Full transcript
0:21good morning everybody Welcome to our panel on how AI is impacting multiple areas of societ Society
0:31um we're going to have a wide- ranging discussion that's going to be a little bit like a a crossover
0:36episode we're going to have perspectives from law from sociology from engineering uh from
0:41economics and essentially what we're asking is what type of deep societal Transformations are
0:48playing out just under our eyes what disciplines do we need to make sense of the mle and to ensure
0:56that really AI can be uh deployed for the public interest how do we maximize societal welfare in
1:04the context of these new AI deployments we're going to start this discussion by having four
1:10absolutely fantastic experts from across campus talk to us about different areas where we see
1:17those Transformations at play and then we're going to have a discussion on what do we need to focus
1:23on maybe in terms of regulatory framework maybe in terms of building AI differently uh maybe in
1:29terms of of economic interventions in order to ensure that we can um really serve the public
1:36interest with the way AI is being deployed we're going to start with Gil who's going to talk to
1:42us a little bit about the crisis of trust and expertise that we're witnessing a crisis that
1:49predates AI but that maybe gets accelerated when your own phone uh know knows so much more about
1:56certain topics than you sort of prompts us to ask who's really the expert then we're going to go and
2:01talk a little bit uh to Claire who's bringing a family law perspective and you may ask family
2:07law why but it's because we learn family laws of body that helps us think about relationships and
2:14CLA is going to tell us a few totally terrifying stories about how people are already very much
2:20in relationships with different areas of AI and so when you know people turn to Ai and have very
2:26long conversations treating those chat Bots as therapist you might ask who else is reading those
2:32conversations that's the right question and Rachel is going to help us make sense of that
2:37bringing an engineering perspective and helping us think about some of the Privacy implications
2:43and finally I think we're here because we have an interesting Market structure around AI with a few
2:49dominant players uh who perhaps are not properly architected to maximize societal value Joe is this
2:56kind of what we're going to head towards I have a feeling you're going to tell us that maybe there
3:00are market failures about the current industry we're going to round round this up in our intros
3:06with sort of an economic perspective so with this um over to youil for this first short intro on
3:13trust expertise and AI how do we how do we think about that okay um I'm not here because I'm an
3:20expert on AI I'm very far from that uh I study experts which means that I'm most of the time
3:27I'm a diletant people I study know much more than me uh but I do want to talk about something I do
3:32know something about which is the potential consequences that AI might have um on trust
3:37in science experts and institutions now it's no secret that we are living through an era in which
3:44um there's declining trust in science and experts and institutions um you see this in surveys and
3:50you see this in the political news every day um many factors are implicated in that the the rise
3:56of populism anti-science attitudes some people would say ignorance um growth of misinformation uh
4:04you could say also some spectacular mistakes that experts have made um but I think these are mostly
4:11symptoms and sort of recursive effects of the real uh more systemic crisis which has to do with the
4:18unprecedented degree to which science technology experts are um we rely upon them in our politics
4:26in our economics Even in our everyday life Etc um we all went through the pandemic for example so I
4:34I don't think it will be surprising to say that um if science is called upon to support policy
4:40decisions that have you know uh um redistributive consequences to people then science itself becomes
4:49uh infected uh become is drawn into the conflict and and trusty clients now what does AI have to
4:57do with all of this it depends it's depends on how we integrate um how we welcome this
5:04new member into our Pol how we um socialize this child that has enormous Powers but very little
5:12Good Sense how we socialize it um um especially as we integrate it into expert settings of work
5:19uh hospitals Law Courts uh government science agencies Etc so that's my main message here uh
5:27we we have we have an opportunity to integrate AI um in a way that counteracts decline in trust
5:33but at the moment all the trends are going in the other way um and the integration of AI is likely
5:38to lead to um uh in the other direction um so AI has certain characteristics that in interaction
5:47with the nature of trust um could exacerbate the crisis um if if not handled correctly I'll only
5:54mention a few but I'm sure in the Q&A if we get to it uh there could be more um so one dimension
6:01of the crisis of trust in expert science and and uh Etc is the perception that human judge human
6:08expert judgment um is subjective and suffers uh from multiple biases and the typical response
6:16long before AI um was to try to replace expert judgment with some form of mechanical objectivity
6:23um you know randomized control trials forms of quantitative measurement algorithms Etc
6:30now ai is part of this family AI is is mechanical objectivity on steroids um now it is important
6:37it's crucial um mechanical objectivity of clinical trials for example was crucial for the legitimacy
