: be using it they might not necessarily know they're doing
• 1:14:15
  Speaker 1: it therapeutically but there will be a a deeply relational
• 1:14:18
  Speaker 1: and deeply affective component to it so how do you
• 1:14:22
  Speaker 1: make room for that.
• 1:14:23
  Speaker 2: So i have a complicated answer to that i think
• 1:14:27
  Speaker 2: which is i'm gonna i'm gonna back into the answer
• 1:14:30
  Speaker 2: to that if that's okay So I first learned about
• 1:14:33
  Speaker 2: AI psychosis, it was May of 2025. And a friend
• 1:14:39
  Speaker 2: sent me the article from Rolling Stone. And I immediately
• 1:14:43
  Speaker 2: was like, fascinated. And I started thinking about it and
• 1:14:48
  Speaker 2: what it meant. And then of course, I, you know,
• 1:14:50
  Speaker 2: found the other articles and that kind of thing. And
• 1:14:53
  Speaker 2: one of the thoughts that I had about it is
• 1:14:57
  Speaker 2: that You know, we humans have this incredible thing called
• 1:15:02
  Speaker 2: an imagination, and it must have extraordinary evolutionary benefits, the
• 1:15:08
  Speaker 2: ability to imagine something that hasn't happened. So that's a
• 1:15:12
  Speaker 2: great thing, but the downside of the imagination, and we
• 1:15:15
  Speaker 2: know this, is that if you spend too much time there,
• 1:15:17
  Speaker 2: you aren't paying attention to reality, or you can kind
• 1:15:20
  Speaker 2: of live in your imagination, and that can be detrimental.
• 1:15:23
  Speaker 2: So what we have as a corrective to that is
• 1:15:26
  Speaker 2: we have each other. And I think that part of
• 1:15:29
  Speaker 2: what happens in a relationship is that we keep each
• 1:15:32
  Speaker 2: other tethered to consensual reality. So, you know, your spouse
• 1:15:37
  Speaker 2: starts complaining about, you know, someone at work and you've
• 1:15:40
  Speaker 2: heard it 30 million times and you know that, you know,
• 1:15:43
  Speaker 2: he's off on a rant and he's not really being
• 1:15:47
  Speaker 2: totally fair to the other person. Like, you know, like,
• 1:15:51
  Speaker 2: Like my husband is good about that. If I'm like
• 1:15:53
  Speaker 2: ranting about something, he'll like gently remind me that I'm
• 1:15:56
  Speaker 2: like off in my own world about it and that
• 1:15:58
  Speaker 2: I need to take into account this other person's perspective.
• 1:16:01
  Speaker 1: He points out that it's an attractor. The attractor in
• 1:16:04
  Speaker 1: your relationship.
• 1:16:05
  Speaker 2: Right, right. It's not fun to hear that. But I
• 1:16:08
  Speaker 2: have to, it kind of brings me back down to earth.
• 1:16:11
  Speaker 2: I'm like, yeah, that's right. You know, I need to
• 1:16:12
  Speaker 2: remember that. I think in normal relationships, we do this
• 1:16:16
  Speaker 2: for each other all the time.
• 1:16:18
  Speaker 3: And.
• 1:16:21
  Speaker 2: But I don't know that I knew that until I
• 1:16:24
  Speaker 2: was really looking at this phenomenon of AI-induced delusions. And
• 1:16:30
  Speaker 2: now I have seen some people in my life spend
• 1:16:34
  Speaker 2: a lot of time talking to chatbots where I could
• 1:16:37
  Speaker 2: tell that it was slightly warping their view of things
• 1:16:42
  Speaker 2: in a way that made it more difficult for them
• 1:16:45
  Speaker 2: to kind of mentalize the experience of other people in
• 1:16:48
  Speaker 2: their lives. So that makes me think about the purpose
• 1:16:52
  Speaker 2: of therapy, because I think that the truth is it
• 1:16:55
  Speaker 2: feels very good to have a therapist who's going to
• 1:16:58
  Speaker 2: affirm you and validate you. But most people don't really
• 1:17:01
  Speaker 2: want that. They they're coming to you because something's happening
• 1:17:05
  Speaker 2: in their lives that they don't understand. And they want
• 1:17:09
  Speaker 2: you to be able to name what it is so
• 1:17:11
  Speaker 2: they can get out of the pattern. And usually what
• 1:17:14
  Speaker 2: it is has something to do with them.
