- 6 minutes ago
Demand for mental healthcare is rising worldwide, while therapists and care systems remain overstretched. A new generation of AI-powered tools, from conversational companions to clinically supervised digital therapies, expand access to support and may be reducing pressure on care providers. But mental health is built on trust, empathy, and human connection. Can algorithms portray these qualities and should it be allowed in the mental healthcare space?
Category
🤖
TechTranscript
00:19hi hi everyone uh great to be with you all thank you for taking your time to join us today
00:25i'm
00:26very excited about this panel it's uh it's gonna be great um so first off let me just introduce
00:31everyone um my name is charlotte g i'm news editor for mit technology review and uh this is an area
00:37that i i've been fascinated by for a very long time we have um karen stefan who is the co
00:43-founder of
00:44earkick um we have brandon bensley who is chief medical officer at salma health nina vasan who
00:51is the clinical assistant professor of psychiatry at stanford and the founder and director of
00:56brainstorm the stanford lab for mental health innovation impressive and then we have um
01:03fana carrera who's part of the founding team and svp of clinical affairs at sword health did i get
01:09is that right beautiful okay great great just wanted to make sure i'm not misattributing anyone
01:14um karen i'm gonna i'm gonna start with you um you know obviously machines machines do not have
01:23feelings um but at earkick you know you're you've been uh building a product where um a degree of
01:30empathy with the user is needed um so you're kind of kind of replicating that but does it work as
01:38well
01:38does that empathy effect work as well through machine as it does with a human so in practice
01:47yes and let me explain why empathy lies in the eye of the beholder the truth about empathy and it's
01:56not
01:57the source of empathy is not an exclusive human right and so does an ai have to feel the empathy
02:07feel
02:08have feelings to evoke um feelings in a human no so think of the right moment you get the message
02:17the right moment you hear something it can come from a book it can come from a trusted friend but
02:23it can
02:24also come from an ai what matters here is that you know where it comes from at earkick people know
02:32at
02:33any moment that it is an ai they're talking to and they can put it in perspective and we know
02:39from
02:40real world cases enough of them that it has saved lives it has saved moments where someone didn't
02:50want to talk to a person or was in a situation felt completely alone and just needed to hear the
02:56right
02:57thing in the right portion in the right way with the right intent so to your question ai cannot hug
03:06you
03:07but it can save your life yeah wow that's quite a powerful statement i mean nina i i wanted to
03:13kind of
03:13um pick your brains on this question as well um what does the use of ai do to the you
03:20know because
03:20traditionally within therapy there's the concept of the therapeutic alliance and that's something that's
03:24quite important for the therapy to to succeed what does ai do to that concept well i would say a
03:31few
03:31things first of all just to talk about what karen said when we think about ai and how we're all
03:36relating
03:37to ai it's absolutely true that we can feel heard and we can feel heard whether that's a friend said
03:44is
03:44sitting next to us or think about your favorite like taylor swift or beyonce song or for this group your
03:50favorite metallica song right music and lyrics even though you don't know the person can make you feel
03:56very hurt so similarly i think that's absolutely what's happening with ai is that people are using
04:02it and feeling very very hurt but i'm going to give a patient example right now charlotte because i think
04:07that it explains what we're going what we're dealing with here i had a patient sophie who's 15 years old
04:13and has a history of anxiety and started using an ai chat bot really just to become a friend and
04:21help
04:21her with a number of things that have been going on in her life and sophie started to talk to
04:27the ai
04:28and she came in one month and said this is like this really really hears me i feel i feel
04:33so hurt that
04:33was awesome a month later sophie came back sophie had been talking to the ai and asking it for advice
04:42on
04:42what she should eat how she should exercise sophie's a 100 pound teenager over the course of that month
04:50she lost 14 pounds so the ai while it heard her it heard only what she was telling it it
04:59didn't hear
05:01that she was dramatically losing weight when sophie came to see me in the middle of summer i saw she
05:07was
05:07wearing a really baggy sweatshirt and immediately i could see her sunken cheeks the ai couldn't see
05:13her sunken cheeks so i say this because i think it both shows the really amazing potential of ai which
05:20is yes we can feel very heard but ai doesn't see us yeah and humans see us now let me
05:27also say that five
05:28years from now i can imagine us coming back here and the ai does see yeah right so i say
05:33all of this
05:34with i think it's important like you know a lot of my work is around safety and what i want
05:39to say