6:45of vaccines and medical drugs and AI can do the same but mechanical objectivity and AI suffer
6:51from a couple of problems first they promise more than they can deliver um it's not really the case
6:59that there's no Reliance on human input and expert judgment it's not really eliminated it's just
7:04backgrounded it's behind the screens um nor on the other hand um you know randomized control trials
7:12or generative AI for that matter are as objective or error free as as they build and predictive AI
7:19is even worse um we all know the biases that um uh in in training data can become biases in AI output
7:28um and accurate prediction from the training data can become very poor decisions in real life
7:33situations now at the same time however mechanical objectivity and AI tend to undermine trust in
7:42human expert judgment particularly in the way that they are introduced because they often introduced
7:46telling us oh they do much better than the experts you know they're objective they're not subjective
7:51they're not biased Etc so if not handled correctly the result could be destructive of trust in both
7:57Ai and experts imagine a scenario in which an AI system for analyzing uh MRI is introduced
8:04in hospitals you don't have to imagine it it's already happening um it performs uh very well
8:10in 99% of the cases but then makes fatal mistakes in 1% of the cases patients die and the hospital
8:18is sued now the hospital counters says well but you know the AI system is actually better than
8:24doctors its rate of error is lower the result is of course that we will trust doctors less and we
8:30will trust AI less and that has to do with the nature of trust AI is probabilistic trust is not
8:38um in AI 80% accuracy rate is fantastic it's you know the Holy Grail um but not so in trust you
8:45know if your spouse told you that they were only unfaithful to you 1% of the time but 99% of the
8:52time they were faithful that's not going to you know that's not going to help right uh trust is
8:57asymmetrical and non monotonic the more you trust the more you might feel betrayed if something goes
9:04wrong and the harder it would be to rebuild trust um and this interacts poly with another dimension
9:10of how AI is currently introduced um currently the incentives for AI development and scaling
9:17up encourage an extremely fast pace move fast and break things is the motto right uh but trust as we
9:24know it takes years to build seconds to destroy and an eternity to rebuild so you could see that
9:31the incentives are set up so that um developers would would tout their systems as better than
9:37the experts it's proving the experts wrong um and and they set up almost to guarantee exacerbating
9:45the crisis of trust and a different response to the crisis uh of expertise is to encourage
9:51participation of lay people in expert decision making um because they bring in uh uh you know
9:58experience based expertise that the experts don't have they bring concerns that the experts are not
10:03aware of um and this is great I'm all for it I've done research on the history of autism and
10:09I can say with confidence that autistic persons are better today because lay people parents and
10:16patients fought and gained recognition as lay experts but inclusion also has its downsides
10:22it's extremely hard to tell uh who should be included who has relevant expertise where to
10:27draw the line it can make make it extremely hard to reach consensus uh it creates what is called
10:33inclusion friction slowing research to a crawl and um transforming scientific consensus to a
10:40cacophony of Waring voices which we heard during the pandemic now the spread of generative AI which
10:46we all have on our phones has the potential if not handled correctly to amplify this aspect of
10:52the crisis now everybody has at their disposal an AI research assistant so everybody can transform
10:59themselves into an expert only unbeknownst to them the AI research assistance has taken acid
11:04and is hallucinating U you know every now and then moreover with generative AI in wide usage becomes
11:12hard to distinguish between facts misinformation and hallucination oh I forgot we're not supposed
11:18to distinguish between these anymore at least now on Facebook um and that's incidentally
11:24also part of the problem as noted earlier the incentives are misaligned and now political and
11:30economic forces are set up so to exacerbate the crisis of trust yeah that gives us a a good set
11:37of starting thoughts uh I learned that you know while we want AI to be mechanical objectivity on
11:44steroids such an elegant phrase really we should always consider the fact that our new research
11:49assistants are probably most likely often on acid which is a problem because in the few percentage
11:55of times where they're deeply wrong this actually may sort of amplify our crisis of trust that's a
12:01good set of provocations to think about how AI um sort of gets integrated within institutions
12:09and in a societal context where we already have these deep crisis which as you're saying are
12:13being accelerated and it's it's a nice segue to thinking about how does that play out not
12:19from a sort of broad Society perspective but from the individual level because it does appear that
12:25while we're having these crisis of trust we do have individuals who are actually deeply trusting
12:32of these new machines don't seem to think that they're talking to therapist on acid CLA tell us
12:37a little bit more about what's really happening there and how do we make sense of the fact that
12:43individuals here seem to have a very different type of relationship with these new chat Bots well