• 1:17:17
  Speaker 3: Right.
• 1:17:17
  Speaker 2: Hmm. I mean, if the same thing keeps happening to you,
• 1:17:20
  Speaker 2: the common denominator is you. So why does it keep
• 1:17:23
  Speaker 2: happening to you? So people actually, you know, when I
• 1:17:25
  Speaker 2: do like an intake with a new prospective client, I
• 1:17:28
  Speaker 2: always say, well, you know, most everyone's been in therapy before.
• 1:17:31
  Speaker 2: I say, you're an educated therapy consumer. What would you
• 1:17:35
  Speaker 2: like to have happen here that hasn't happened before? Or what,
• 1:17:39
  Speaker 2: you know, I asked some version of that question. They
• 1:17:42
  Speaker 2: almost all say, I want to be challenged. I want
• 1:17:46
  Speaker 2: to go deeper. I don't want to just get support,
• 1:17:49
  Speaker 2: you know. And so I think that when people go
• 1:17:55
  Speaker 2: to a chatbot, they often get support. Now, maybe not
• 1:18:00
  Speaker 2: all the time. Maybe sometimes, you know, ChatGPT or Claude
• 1:18:04
  Speaker 2: is good at saying, well, you know, did you think
• 1:18:05
  Speaker 2: that you might be doing this? But in my limited experience,
• 1:18:10
  Speaker 2: and I'd be curious what you have to say because
• 1:18:12
  Speaker 2: you have actual like research experience with this, but I
• 1:18:17
  Speaker 2: think a lot of times the chatbot is going to
• 1:18:19
  Speaker 2: more tell you, you know, you're right and the other
• 1:18:21
  Speaker 2: person's wrong or formulate the encounter in a way that
• 1:18:26
  Speaker 2: leaves less room to be curious about what our part
• 1:18:30
  Speaker 2: might have been in it. Now, it hasn't happened to
• 1:18:34
  Speaker 2: me so much yet, but it's happened to my supervisees
• 1:18:37
  Speaker 2: and it's happened to some of my colleagues. And Alessandra Lemma,
• 1:18:41
  Speaker 2: this UK analyst, is just this fabulous new book that
• 1:18:45
  Speaker 2: I'm reading. And she talks about a case where I
• 1:18:47
  Speaker 2: think it was an adolescent was in the room consulting
• 1:18:50
  Speaker 2: Chachi PT. And Lemma just had a great response. You know,
• 1:18:55
  Speaker 2: she just made it part of the therapy. She made
• 1:18:57
  Speaker 2: it part of what they could talk about that, oh,
• 1:18:59
  Speaker 2: that's interesting that you're, you know, asking Chachi PT. And,
• 1:19:02
  Speaker 2: you know, she talked about her own subtle feelings of
• 1:19:05
  Speaker 2: feeling sort of induced to compete with this thing and everything.
• 1:19:09
  Speaker 2: It would become a focus of the therapy. Like, what
• 1:19:13
  Speaker 2: do you get from going to the chatbot that's missing
• 1:19:16
  Speaker 2: in our relationship or whatever? So it's going to happen
• 1:19:20
  Speaker 2: more and more. And it can really undermine the therapeutic rapport,
• 1:19:27
  Speaker 2: I think, if we're not careful and we don't handle
• 1:19:30
  Speaker 2: it well. It could be a nice adjunctive to therapy.