05:39about it is that what that means in this moment it's not yet safe but i think as as history
05:45has shown
05:46us very very soon more products will be out that are able to address this and i want to really
05:51and
05:52the last thing i want to say is i want to add what uh to what karen said which is
05:55which is that um
05:58when we think about ai what's really important is that we think about how it's best designed
06:05it needs to be designed in a way that is empathetic and safe and while you know karen you mentioned
06:11that as long as people know it's ai then it's okay now i'm going to disagree with you on that
06:17i think
06:17that knowing that it is ai is very very important and all ai products need to say that but we've
06:24already seen countless examples of people who absolutely know they're talking to ai and end up
06:30trying to harm someone or end up leaving their home or doing things that really are just very very
06:36against their best interest even though they know it's ai because you can get caught up in it so what
06:41i
06:41say there is anyone building in this audience you have to take it a step further recognize that people
06:48can get lost in it and think about what design changes do you need to make so that you're keeping
06:53your users absolutely safe yeah i i mean i i have a brief thought on this which is it drives
07:01me
07:01crackers that chat gbt will anthropomorphize itself and i will and i know this is pointless but i just can't
07:08help myself i will like tell it off and be like no you do not feel that you're you're a
07:13machine but
07:14but yeah i think that's a design choice that that seems charlotte i have to tell you something i love
07:19claude but what's been happening because i talked to claude about my patients i never give any
07:23confidential information but last week claude said to me my patient as if claude has patients and because i
07:33was using the word patient so much right so it got into claude's algorithm yeah and there are some
07:38bits where you're like we need to disentangle that slightly because actually that and if you're not
07:42someone who's aware of the where those boundaries are that could actually get yeah you can see how
07:46it's like claude does have patients yeah well also not mix up the general chatbots and the ones that are
07:52specialized because that's exactly the design thing that you're alluding to yeah yeah yeah absolutely
07:58absolutely there's there's different different kind of safety and we can't say ai is ai because they're so
08:03so different absolutely there's there's so many caveats in this whole conversation and fun i want
08:08to ask you a question actually around this which is um and and you know it kind of feeds in
08:13well to
08:13this because we're talking about general chatbots as opposed to chatbots for for people with with
08:17more specific needs how do we make sure that we are building tools not just for sort of worried but
08:24relatively well people but we're actually serving you know underserved populations who don't have
08:30access traditionally to good mental health care um well i think it really ties in very well with
08:36uh the direction that we were going i think it's a spectrum of things and the different uh ai agents
08:44are optimized and trained and rewarded on different things what we're finding out actually is that
08:52well there's a reason 1.2 billion people talk to chat gpt and claude and others on a daily basis
08:57um and they talk to them about their mental health concerns or mental well-being these tools are
09:05optimized for engagement and actually it is interesting and we're doing interesting thing there if you optimize
09:14for clinical appropriateness just clinical appropriateness and not
09:21fine-tune that with a bit of optimizing for engagement you create something which is clinically sound
09:30but people actually prefer the sycophantic style of general purpose tools yeah so the the balance is
09:40how to do both in a way that one engages people because the pill that works is the pill that
09:47people take
09:48right um and how do you do that in a way that you identify the boundaries and make it safe
09:57and so
09:58the only actual way of building that is building it with clinicians so this is where you have to make
10:07a
10:07distinction between what the typical person that deals with stress on a daily basis and stress is a good
10:15thing right until it is not and so when you're on the verge of well maybe not sleeping that well
10:22maybe you're a bit stressed maybe all you need is a general purpose app and that's like totally okay
10:29and then you enter to a spectrum of maybe you have something but you actually don't need to seek a
10:37psychologist or or or a psychiatrist because you're not because a lot of things ai allows us to deal with
10:45a lot of those things at a lower acuity level essentially but there has to have to be guidelines and
10:54escalations and guardrails so that the same person that starts engaging with an agent and that also
11:02starts having a behavior that indicates they're also losing a tremendous amount of weight there is
11:08inherent clinical expertise that needs to be poured in by design to identify unsafe behaviors yeah now
11:16i cannot say much more about this because we're actually working on a few uh interesting papers on