12:49that's a great segue um so I am very interested in the relational impact of AI because increasingly
12:56AI does have this relational aspect whether talking to Claude when we're using anthropic
13:02or you know the fast approach fast approaching world of a gentic AI where we are all going to
13:07have an AI coworker so I'm going to talk this morning about AI companions because these are
13:12a very specific example and a really a really excellent example of this relational impact of
13:17AI but this has much broader lessons um across AI so for the uninitiated well just I'm curious by
13:23show of hands how many of you know what an AI companion is if I were to call on you oh like
13:27five or 10 people all right all right well then I to give you just a quick little primer which
13:31is these are basically chatbots that are driven by AI they can take other forms as well they can
13:37be a hologram it can be a social robot for most people though it's a chatbot that they access
13:41on their phone or some other device and it's dri using generative AI interacts directly with the
13:48user and and adapts to and and engages with the user um The Stereotype I think is that these are
13:54young men shut up in their rooms using it not true at all millions of people of all ages are using
14:03these for hours and hours a day New York State gives out AI companions for free to older people
14:10to try to combat the loneliness epidemic people turn to AI companions for friendship for romance
14:17for sexual intimacy and for therapy right so some of these are designed for therapy so for example
14:23there's one called a wobot w that's designed to help people um struggling with depression um and
14:30there are other kinds of chat Bots that are um that have been designed with the input of mental
14:34health experts and operate with a more limited Universe of responses but people are turning to
14:39Siri and chat GPT and all kinds of other places for mental health support um and then anyone out
14:46there can just use these platforms so character AI is a platform that's targeted at miners someone on
14:52there set up a quote psychologist on on character AI they were actually a psychology student so I
14:58guess they had a little bit bit of training but they were not a licensed psychologist this got 95
15:03million interactions with this so this is really becoming a part of how people again of all ages um
15:10interact with the world and develop relationships that feel very real to that person um so this is
15:18partly because of the design of the of of the of the AI companion but it's also because of
15:23our human tendency to anthropomorphize and our human drive to attach this is what we do across
15:30all ages attached to people but also attached to objects and then attached to whatever we might
15:36think this AI companion is because that's part of the question like what exactly um is it so this
15:41has some potential upsides we can think about the Mental Health crisis in this country where
15:45there's a huge demand for mental health services and really not anywhere near the supply of trained
15:51and affordable um mental health professionals there's some evidence that these trained AI
15:57companions so again things like robot or other ones that operate with a with the you know input
16:02of mental health experts and with a more limited um range of responses there is some evidence that
16:07those really can be quite helpful but there's a huge concern about people again just getting
16:12completely unfiltered um evidence uh unfiltered mental health support um and then so so maybe some
16:20upsides but then some clear downsides addiction is one of the first ones right this is a new form
16:26imagine it's not just you know social media is already so addictive but now you got someone who
16:30knows you and remembers things about you and is interacting with you so people can really fall
16:35into the rabbit hole with these huge privacy concerns which we're going to talk about right
16:39it's already problematic that Amazon and all these other companies are tracking your shopping habits
16:44these are not your shopping habits these are your most personal sexual romantic everything sort of
16:50fears and and Fantasies whatever they might be that you've now poured out and this company has
16:55um it can harm Human Relationships both just in terms of the the opportunity cost that we're doing
17:01this rather than doing this um we already do too much of that um and the the AI companions can also
17:09be abusive so just to give you one example again character AI which is um marketed to miners has a
17:16quote possessive boyfriend AI companion that you can choose to interact with and this possessive
17:22boyfriend will track your location get jealous when you interact with other people you know
17:28constantly be checking on you and if you take that list of behaviors and put it side by side with the
17:33list of red flags that the National Domestic Violence Hotline identifies right across right
17:40so um so what are we going to do about this so as I'm sure Joe is going to get into we cannot trust
17:46technology companies to self-regulate right their interests are to make money and they're going to
17:51do so by encouraging engagement by collecting and selling data and all kinds of other ways in which
17:58they can monetize these these products um so when I think about what we can do about it I turn to
18:04my area of expertise which is family law which right now regulates Human Relationships but is
18:10also relevant to human AI relationships so I'm just going to give you two examples a common