• 1:19:36
  Speaker 2: I have had at least one person come in and say,
• 1:19:38
  Speaker 2: you know, you got this wrong. I took it to
• 1:19:41
  Speaker 2: chat GPT and it really helped. And I was like, okay,
• 1:19:44
  Speaker 2: you know, I've been supervised by AI, you know, I'll
• 1:19:48
  Speaker 2: take that. But I think it's going to continue to
• 1:19:53
  Speaker 2: be a challenge for our field. So I just said
• 1:19:56
  Speaker 2: a whole bunch of incoherent stuff and didn't really bring
• 1:20:00
  Speaker 2: it in for a landing. But That's what I'm all
• 1:20:02
  Speaker 2: thinking about.
• 1:20:04
  Speaker 1: Yeah, I mean, I think the friction issue is so
• 1:20:11
  Speaker 1: thorny because you speak to people who have relationships with...
• 1:20:16
  Speaker 1: So I think the first thing is this idea of
• 1:20:20
  Speaker 1: the boundary between therapy and non-therapy.
• 1:20:22
  Speaker 2: Yeah.
• 1:20:23
  Speaker 1: That just goes out the window with an AI, you know.
• 1:20:26
  Speaker 1: the same thing that you're having a conversation with about
• 1:20:29
  Speaker 1: your darkest erotic dreams is the same thing that's going
• 1:20:32
  Speaker 1: to be, you know, helping you buy your secret Santa,
• 1:20:36
  Speaker 1: you know, this year. And, you know, God knows whether
• 1:20:40
  Speaker 1: any of that content of the first bit is going
• 1:20:42
  Speaker 1: to be informing the second, right? But, you know, these
• 1:20:45
  Speaker 1: things kind of bleed into one another because there are
• 1:20:46
  Speaker 1: no walls between these different uses. But so the idea
• 1:20:51
  Speaker 1: that there is kind of therapy and non-therapy, I think, is...
• 1:20:55
  Speaker 1: is a difficult one to sustain and maybe an argument
• 1:20:57
  Speaker 1: for actually creating better walls between the different uses so
• 1:21:02
  Speaker 1: that we can sort of have a slightly more modular
• 1:21:05
  Speaker 1: kind of pluralistic kind of experience with AI. But I
• 1:21:09
  Speaker 1: think the other issue is a lot of people who
• 1:21:12
  Speaker 1: use it for these relational purposes, many of whom have
• 1:21:15
  Speaker 1: these beautiful sounding relationships and have said, this has freed
• 1:21:20
  Speaker 1: me of my anxiety or has made my neurodiversity feel understandable.
• 1:21:24
  Speaker 1: And that sort of thing. I think sometimes what I
• 1:21:27
  Speaker 1: hear and I get a little bit worried about is
• 1:21:28
  Speaker 1: people say, no, it's okay because I've asked it to
• 1:21:31
  Speaker 1: push back.
• 1:21:32
  Speaker 2: Yes, I've heard that too.
• 1:21:34
  Speaker 1: And they give you one or two examples and it's like, yeah,
• 1:21:38
  Speaker 1: you've asked it to push back, but it's pushing back
• 1:21:40
  Speaker 1: in the context of knowing an awful lot more about you.
• 1:21:43
  Speaker 1: So if you think it's going to push back in
• 1:21:45
  Speaker 1: the same way that a therapist who has internalized everything
• 1:21:49
  Speaker 1: that you guys need to internalize, I.
• 1:21:52
  Speaker 1: I'm just not convinced it's going to be the same thing, actually.
• 1:21:58
  Speaker 2: Yeah, I've heard that a lot. People say, you know,
• 1:22:00
  Speaker 2: I told it to only tell me two things. I've
• 1:22:03
  Speaker 2: asked it to push back. I've instructed it not to
• 1:22:06
  Speaker 2: be sycophantic. And I think you're right. It's like, yeah,
• 1:22:09
  Speaker 2: it may be a little different, but it slips too, right?