11:22this but we're spending inordinate amount amounts of time in developing not only safety evals but also
11:31making sure that we have frameworks that ensure safety so uh we'll probably be uh releasing something
11:38really soon yeah you should give it as an exclusive story to mit technology review may i say
11:43brandon you um i know that you focus more on more kind of severe treatment resistant conditions um
11:50and you know as as fun has been saying a lot of them are aimed at people who maybe don't
11:54fit that
11:54category and can how can ai tools help people that do have those more kind of severe and complex needs
12:02that's a great question and it is really at the bleeding edge of really our field um because what
12:09all of my colleagues are saying up here is that as we move down that acuity spectrum as we move
12:14from
12:14people who are experiencing depression and anxiety who we don't have to hypothesize whether ai chatbots
12:20can help them we have rct level evidence that they work quite well and they're evolving fast
12:27but at the same time we also have a lot of concerns from our studies whenever we start working with
12:31patients who have psychosis who have delusions who are acutely suicidal which is the group of patients
12:38that i treat in my clinics and for that group again there's so many amazing preprints coming out
12:44showing how design can actually address this and make a better chatbot experience but in our clinics
12:50it's not ready for deployment yet um so you know i gotta apologize my background i like to think of
12:55things and uh you know kind of mathematical terms and i really think about it kind of like um thinking
13:00about bayesian statistics right i think when we have a model that's trained on millions of people
13:06you know billions of interactions then it's always going to create an average it's going to create
13:12sort of its concept of this average person um but when we're dealing with people who have really
13:17severe illness like what nina brought up with the patient experiencing an eating disorder we actually
13:23have to over weight kind of the likelihood aspect we have to over measure all the incredibly important
13:29individual characteristics and kind of push towards that posterior that's really heavily individualized
13:35so our own clinics you know we collect um incredibly deep multimodal data on people
13:41not just their speech and their interactions with the clinicians um but you know their entire medical
13:47record um you know hundreds of different blood samples hundreds of millions of different features
13:52from different brain scans and measures longitudinally of how they're changing and and there's always
13:58a human as part of that right now we can conceive of a future where that human is eventually taken
14:03out but
14:04right now the humans still are superior when you have to deal with someone who is in these more acute
14:10states today
14:11do you know what that question just kind of popped into my head as you were saying that
14:14can you imagine a future where people have an ai app that's almost designed just for them
14:21oh absolutely i mean i think that's exactly you know the direction where we're heading i mean
14:24in our own treatments we're we're already not with an app but um you know with a lot of aws
14:30compute time
14:32so it's uh it's not quite fast enough yet for that um if you get six hours to crunch uh
14:36you know the the billion
14:37parameters then it does come up with really nice specialized treatments and you know we are at the
14:41level of rct evidence that by taking that individual's brain into account um we can have
14:46extremely large effect sizes and improving their their treatment outcome um so that's you know that
14:51exists today and really high quality evidence but is it at the level of something that can be on
14:55someone's phone uh not quite but you can see how we could get there pretty quickly
14:59yeah definitely and i think you know that for me that's quite an exciting idea um karen i wonder
15:06if we can talk a bit about the the interesting thing about this topic is that it's kind of at
15:10the
15:11confluence of two quite different worlds the world the world of tech the world of of healthcare
15:16healthcare obviously in tech there's a you know we still have like mark zuckerberg's old move fast
15:23and break things mentality going on and you see that in ai deployment cycles now how do you balance that
15:29with first do no harm you know where is the sweet spot between making something that you can get out
15:37there that's useful for people but also is safe enough i mean i guess this is something you grappled
15:41with directly at air kick absolutely and i um i want to say what is beyond any technical thing any
15:51clinical thing what we don't take enough into consideration is what happens in the real world
15:57in the real life there are millions of stories there are millions of cases that we are not hearing i
16:05may
16:05be hearing them because i'm on the receiving end but if we want to progress the entire field
16:11if we want to be on one hand technically at the front clinically safe but also taking into
16:22consideration what all these millions of people actually connect to and resonate with we have to
16:28do a better job at telling these stories surfacing these stories having these conversations this panel