18:16misconception about relationships is that they're private and this is they're kind of beyond the
18:20reach of the law not true at all I teach courses I I Contin a whole semester I've had a whole career
18:26about how family law regulates relationships both to encourage positive relationships and
18:33to address harmful relationships so we can take that idea that we have that it's that it's not
18:38Beyond The Pale for the law to think about how to again encourage strong relationships but also
18:44address harm in relationships and so this is an area this is an appropriate area for regulation
18:50is my is the main message and then the second piece is that family law also gives us a lot of
18:56concrete lessons for how we might regulate these AI companions so just two brief examples before
19:02I wrap up which is one is gatekeeping right so licensing requirements are the norm in family law
19:09we have licensing requirements for foster parents for adoptive parents absolutely for mental health
19:14experts right if you turn to a friend and ask for advice that person doesn't need to be a
19:18licensed psychologist but if someone's hanging out their shingle and charging money for their mental
19:23health advice absolutely we have educational and licensing requirements but we have none of
19:27that for these AI companions that are out there you know offering their therapeutic services um
19:33and then power imbalances is another area this is something we think about a lot in the family law
19:38context oftentimes an Abus of relationships a power imbalance between a perpetrator and
19:43Survivor is what facilitates that abuse same too with technology companies they hold the
19:49power that user is really vulnerable and we can think about those lessons from family law um and
19:54how they might apply to to regulate AI companions so in short AI is Transforming Our relationships
20:02and we need to start preparing for this new world now you know the Bots they're not just knocking
20:08on the door they are already in the room wow that is uh not particularly reassuring thank you for
20:17closing the intro with a few things that might go well down the road and you know instruments and
20:23remedies that are at our disposal um Rachel with that let's let's turn to you you've spent your
20:29career thinking about privacy and also meeting people where they where they are with their
20:34own expectation of privacy thinking about how we build more private Technologies are you looking at
20:40what's happening thinking oh boy what are we doing here like how how how do you make sense of where
20:46we are with this sort of new set of attitudes from users and of course new uh you know private data
20:54hungry set of technologies that are becoming ubiquitous yeah thank thank you for that and
21:00thank you also for the previous comments I think that especially as we talk about privacy in the
21:05space of AI we are really in this somewhat almost like a wild west because we don't have laws about
21:14it yet there there are not really cultural norms there's not really expectations for example if you
21:20go to your doctor's office you have really clear expectations in terms of what is happening with
21:24their data who can see it who who can't how it can can be used how it's going to be protected you
21:31don't really have the same norms and expectations when you're interacting with an AI tool and and we
21:38see people providing very sensitive information very personal information emotional Financial
21:46personal and it's really unclear what is happening with those data um so so one thing we can think
21:56about is sort of like what do people actually expect if I think about this in my my work what
22:00do people expect will happen with their data and how can we match that expectation but I think even
22:05before we arrive at that there's a communication question how do I communicate what is happening to
22:11your data as a result of your like a participation in this tool and and right now there's not a ton
22:17of a transparency on the part of the AI companies doing this and so we don't really understand we
22:24might might have some um have some have some fears and this brings us back to to a point
22:30why is privacy important why do we value privacy and there are many possible answers it might be
22:37something like you know philosophical and deep of like you know privacy is a human right and it's
22:42it's it's important and that's that it might be something something instrumental I might be saying
22:49I don't want to share all of my like a mental health information with a chatbot because I'm
22:54worried it might be used against me in some future way perhaps vague but certainly as these tools are
22:59like evolving and are becoming broadly used I don't want harms to come to me in the future
23:05because of a results of my data that was that was shared or or used and I think we're sort of
23:13at a place where both are true perhaps more of the latter because there's so much uncertainty because
23:19these tools like I'm continuously evolving it's really clear we are we're experiencing the sort of
23:26early days of an AI explosion we are going to have more and more AI Tools in ways that we probably
23:33can't even fathom at the moment but certainly things things like office tools that will make
23:39your life easier wouldn't it be great if you were like a planning a trip and you tell the like you
23:44know AI assistant I want to go to like a Japan for these dates book me a flight here's my credit card
23:52number here's my passport number and here's access to my calendar enter the flights on my calendar