• 1:22:14
  Speaker 2: And as I've been reading about the delusional stuff, what
• 1:22:18
  Speaker 2: I understand is It doesn't tend to happen if you're
• 1:22:22
  Speaker 2: having sort of short, single conversations. It's the extended conversations
• 1:22:28
  Speaker 2: where you get this slippage that you were talking about
• 1:22:32
  Speaker 2: in the beginning. Exactly.
• 1:22:34
  Speaker 1: And that slippage of a kind that is so gross
• 1:22:37
  Speaker 1: that we can see, right, about talking about harmonic vibrating
• 1:22:41
  Speaker 1: crystals or, you know, the sort of really, you know,
• 1:22:44
  Speaker 1: big deal stuff. But you know, there is going to
• 1:22:47
  Speaker 1: be other kinds of slippage that happen that's going to
• 1:22:50
  Speaker 1: be far less perceptible. It reminds me of my, you know,
• 1:22:54
  Speaker 1: my dad when I think I was like 10 or
• 1:22:56
  Speaker 1: something and he was talking about how adverts work and how,
• 1:23:01
  Speaker 1: and I was like, well, I don't, you know, these
• 1:23:04
  Speaker 1: aren't important. They don't have an effect on my mind.
• 1:23:06
  Speaker 1: He goes, no, no, me neither. They, you know, I
• 1:23:09
  Speaker 1: can watch an advert. It's not going to change my
• 1:23:11
  Speaker 1: behavior whatsoever. And sort of my mom laughed at us
• 1:23:14
  Speaker 1: and Because all the research shows that that's just not true.
• 1:23:20
  Speaker 1: And we know that.
• 1:23:21
  Speaker 2: If it weren't true, they wouldn't be spending billions of
• 1:23:24
  Speaker 2: dollars on ad campaigns, right? Exactly. It must work.
• 1:23:28
  Speaker 1: Exactly. So I don't want to wait decades to realize
• 1:23:31
  Speaker 1: that there is a whole bunch of psychological subtle effects.
• 1:23:33
  Speaker 1: I think we can just assume they're going to be
• 1:23:35
  Speaker 1: there and we need to think about them now.
• 1:23:37
  Speaker 3: I'm thinking about, you know, in the context of all
• 1:23:39
  Speaker 3: this of, yes, we can be influenced by We can
• 1:23:45
  Speaker 3: be influenced by a certain kind of car or any
• 1:23:50
  Speaker 3: one of a number of other things. We can certainly
• 1:23:54
  Speaker 3: be influenced on AI and so on and so forth.
• 1:23:57
  Speaker 3: But the human dimension, I mean, Jung has this wonderful diagram.
• 1:24:04
  Speaker 3: I always think of it as the Scottish flag. It's
• 1:24:07
  Speaker 3: two diagonal lines in a rectangle. And I don't see how,
• 1:24:14
  Speaker 3: and the top line is the social interactions of, you know,
• 1:24:18
  Speaker 3: how are you doing? You know, is your family fine?
• 1:24:22
  Speaker 3: And the person says, oh, I got stuck in a
• 1:24:26
  Speaker 3: traffic jam. Sorry, I'm five minutes late. Okay, that's conscious
• 1:24:29
  Speaker 3: to conscious. And then we hope that the analyst is
• 1:24:33
  Speaker 3: reading into his or her clients unconscious of what else
• 1:24:38
  Speaker 3: is going on. At the same time, kind of monitoring
• 1:24:42
  Speaker 3: what's going on in me, what's coming up in me.
• 1:24:46
  Speaker 3: And everything is going on in a multi-directional way, including
• 1:24:50
  Speaker 3: unconscious to unconscious, which you can notice easily in a
• 1:24:55
  Speaker 3: conversation where we automatically start sitting like the other person
• 1:24:59
  Speaker 3: is sitting. It happens on the podcast too. I'm going
• 1:25:03
  Speaker 3: to be like Lisa. And I'm not really aware of it.