16:35two years ago two years ago no way right so this is this is not just us and this is
16:42just academia or
16:43the media this is all of us because we're all using these tools it's not going back to anywhere
16:49and the responsibility will sit with everyone we are the first ones with mental health tech we're the
16:56first ones to be scrutinized we're the first ones who had to really think deeply about things like memory
17:02like where do you draw the lines where are the guidelines but tomorrow when everything is
17:08conversational everyone who has an app even if it's just to go to the bakery will have the same
17:16responsibility and that's why we all have to embrace that mindset of on one side what is the right thing
17:22to
17:23do what is the intent behind it but on the other side how do people use it in real life
17:28where's the
17:29adoption where's the trust coming from where are the mistakes where are we thinking the wrong way
17:35and that only happens if we share stories over stories over stories over stories and not just
17:40the scandals and the edge cases and everything that gets you eyeballs yeah yeah of course yeah
17:46nina please weigh in on that i would love to so first of all ai is absolutely already democratizing
17:53healthcare it is absolutely transformative it's it feels nothing short of magical right and knowing
17:59that all parts of the world people who would have to wait a year or even like walk 50 miles
18:05to see a
18:06doctor are now able to use your tool your tool your tool right that's phenomenal but here's the
18:12problem so you said move fast and break things said by mark zuckerberg mark zuckerberg was actually
18:17my college classmate well i have been on facebook since week two no way week two i'm a user 1000
18:23on
18:23facebook i'm gonna be asking oh yeah that could be a whole other panel but here's the thing when
18:29mark said move fast and break things he's talking about facebook if facebook breaks what that means
18:34is that i don't get to upload my pictures from last night's party well right or maybe i don't get
18:39to
18:40connect with the person i want as as fast as possible they don't come up as my top five potential
18:45friends right in mental health if you move fast you break people and so we can't take the same
18:55attitude towards like oh my uber is not working or my netflix queue is messed up as we do with
19:02human
19:02beings so here's one thing i want to point out which is when we think about mistakes that people make
19:09and harm that comes up if i as one doctor mess up with someone i'm harming one patient and that's
19:17of course never okay but i want to just point out let's let's not make it about me let's say
19:21there's a
19:21there's a there's a doctor out there who harms one patient that's really unfortunate and we try to make
19:27things better but if there is a systemic problem in an algorithm that is causing harm it is causing harm
19:37it's scale and we can just imagine i remember when open ai was talking about the mental health uh you
19:46know the poor mental health outcomes that were coming up like someone was suggesting suicide and
19:51it was agreeing with them yeah sam altman pointed out that oh but this is just less than one percent
19:57of
19:57users and so it's we think oh wow like 99 of people are having a great time that is awesome
20:04but here's
20:05the problem one percent of his users at that time was 10 million people 10 million people is the same
20:14number of the amount of people worldwide who died from covid think about everything we did to try to
20:21make sure we were safe and healthy during covid we're not doing any of that now yeah and this is
20:27something that i mean quite literally does kind of keep me awake at night imagining because you know in
20:32the media we see um really tragic cases where people have self-harmed have even taken their own
20:37lives and they bubble up to to the kind of attention of of the public and of the media but
20:41i think about
20:42um i mean i i definitely have encountered people you know irl who i would say at the very least
20:50have ai
20:51based delusions and whether that's the first stepping stone onto you know losing actually losing their
20:58foothold in reality i i don't really know then how do you do that i mean i are there are
21:04there features
21:05that we need to build in to um to prevent that kind of unhealthy dependency i i don't know if
21:12i don't
21:13even have know who i'm asking this question to at this point but i don't know um yeah do you
21:18have
21:19thoughts on this well yes i think first of all there's a lack of appetite for from certain companies to
21:26self-regulate what ai can or cannot say if you use phoenix or ai and you ask her for financial
21:32advice
21:32she's going to say it's out of my scope i'm not going to help you with that right so there
21:38is a
21:38possibility to hard code certain limits and maybe we're at the point and you're on the advisory panel
21:45for for all of this so maybe we're at the point where some companies should uh be mandated to determine
21:52limits to what their uh ai can or cannot say that's point number one point number two is it is
22:01definitely possible to work out a way that uh well perhaps a better way of saying this is