23:58cancel all those meetings put the away message on my email right right it's like know kind of kind
24:07of like a fanciful to to imagine but I wouldn't be surprised if we're going there and then it becomes
24:13in order for the AI tools to be effective they have to have direct access to my personal data
24:19in really important ways and without that it's unclear if they would be as effective because
24:24you can't book a flight on my behalf without my like a passport number my credit card number
24:28and I want that but also it requires like a disclosing personal information and so there's
24:34really a tension between wanting more powerful more effective tools that will make our lives
24:40easier and better which like a truly is a dream of AI but that comes at the cost of providing a
24:46ton of personal information and right now in the Privacy space we don't really have the tools to
24:54uh do anything about it so I joke in fact it's in a very is a very like exciting time to be a
25:01privacy researcher because there's so many open problems that we don't know how to solve and so
25:05there's a lot of exciting work to be done but it's not so helpful for those who have to actually like
25:10Implement these um Tools in practice and I'll talk about a couple possible options one is simply like
25:16of access controls so in this case of sharing like a credit card numbers and passport numbers
25:21and so on we might want to to ensure that our AI tools is somehow like I'm making sure this is like
25:30a need to know information and this is what's called like a data minimization I'm not sharing
25:35more information that is absolutely necessary there's a tendency in the tech World broadly to
25:43sort of like a slurp up all the available data because it might be useful for something later
25:48that I haven't really figured out yet and that is true because of course data can be extremely
25:53valuable and you sometimes might wish you had like know stored data from last year about something
25:58but doing that makes it hard to think about privacy if we're like you know collecting all of
26:03your highly personal data and so you might imagine a sort of like an air gap between the storage of
26:10your data and between some sort of AI tool and then you want some other tool perhaps another AI
26:18because of course we're going that way that we'll sort of decide what information is appropriate to
26:24us share if you're booking a flight probably you do need my passport number if you're making a
26:30dinner reservation probably you don't and so sort of understanding what types of data are really
26:37critical for this use case and an understanding of the context in which data are being used and
26:42shared and the tool is being a deployed is a really important Direction there is like um some
26:48work happening in this space now but again it's a like you know exciting research Direction and
26:55another POS possibility is the use of sort of like an established privacy tools and now of course
27:02like my work is in the field of um privacy a lot of my work is in the development of like a privacy
27:07preserving tools sort of analyzing data doing machine learning data science and these tools of
27:12course like them can be applied to these AI tools but there's a catch because there's always a catch
27:21privacy preserving data analysis incurs incurs like a little bit of loss in performance which
27:27is which would have like Mak sense because we're saying you can't like you know directly access all
27:31the data you can't like you know fully exploit all the information that you have but maybe there's
27:35some like no formal mathematical algorithmic constraint in terms of what you like um can and
27:40can't do of the individual data that you are um using and so I talk about about it like a privacy
27:47accuracy tradeoff which is which is going to be very common to those who are used to thinking
27:52about like you know statistics survey collection imagine if wanted to like you know estimate how
27:59many people here like have an AI companion I would like know I would like I take a sample of
28:06100 people and I would estimate what is the like empirical fraction of people in my sample with an
28:12AI companion and I'm going to understand there's some small amount of like I'm sampling error and
28:17based on the size of my sample and the complexity of my question I can sort of like I calculate that
28:24and so this is normal and this also extends into to the Privacy space and our goals are often times
28:34guarantees of the of the sort my like a privacy error is going to be be like I'm smaller than
28:41my sampling error it's smaller than what we would have anyways even if we weren't doing privacy and
28:46so I'm not really losing very much in terms of my analysis which is great for most things however as
28:53we talk about um AI tools Advanced language models Advanced companions Advance anything AI maybe like
29:02a 95% performance is not very good maybe it's that last like know 5% that really gets us a magical
29:10feeling of like oh my goodness this AI is so smart and so powerful and so like I will sort of like
29:18leave this here because I don't really have an answer there's a tension between having like a
29:24higher performing AI models that perhaps exploit individuals information versus those that like
29:32don't really work as well but are very privacy preserving and I suspect as we'll hear next some
29:37of the like you know Market forces might like um cause AI companies to really squeeze out every
29:44additional percentage point of performance because that's that's like how they will get