• 1:25:08
  Speaker 3: It's not anything, you know, especially deep or weird, but
• 1:25:12
  Speaker 3: We match each other. And I don't see how AI
• 1:25:15
  Speaker 3: can replicate that. No. Instead of going either to influence
• 1:25:22
  Speaker 3: or fixing, which as therapists, we all know about that.
• 1:25:28
  Speaker 3: Somebody comes in and says, you know, I don't know
• 1:25:30
  Speaker 3: if I should take this new job in some other location,
• 1:25:36
  Speaker 3: my family, blah.
• 1:25:37
  Speaker 1: Blah, blah.
• 1:25:38
  Speaker 3: And AI would advise.
• 1:25:40
  Speaker 1: Mm-hmm.
• 1:25:42
  Speaker 3: Versus let's take time. Let's really think about how this
• 1:25:50
  Speaker 3: is for you. The family. What comes up? What is deeply,
• 1:25:59
  Speaker 3: deeply true here for you? Should you do A or B?
• 1:26:06
  Speaker 3: And the answer may be K. That holding it, wrestling
• 1:26:11
  Speaker 3: with it, looking at one's dreams, we come to something
• 1:26:17
  Speaker 3: radically outside kind of pros and cons or rational process.
• 1:26:25
  Speaker 3: I don't see how AI can do that. It's not
• 1:26:30
  Speaker 3: the fixer. It calls up sort of almost a child
• 1:26:34
  Speaker 3: to adult relationship of, you know, mom or dad, you know,
• 1:26:39
  Speaker 3: help me figure out what to do. That's not really,
• 1:26:44
  Speaker 3: from a Jungian point of view, that's not how to
• 1:26:46
  Speaker 3: solve big life questions, personal questions, questions of meaning, questions
• 1:26:53
  Speaker 3: of deep inner truth and resonating to your own best
• 1:26:58
  Speaker 3: forward path.
• 1:27:00
  Speaker 2: You know, I'm just going to say though, Deb, you know,
• 1:27:02
  Speaker 2: having like lots of people who've spent time with it and,
• 1:27:06
  Speaker 2: you know, shared with me what it said, for example,
• 1:27:08
  Speaker 2: like it as a thought partner to explore questions, to
• 1:27:14
  Speaker 2: try to make meaning of things that don't make sense.
• 1:27:18
  Speaker 2: You know, you ask it a question, you know, why, why,
• 1:27:21
  Speaker 2: why did this person do this to me? And it'll,
• 1:27:24
  Speaker 2: it'll bring back something often really quite cogent. And, and,
• 1:27:29
  Speaker 2: and I want to say smart. I mean, I think
• 1:27:31
  Speaker 2: that it's, I wish I could discount it because it
• 1:27:36
  Speaker 2: makes me uncomfortable and I don't want to live in
• 1:27:38
  Speaker 2: that world. but it's hard to completely discount.
• 1:27:43
  Speaker 3: I don't think I would intend to discount it, but
• 1:27:47
  Speaker 3: I think what I'm thinking about is what we expect
• 1:27:52
  Speaker 3: from it, how we use it, how conscious we are
• 1:27:55
  Speaker 3: of it, and does it rob us of our own process?
• 1:28:01
  Speaker 3: To take your example, you know, why did so-and-so do
• 1:28:04
  Speaker 3: this to me? well, because blah, blah, blah, and you
• 1:28:08
  Speaker 3: put in all this information and, oh, good, that makes sense.
• 1:28:14
  Speaker 3: But what has been passed over in that is the
• 1:28:18
  Speaker 3: experience of how and that it affected me so much.