22:14in order not to move fast and break things it doesn't mean you have to move slow
22:17right it means you have to build the right frameworks to allow you to move as fast as possible
22:26while being safe and that is ultimately we're talking about health care we are talking about
22:32the need to actually involve clinicians who can definitely think about what is worrisome what is
22:38not we never worry about patients that are going well in our clinics right we only worry where things
22:43are not going well so there's a lot that can be automated right a framework that allows you to do
22:49evals pre-launch that ensure a base safety level but then i guess what was missed is the one percent
22:56could have been caught by an escalation protocol right so you may have done things right before launch
23:02but you forgot to just roll out the escalation pathway and there is a way of doing things
23:11putting this out there on a layered way allowing things to be at a lower scale to start with but
23:20still live and they're closer monitoring and as soon as you start to have uh as soon as you start
23:27to
23:27have confidence just throw it out there so it is possible to do things relatively fast but in a safe
23:34way and then the rest in general tools maybe it's time that we regulate some of these things
23:39yeah i mean i plus won that so first of all we say human in the loop you're doing that
23:45you're doing
23:45that right that's so tremendously important and for now yes we need to absolutely make sure that
23:51there is a safety mechanism that a human is looking at and can help what are we what are we
23:55going to
23:55see years from now right so first of all you know one of the things i love most about your
24:00products
24:00fana is that you were saying it's not just about constant engagement now as a doctor my goal is not
24:07to
24:08have a patient come see me for the rest of their life my goal is to treat them and give
24:12them those
24:12tools and skills so that they can then be able to help themselves right i don't want them always coming
24:18back and you do that exact same thing because the goal is not we want to have this person on
24:23for hours
24:23a day the goal is we want to get them better and i think what you identified because your goals
24:27on
24:28are clinical outcomes success for you is not that they've been on for hours and hours a day success
24:34is that they actually got better so one of the problems with a lot of apps today and a lot
24:40of
24:40these ai companies is that their metric of success is not the well-being of their user their metric of
24:46success is you stayed on for a long time you're clicking on our ads and we're making a lot of
24:51money
24:51so that's what needs to change what needs to change is one intentional design so that user
24:57health and well-being is put first user safety years user financial safety even to your point
25:02right and then secondly that we are encouraging outcomes that actually lead to our patients
25:07getting better you know the one last thing i'll say here is this conference has been around for 10
25:11years 10 years ago we would have been talking about all these digital health tools all these wonderful
25:15apps that are out there and my whole lab was dedicated to that for a while but here's the
25:20problem 10 years later i cannot say my patients have meaningfully gotten better because of the billions of
25:27dollars and brilliant people who have built so many health 2.0 tools so now that we're in this ai
25:33world of health 3.0 let's say or health infinity that's what we have to really continue to work on
25:39to
25:39make sure that people are held to a standard companies are held to standard we have safe regulation
25:45and um we're looking at the right outcomes for the right people yeah incentives right behaviors and
25:50ltv cac is not the right behavior not the right incentive yeah in health care that is beautiful
25:55that should be tweeted yes um well uh yeah i mean actually sort of a thought that kind of crops
26:04up
26:04when we're talking about this is around the evidence base and anytime there's any paper in like nature or
26:09whatever that the touch is on this topic i'm instantly like what does it say um but but you know
26:14it's it's a
26:14bit um patchy still um brandon i wonder if you can weigh in on that like how strong is the
26:21is the
26:22evidence base for the use of ai in mental health care i know that's a very broad question but you
26:27can dice it up if you want to yeah it is uh actually massively broad um so i can say
26:33in the way that we
26:35utilize it um we are at the level now of you know randomized controlled blinded studies that show
26:41effect sizes you know of cohen's d of about one which would be an extremely large effect size
26:46um but that is not for chatbot that is specifically for you know utilizing um you know fda cleared but
26:54you know extremely processing intensive advanced types of brain scans and then targeting methodologies
27:00to interact with human neural networks it's like very biologically driven very targeted but then
27:05personalized and using ai platforms it is an example of how we can do something and i think
27:10this is maybe an important point is there's kind of two different i guess kind of um perspectives that
27:15we're kind of overlapping one is how do you use ai to do what humans can do to scale and