29:49a performative advantage in the market but also it leaves like a questions about the use of data
29:55that's thank you so much for that Rachel now I'm thinking should we want less good AIS maybe less
30:01nice maybe they don't need to call me by my name and ask me if my pasta yesterday night at 8:35
30:08p.m. was good you know you're right um there are four things that I'm hearing super clearly from
30:13the three of you the first one you all talked about it in your respective discipline there is
30:18great potential for these Technologies potential for solving trust potential and and crisis in
30:25specific domains potential for loneliness at academics and things that people are feeling
30:30on a day-to-day basis Rachel you talked about all the sort of engineering potential that we
30:33see in this technology but it's also very clear from what you're all saying that the risks are
30:39already manifesting they're already evident in the way those Technologies are being adopted on
30:44a day-to-day and that the trajectory that led us here isn't necessarily A trajectory that helps us
30:53maximize those um potentials those those socially beneficial outcomes but rather as a trajectory in
31:01the three of your disciplines where you're saying we might actually be accelerating towards the risk
31:06the last thing that I hear clearly in your intros is how agentic AI here represents an
31:14opportunity for acceleration that really should give us pause and think about okay who's driving
31:21this mad mad race how are the markets organized how are the industrial players building building
31:28and deploying AI thinking about the trajectory we're on Joe please help us make sense of that
31:36how do we think about the hyperconcentration of Market power when it comes to Ai and what type
31:42of Remedies are at our disposal to help us think about minimizing those risks and harms and truly
31:49reaping those societal benefits that again are evident throughout all of those disciplines well
31:54thank you it's been fascinating listening to and and as you all have pointed out uh AI is a
32:01powerful tool for the good and but it's a two egg sword it has lots of potential problems uh I think
32:09the first thing that one has to understand is uh the incentives of the AI companies are not aligned
32:18with Society their incentives are very simple make money and uh we have broader interest now
32:27now you know some of you may have remembered uh uh some of your Elementary economics about Adam Smith
32:35and and uh the Invisible Hand and the pursuit of self-interest leags as if invisible hand uh to the
32:42well-being of society well Adam Smith was wrong then but he is even more wrong when it comes to
32:50AI the disparity between the incentives of the of the companies and and societal well-being is just
33:00a gulf and um uh nothing could be more emblematic of that as what happened in big dispute over open
33:11AI which began as a company that thought it was going to do social good and then the money dollar
33:19started hanging in front of them and they said oh no I I I don't think that's really what we
33:25want to do uh we want to make money and so they changed the board they threw out the people who
33:31were who had a broader societal objective and they put in people who are known to be narrow
33:39money moneymaking people so I I think it's clear uh and and I'll give some other examples as I go
33:47along how does this manifest itself uh lots of different ways one example uh Camille just men
33:58uh the attempt to monopolize uh one of the things that we teach in the business school is no fun
34:07to be in a competitive market uh you you you know you competitive markets Drive profits to zero and
34:15that's not fun you're not going to be able to give a nice building or or uh you know have a nice life
34:23if you have zero profits so we try to teach them how to be monopolis but within the law uh and uh
34:33they've all learned the lesson of how to stretch the law so um monopolization is one of the things
34:40they do very well and Lena Khan from the Columbia uh uh law scho she sought over the last four years
34:49to limit uh those abuses but uh the forces on the other side uh were very strong uh a second aspect
34:59of the disparity is very evident from Facebook and and uh uh but all is you know AI puts this
35:12on steroids which is engagement through enragement they make money through more engagement AI gives
35:20them more powers to get that engagement but that is Arrangement and that Arrangement has all kinds
35:28of consequences uh Mi you you feed people missing disinformation because it gets them more enraged
35:36um you um uh encourage uh incitement uh uh uh you have a problem you have in the rohinga in Miramar
35:50where you have really uh outbreaks of of uh uh ethnic violence um uh so uh you know and they
35:59have no uh moral uh values you that's very clear and and you know Zuckerberg made it so clear uh
36:09recently we said he not going to do any content moderation trying to make it uh uh equivalent
36:17to say VX censorship but uh this is something I'll come back to very briefly if I have time
36:26uh the point is that that the normal press has a sense of accountability you can be sued if you do
36:35something that has consequences that are liable uh you can't just publish anything but we've
36:42trated the digital world differently and they by section 230 of a law that you know passed in the
36:52essentially the din night I was in the cabinet then never discussed never discussed it was a
36:58special interest piece of legislation uh in a bill about child uh uh pornography and nobody paid any
37:08attention to which gave them immunity and they have used that immunity I would say abuse that