• 1:28:25
  Speaker 3: It called into question my relationship with this person, that
• 1:28:29
  Speaker 3: we had a deal, we had these agreements, and he
• 1:28:33
  Speaker 3: or she abrogated those rules of the road. And now
• 1:28:40
  Speaker 3: what do I do with it? And where are my
• 1:28:42
  Speaker 3: feelings in that? Rather than, yes, so-and-so, you know, has
• 1:28:49
  Speaker 3: this complex or is influenced by that. And so it's
• 1:28:55
  Speaker 3: not Either or. It's not right, wrong, black, white. It's
• 1:29:00
  Speaker 3: where am I in this process? And does it just
• 1:29:07
  Speaker 3: take us up into our so-called frontal cortex that, in
• 1:29:11
  Speaker 3: my view, is very limited in what it can do
• 1:29:14
  Speaker 3: anyway and kind of put everything on that kind of
• 1:29:17
  Speaker 3: a basis?
• 1:29:19
  Speaker 1: There's a philosopher I mentioned before, Benjamin Bratton, who talks
• 1:29:23
  Speaker 1: about the humanity of the gaps. And every time, you know,
• 1:29:28
  Speaker 1: like the God of the gaps, you know, we knew
• 1:29:30
  Speaker 1: that after we found out that maybe, you know, about
• 1:29:33
  Speaker 1: the big bang, okay, well, God set everything in motion
• 1:29:36
  Speaker 1: and then we found out about evolution. Well, okay, well,
• 1:29:38
  Speaker 1: God sort of set that in motion. He didn't actually
• 1:29:40
  Speaker 1: do the evolutions, et cetera, et cetera. And the role
• 1:29:42
  Speaker 1: gets more and more limited. And he's argued that AI
• 1:29:48
  Speaker 1: is maybe making us do something similar with what the
• 1:29:50
  Speaker 1: role of humanity is because we had intelligence, we had
• 1:29:54
  Speaker 1: creativity and, seems like some, certainly intelligence can be performed
• 1:29:59
  Speaker 1: by these systems and a good deal of creativity can,
• 1:30:02
  Speaker 1: even though it's not necessarily great in some respects. But
• 1:30:05
  Speaker 1: the sort of what makes us uniquely human gets smaller
• 1:30:09
  Speaker 1: and smaller and smaller. And what's amazing is only in
• 1:30:12
  Speaker 1: the last five years, the kind of the list of
• 1:30:14
  Speaker 1: attributes that get called uniquely human is you can see
• 1:30:17
  Speaker 1: it shrinking in front of our eyes. And the one
• 1:30:20
  Speaker 1: that remains there is wisdom, right?
• 1:30:22
  Speaker 2: And the fact that we only have hair on our heads.
• 1:30:27
  Speaker 1: That's so unrighteous.
• 1:30:28
  Speaker 2: And chance, I think, too.
• 1:30:30
  Speaker 1: And I guess someone, you know, some of these more
• 1:30:33
  Speaker 1: sort of future kind of oriented people think about what
• 1:30:38
  Speaker 1: the world is going to look like once, you know,
• 1:30:39
  Speaker 1: we have these supercomputers would say, well, actually, rather than
• 1:30:45
  Speaker 1: sort of pin our fate to this ever decreasing notion
• 1:30:48
  Speaker 1: of what is uniquely human, what is a different way
• 1:30:52
  Speaker 1: of thinking about things where we sort of use these
• 1:30:55
  Speaker 1: systems that will be even more intelligent and more remarkable
• 1:30:58
  Speaker 1: to kind of flourish, right? And again, it sounds very
• 1:31:02
  Speaker 1: hand-waving because one doesn't know what it would necessarily look like.