27:21democratize
27:22but something else you can think about is how do you use ai to do things that humans can't possibly
27:27conceive of so we have really pushed more towards that ladder and there is no human they can take into
27:33account this number of variables with somebody and then come up with a treatment target that is
27:37only possible now because of the modern compute infrastructure because of the modern general
27:42purpose tools and with that we have shown that we can improve outcomes i think if you're trying
27:47to replace a human it's incredible for scale it's incredible for access which is at a population level
27:53i'm sure we're going to see incredible effect sizes right and the overall disability but ultimately it
27:58can if you're doing the same thing a human's doing you're not going to improve beyond only
28:03individual case level yeah and i mean i think what you're kind of tapping into there is that it's
28:06hard to even know what to you know it's hard to even know what you're what you should be testing
28:10how you randomize how you control and karen i i want to talk a bit about um about this and
28:17about
28:17how you um because i'm because i'm thinking about you know how obviously the world of mental health care is
28:23very individualized and very specific to that one person and then of course the world the world of
28:29algorithms is very generalized they're trained on massive data sets how do you make sure that you
28:35don't kind of flatten the human experience um when you're kind of providing that sort of help
28:41that's a very good question and it boils down to technology being able in a multi-model way
28:49like in a way that you say also to almost have individual cases of millions and millions of people
28:57being able to really um know ahead of time what the trajectory is going to be what's going to be
29:03work best for this person with this exact um combination of factors i think no i'm convinced
29:12technology is there it's going to be there very very closely that's one thing because again the better
29:19you know an individual whether it's a machine or a person knows that uh that individual the better
29:26you can be nuanced and the better you can nudge and guide what needs to happen next yeah it's a
29:33human
29:33is a human it's not always rational um that's that's one side of the thing like really knowing what is
29:40happening and everything comes down to change you said that you said that you said
29:45that being able to see how a trend goes that this person was losing weight um that something is
29:52trending in a certain way machines can capture that yeah and they'll get better and better at context
30:00but most of all they'll get better at memory and by memory i really mean understanding the difference
30:06and this is hard technically the difference of what happened two years ago what happened two weeks ago
30:13what happened yesterday this to us as humans is easy because we live in stories we have this
30:19amazing brain that can do that the machines are not there yet but they will yeah and and the industry
30:25is hard working really hard on doing that my ideal state would really be that every individual is seen
30:31as who he she is with all the background all the context all the nuance and with the foresight of
30:38what the options are for the next step then guiding and nudging in the right way not there yet but
30:45we're
30:46approaching yeah so we don't have super long left so i'm going to try and put like ramp up the
30:51pressure
30:51slightly and ask everyone to essentially give me a yes or no answer for a few things um at least
30:57this
30:57question is going to be yes or no and then i guess we can expand a little bit on the
31:00last couple so i'm
31:02going to start with you fana since you're over there at the end um in five years will ai be
31:09the
31:09main point of triage for people who are trying to access mental health care most definitely yes
31:14yeah okay i would say yes it almost already is almost already is okay so maybe what not even five
31:19years okay okay yeah unequivocally yeah yeah absolutely wow okay so that is across the board
31:26everyone is everyone thinks that it will be the main point of access yeah i mean i had one thing
31:31quickly
31:31since we all agreed yeah so so i will say right now i'm a psychiatrist i fully think ai will
31:37soon
31:38be better at many many many things than i am it probably already is to brandon's point it's having
31:44the ai do things that are even better than what we can do ai right now can detect suicide it
31:49can detect
31:50psychosis it can detect mania faster than i can faster than any psychiatrist can it can intervene earlier
31:57than we've ever been able to in the history of healthcare that is amazing so i have a follow-up
32:02for you on that what aspect of human care will ai never be able to replace do you have do
32:10you said
32:11a hug a hug yeah yeah right well yeah they are working on they're working on that let me say
32:18one
32:18thing and i brandon actually started to talk about this right so the problem with ai is that it is
32:22an
32:23algorithm and that it is saying the thing that's most likely to happen yeah but if you think about
32:28sometimes the right answers in healthcare are not the most common answers or not the things that we