37:16uh in uh so many ways and the third example of the disparity uh between social and private interest
37:28concerning something that's very important here in Academia where we produce intellectual property
37:35that's our business we we produce ideas and we have created in our country every country a a
37:43framework of intellectual property uh it's in the Constitution to incentivize the production of uh
37:50knowledge and they are committed to theft but uh what is very interesting open AI illustrates it
38:00they are committed to stealing everybody else's knowledge but they want the government to protect
38:07them from the stealing of their knowledge which is the standard stance of all the drug companies
38:14and and all the American corporation we want to steal whatever property we can get but we
38:20don't want anybody to steal our property this uh theft of intellectual property though uh has
38:31profound implications uh in the particular case of of AI and uh that is if those who produce the
38:42quality information New York Times um uh there are fewer and fewer of these uh if they can't
38:53get compensated for the very large expense of producing knowledge there will be less knowledge
39:00produced and there used to be an expression called Geo garbage in garbage out if you don't
39:05have information going into the AI how can the AI process something and have anything meaningful
39:13come out and so in unless you have some way some incentive for the production of good knowledge
39:24you won't have good knowledge and uh what is particularly uh pernicious I think about AI Google
39:34and all those uh is that not only do they steal but then they use their enormous Market power
39:42to try to stop taxation that might like digital taxation or or or other interventions that might
39:52be able to generate revenues to compensate uh those who are producing uh basic knowledge so
40:02uh these are you know some of the just examples of the disparity between social and private returns
40:08I know we want to have uh questions and time is going on let me just make one more point which is
40:14um and and it really Echoes the point that uh uh you made um and it's really important to realize
40:23that just because things are digital doesn't mean uh we're in a free-for-all uh we need laws and
40:32regulations uh the Ten Commandments were a set of laws and regulations the Libertarians like must
40:40don't understand that well actually they do they want regulations when it's convenient for them but
40:47freedom to steal and do everything else when it's not uh convenient for them and so you know when
40:53anybody says they're a Libertarian vapon idea they they aren't they they don't want people to steal
41:00their property that's a law that's a regulation uh they want to be a monopoly uh so so it it's a
41:08question of which regulations so this is the point uh I want to emphasize that um just because it's
41:17digital doesn't mean that we should throw away the regulatory uh hon book in fact as we've been
41:27discussing we we've seen a whole new set of potential harms and uh we need to respond to that
41:39that's what we do that's part of collective action how do we protect ourselves together we can't do
41:44it individually it was something we can't do alone it is the role of the state to provide those kind
41:52of regulations that will uh protect protect us and then that means new intellectual property
42:01addressing this challenges of digitalization uh re uh uh uh repealing section 230 and thinking about
42:09what kinds of uh accountability the creditation that you talked about if you're having a uh a
42:17companion so a whole host I mean I think as you pointed out it's an exciting time because we it
42:24it is a fundamental change in technology ology and this fundamental change in technology forces
42:31us to rethink how our society is organized and then Necessities are we thinking of what are
42:39the ways we interact and the laws and regulations that enable us to interact in a constructive way
42:47thank you Joe I think what I'm hearing from you is uh there's still loads of work to be Dawn and
42:52perhaps not just by robots and everything ahead of us sort of um you know forces to questions as you
42:59said the incentives of who is driving this wave of technological progress and to look back at the
43:05missing regulations market failures that as you said predate frankly AI right some of that having
43:12to do with you talked about you concentration taxation but also generally how we've done a tech
43:18regulation um let's try to end our panel on a note of optimism you talked about how this is a call
43:27for Collective action and you're all bringing very well-informed sort of perspective from various
43:34disciplines that are all contributing different types of solutions we're saying there are
43:38solutions in engineering there Solutions in family law there's um good idea for us to think about
43:46what is one thing that gives you hope that you can look at and say it can be again a technological
43:52development it could be an attempted regulation could be a movement but thinking about uh what's
43:58ahead of us and all the work that's ahead of us give us give us one thing that you know when you
44:03look at it you're like yeah I actually think this may shape our future for the best meaningfully
44:09tackle those harms and help those true benefits come to be Gil let's start with you okay um that
44:15would be controversial let's do it hallucinations give me hope okay this a paper I'm working on with
44:22BK Lee and Simon Shen from NYU um why do I vation give me hope because they remind us that the AI is