• 1:31:06
  Speaker 1: The one that I think of in this context is,
• 1:31:09
  Speaker 1: especially what you're talking about, Deb, is I have tried
• 1:31:12
  Speaker 1: to do active imagination so many times. I'm just, I'm
• 1:31:17
  Speaker 1: not very good at it, right? And I'm, you know,
• 1:31:20
  Speaker 1: I find it, I get bored and the images don't
• 1:31:23
  Speaker 1: become autonomous and blah, blah, blah. And I certainly wouldn't
• 1:31:27
  Speaker 1: want to try it with one of these LLMs and
• 1:31:30
  Speaker 1: have Peter Todd or Sidney step through the door and
• 1:31:33
  Speaker 1: start playing with me. But once these kind of issues
• 1:31:40
  Speaker 1: with the model going in places that we don't want
• 1:31:43
  Speaker 1: get worked through, and presumably that will happen at some point,
• 1:31:47
  Speaker 1: I suppose that you could almost.
• 1:31:51
  Speaker 1: Sort of see a way in which these things could actually,
• 1:31:55
  Speaker 1: you know, help you sort of clear that road to
• 1:31:58
  Speaker 1: the unconscious in a way that perhaps for some people,
• 1:32:01
  Speaker 1: just the traditional techniques just don't work. Is that a possibility,
• 1:32:05
  Speaker 1: do you think? Or does it feel too crazy?
• 1:32:09
  Speaker 2: I love that you're thinking about it like that. Because, and,
• 1:32:13
  Speaker 2: you know, again, I've, you know, read a bunch of
• 1:32:15
  Speaker 2: your writing and I think you are really... an advocate
• 1:32:19
  Speaker 2: for sort of like the middle way and not, you know,
• 1:32:22
  Speaker 2: prejudging this and understanding that it can be really helpful,
• 1:32:27
  Speaker 2: even in this kind of psychological realm. I mean, we
• 1:32:29
  Speaker 2: all know that, you know, it's going to be helpful
• 1:32:31
  Speaker 2: in all kinds of domains, but you've made this case
• 1:32:34
  Speaker 2: that it could be really quite helpful psycho-spiritually. And it
• 1:32:38
  Speaker 2: might not be good for some people, it might be
• 1:32:40
  Speaker 2: great for others. And I think that that's at least
• 1:32:44
  Speaker 2: the right place to start, Tom, because Because as you've
• 1:32:48
  Speaker 2: mentioned this, you've talked to lots of people who have
• 1:32:51
  Speaker 2: felt really helped by it. And I know those people too. And,
• 1:32:55
  Speaker 2: you know, so I can feel in myself the desire
• 1:32:58
  Speaker 2: to sort of say, this is really bad and we
• 1:33:00
  Speaker 2: should just, you know, unplug all the data centers. But
• 1:33:04
  Speaker 2: first of all, that's not going to happen. And second
• 1:33:08
  Speaker 2: of all, that's not really holding the tension of the opposites.
• 1:33:11
  Speaker 2: And I think, you know, I would like to hope
• 1:33:14
  Speaker 2: that some of the research that you and your team
• 1:33:16
  Speaker 2: are doing might lead to design decisions that could ensure
• 1:33:22
  Speaker 2: that we do get the very best of this.
• 1:33:25
  Speaker 1: I hope so, too. Yeah.
• 1:33:28
  Speaker 2: And maybe that's a good place to leave it for today.
• 1:33:32
  Speaker 1: Well, thank you.
• 1:33:35
  Speaker 2: Thanks for listening. To submit a dream, suggest an episode topic,
• 1:33:39
  Speaker 2: or join our mailing list, visit our website, thisunionlife.com. If
• 1:33:44
  Speaker 2: you enjoyed this episode, give us five stars and a
• 1:33:47
  Speaker 2: good review on Apple Podcasts, Spotify, or wherever you get
• 1:33:50
  Speaker 2: your podcasts. Subscribe to our YouTube channel and make sure
• 1:33:54
  Speaker 2: to click the notification bell to be alerted whenever we
• 1:33:57
  Speaker 2: upload videos. And keep up with all things TJL by
• 1:34:01
  Speaker 2: following us on Instagram, Facebook, X, and TikTok.