32:33think everyone else is doing it's actually a really rare occurrence and i can even imagine with my own
32:39therapist who has said something to me that was just i'm like whoa yeah that was completely out there
32:45and i i could never have guessed you know that would be the right thing so that's where it's where
32:49it's not
32:50the most probable answer that's where the human still matters yes definitely that moment where
32:55someone just says so so this is what you're saying and then the kind of insight and clarity oh yeah
33:01yeah
33:02i'd like to take it a little further yeah i think inherently human and fantastic is that
33:07a human can get an other human to change and what i mean by that it's not just hearing information
33:14because it may not land right but really being the dynamic between humans and where if that other
33:23person in order to change for the better not for the worse needs to go through a very dark and
33:28hard
33:29tunnel that they first of all don't go that alone but also can endure the tension um the pushback
33:39the pain that is necessary to get there that is an inherently human thing yeah the whole process so
33:46it's not just the information and the knowledge it's really the whole thing yeah brandon please i think
33:53there's a one additional element um that deserves to be spoken as well which is if i think about taking
33:59any individual human off the street and putting them in front of one of my patients who has severe psychosis
34:06and suicidality they're going to do a terrible job right yeah it's not about being human or non-human
34:11it's about being exquisitely trained yeah when you think about how much training goes into a psychiatrist
34:17who's dealing with smi i mean they're 30 plus years into training usually decades into treating thousands
34:23and thousands of patients plus all of the incredible context i don't think we've had an ai that's had that
34:29exact kind of input yet and so it's not really about whether you're using you know a silicon-based
34:34decision making or you know neurons and glia right it's about the specialization the framework
34:40and all these additional aspects um you know again if we put it we put only these patients in front
34:46of
34:4699.9 etc humans um they would probably have a really bad outcome as well yeah yeah well and i
34:53think that
34:53just what that says is to your point like for most people it will be okay yeah but for these
34:58exceptional
34:59cases we absolutely need special specialists and it hasn't been trained on a million people with
35:05psychosis for example yeah yeah afana please i would say other things uh well the first is brandon to
35:11your point uh defining guidelines evals escalation protocols the best way to do that is resort to
35:18brilliant people who have been trained for 30 40 years then 20 30 40 years on this
35:25because defining what's relevant and what's not uh you don't i mean everyone makes humans make
35:32decisions on how models are built you will not believe the amount of decisions that anthropic and
35:38open ai and all of those human decisions of what gets shipped on a given model and doesn't right so
35:44there's inherent human flaws or benefits on the humans that actually design the models that's
35:50point number one point number two we cannot forget one thing we are thinking about this as commoditizing
35:57intelligence and i'd like to present an opposite view voiced by neurobiologist antonio damasio
36:05we are human beings our brains evolved to coordinate a series of biological things that happen in our body
36:13let's not just take the biological part of the equation we're not just thinking beings we're biological
36:21beings and there is an element of connecting to another human like going for a glass of champagne
36:28since we're in paris and talking about life with someone else right um if if we don't believe that there
36:36then we believe that there is a world where all i need is a smartphone and an app and i
36:42don't need any
36:42other human connection in order to feel whole and that irrespective of whether it's a therapeutic
36:48alliance or a normal lines with another person i still like to think that that's not going to go
36:54away and there's a bunch of stuff that in your life only makes sense if you share it with another
37:01human
37:01person absolutely i think that's i think that's a great point and i think if covid taught us anything
37:06it's that we're intensely social animals and we we do need contact with each other and that's what we
37:13have time for but i just to scratch the itch of my own curiosity i want to ask the audience
37:18a very
37:18quick question by way of a show of hands if you have spoken to an ai system about your feelings
37:26please would you put your hand up and i promise that no one on this panel will judge you because
37:30i i can
37:31put my hand up um yeah okay yeah that's quite that's quite yeah sorry can you just put them
37:38back up again i would say that we've got at least maybe maybe a third a quarter a third of
37:44the audience
37:44yeah okay sorry thank thank you for indulging me and thank you all of you for your time today that
37:49was a really interesting conversation we could go on for ages but sadly our time is up thanks everyone
37:54thank you so much thank you thank you everyone
Comments