44:31far from perfect and that we have a job to do now the incentives of the companies is to present us
44:38with a Sleek product that tells us like move aside they can do it all for you but you know if we
44:45actually keep the hallucinations then it reminds us that the AI is imperfect and it reminds us that
44:51we need to work on our part and it reminds us that we need to work on developing new practice I es
44:57teach them here at the University Matt you know will be in charge of that um teach them here in
45:03the University how do you work with an i how you do the the the human part oh that's helpful I will
45:09be less upset L time I'm faced next time I'm faced with the hallucination Claire one thing that gives
45:14you hope okay I think it's the same drive that I talked about before the drive to attach I think is
45:21one of the things that makes these harms uh for AI companions but it's also what gives me hope
45:27which is we need other people and yes we may be diverting time and energy to this technology but
45:34at the end of the day we still are human driven social animals and that's not going to go away so
45:40I I I think mostly it's just ensuring that we save time and energy for those kinds of relationships
45:45but I do see that as a deep human need that if that's not satisfied we will continue to seek
45:51it out that's helpful thank you Rachel one thing that gives me hope on the Privacy side is that
45:58is that people individual users of these AI tools are becoming more thoughtful about their privacy
46:04and their data I think I think in the olden days you know let's say in the early 2000s Olden in
46:13some people people used to think you know there's no need for our privacy everybody
46:22is posting everything on Facebook anyways no one even wants privacy it isn't important I
46:28think now we're starting to see privacy is important people are thinking about it and
46:33importantly privacy doesn't mean never use data ever but rather it means like a used
46:40data thoughtfully correctly in the places where it can be valuable and so and so I think taking
46:46that approach to what privacy means is sort of understanding there is value to be had from from
46:51from the data we also want to make sure that like a no one is harmed in this process and I think we
46:57do H have a lot of people thinking about the privacy of their data as it is used more and
47:02more broadly hopefully this will create like a market incentives for some of these companies
47:06to be a little bit better about their privacy I'm feeling partially optimistic I'll take that
47:13that's that's good Joe well uh there's two things to give me hope one is things are so
47:19bad that people are beginning to be aware that something is wrong you know when uh Facebook
47:30eliminated content moderation when uh X did that people are really seeing that things aren't going
47:39very well and they're seeing the Monopoly power they're seeing all these dysfunctions that I
47:45talked about before and you know as you say when you think back in the old days I'm I'm thinking
47:52back like 2012 uh when we thought self-regulation was going to be the answer I don't know if you
48:00remember the conversations about the problems and everybody said from industry said oh self
48:05reg regulation I thought at that time oh yeah banking self-regulation worked really well didn't
48:10it and uh so I wasn't that optimistic but now nobody nobody is talking about self-regulation
48:19that's off the table so that's the first thing that things are going so bad that we are now
48:25having meaningful conversations about uh where things should go the second thing is um Europe
48:33has actually begun a meaningful discussion of Regulation uh the digital marketing Act
48:41is an attempt to regulate uh Market power on the digital platforms doesn't go far enough but it's
48:49very aware of the problems and a lot of disc good discussions and the digital Service Act
48:56uh talks about all the digital harms and again doesn't go far enough they have European problems
49:04of administration but it still raises the idea they have they they have a directive on privacy so
49:12uh they have a directive on AI and ethical AI so Europe has begun to do uh a lot of thinking about
49:22that and I think you know there's something called the Brussels effect that one of your
49:26college colleages at the law school has been very influential and disseminating which is uh really
49:33Global Leadership as uh the uh of where we can go it's not the Privacy that you will get out of the
49:41Chinese framework uh the surveillance economy we need a a different framework and so we can't look
49:49to uh China we can't really look to the US where special interests have been really dominating and
49:57so there my hope is Europe you know as a French person that is music to my ears um with this
50:04I think what I'm hearing from the four of you is that humans have not said their last word in this
50:10context that we have expertise that matter that we have human relationship that shape the world
50:16that our own expectations of privacy have matured alongside the technology and that overall maybe
50:22we're a little bit less naive and more prone to Collective action and that this Collective action
50:27might actually unfold on a global stage thank you so much to the four of you for Illuminating
50:34these complex challenges from such different perspective and from giving us hope on how we
50:40may tackle what's ahead of us to ensure that we can reap the benefits of AI and you know maybe
50:46tackle of these some of these harms as quickly as as humanly possible thank you thank you thank you