Is it possible to regulate AI before the next generation pays the price for our inaction?
In this episode of Your AI Injection, host Deep Dhillon talks with Jai Jaisimha, founder of Transparency Coalition.AI, about what it would actually take to hold the AI industry accountable. Jai explains why protecting kids is his top priority, especially as companion chatbots quietly become confidants for millions. He shares one tragic case that's become a rallying point for new legislation: a chatbot conversation that started as homework help and spiraled into self-harm. Deep and Jai dig into why this keeps happening. Safety filtering costs money, and too often companies choose to save it instead of protecting users, leaving AI development stuck in a black box where safety reports have quietly disappeared. Could independent testing force the transparency this industry has avoided for years, or is it already too late?
Learn more about Jai here: https://www.linkedin.com/in/jaijaisimha/
and Transparency Coalition.AI here: https://www.transparencycoalition.ai/
Check out some of our related content here:
Get Your AI Injection On The Go:
xyonix partners
At Xyonix, we empower consultancies to deliver powerful AI solutions without the heavy lifting of building an in-house team, infusing your proposals with high-impact, transformative ideas. Learn more about our Partner Program, the ultimate way to ignite new client excitement and drive lasting growth.
[Automated Transcript]
Jai: And I think where things went sideways was that the, Very often, if you just leave it to the mathematics of it, you'll find that content that is outrage inducing or anxiety inducing or worse, can actually drive more engagement than content that is potentially not as outrage inducing.
Jai: And so the algorithms were let run free. the safeguards, uh, to the extent that they existed, were overlooked. The people who provided, internal warnings were disempowered.
CHECK OUT SOME OF OUR POPULAR PODCAST EPISODES:
Deep: Hello, I'm Deep Dhillon, your host, and today on your AI injection. We're joined by Jai Jaisimha Jai holds an MS in electrical engineering from Tulane University and a doctorate in electrical engineering from the University of Washington, where he is now affiliate professor.
Deep: Dr. Jaisimha is the founder of the Transparency coalition.ai, where his work sits at the intersection of AI and public policy focused on making these systems safer and more transparent. And today we'll be exploring AI [00:01:00] transparency and what accountability in the space actually looks like. Jai, so good to see you.
Jai: Likewise. The really happy we're doing this.
Xyonix customers:
Deep: Yeah. Um, just for everybody else's benefit, Jai and I go way back. I don't even know. 15, 20 years or something is my guess.
Jai: Yeah, those are
Jai: 25 at this point.
Deep: Oh, yeah. Yeah. We've both been slinging AI models and machine learning stuff and running engineering teams and stuff for, for decades.
Deep: So Jai, why don't we kick it off. What problem are you guys trying to solve and do you have any lessons that you're taking from the social media addiction trials?
Jai: Definitely the problem we're trying to solve is you know, there's a gap in the creation of policy where a lot of the people that are trying to make policy don't have domain knowledge or expertise.
Jai: Uh, so they don't actually know what's possible versus not what's hard, what's easy. they also don't understand much about how decisions made inside large companies. And so trying to figure out how [00:02:00] policy can take the existing processes and ways that safety and transparency or practiced inside companies and try to amplify them with their regulations is something that they don't have the insider knowledge to do that kind of thing.
Jai: So we provide that, uh, as a. Compliment to what, uh, lawmakers are trying to do primarily in states. So that's the main thing that we're focused on.
Deep: It's kind of like a sounding board slash translation layer or something.
Jai: Yeah, partly. Originally we thought it was mostly about being a translation layer.
Jai: Uh, now we realized that we actually have to provide the frameworks as well, uh, to create policies. So not just plugging gaps in policies that they're coming up with, but also coming up with entire model frameworks which are called model bills in this space to help people. Uh, you know, it's, it's almost like a policy in a box, uh, ready to go with the levers, thought through you know, some knowledge of state law and how it fits together.
Jai: Certainly, awareness [00:03:00] of the US Constitution, uh,
Deep: yeah,
Jai: it operates, but so that's, that's increasingly become a big part of our work as creating that model policy. And leading with the model policy where that's appropriate or filling in the gaps when they've already,
Deep: yeah. I mean, let's, maybe let's start with the, so I brought up the social media addiction trials 'cause a, they're super timely in the news, but you and I were chatting maybe a week or two ago and I think you said something really interesting, but I kinda want you to maybe bring it up here.
Deep: Like what do you guys feel happened well, or not so well with the social media, you know, technological evolution and what are sort of the lessons from an AI vantage when we start thinking about regulation?
Jai: Um, I think the one thing, there's sort of a couple different things before the trials even happened, right?
Jai: There were some practices around business practices around, uh, how social media companies operated. I know this because one of, [00:04:00] in one of my past lives. I ran a social media advertising targeting company and
Deep: mm-hmm.
Jai: Which I know you and I worked on a little bit together too, so
Deep: We did indeed. We did,
Jai: yes.
Jai: Yeah. But that, uh, one of the things we always knew that engagement was something that was, you know, super prized because engagement usually translated into business results. You know, in the case of advertising, you want engagement because that typically precedes conversion. in the case of, social media companies, they knew that they needed people engaged in order for people to see more ads, and if they saw more ads, they were more likely to click on the ads.
Jai: And that's how the social media companies made their money. It was all about driving people's engagement. And I think where things went sideways was that the, Very often, if you just leave it to the mathematics of it, you'll find that content that is outrage inducing or anxiety inducing or worse, can actually drive more engagement than content that is potentially not as outrage inducing.
Jai: And so the algorithms were [00:05:00] let run free. Uh, the safeguards, uh, to the extent that they existed, were overlooked. Uh, the people who provided, internal warnings were disempowered. Uh, you know, there's plenty of documentation of that now.
Deep: Yeah. And the, and the ramifications were not at all subtle for our children, right.
Deep: Like we, I mean, anyone who's raised a Gen Z kid, um, or even an alpha kid knows the incredible harms that have come from. That lack of regulation in the early days or really much, pretty much ever. We have elevated suicidal ideation rates amongst our teen girls in particular, but also amongst the boys.
Deep: We've got anxiety outta control,
Jai: body dysmorphia. It's just o over and over over again. You know, both you and I are parents of, you know, kids who, who lived through that and, and of course so I think that's kind of, I think what the trials have done is they've finally brought these things to light.
Jai: You know, you had whistleblowers talking about it, but there's nothing quite as [00:06:00] transparent as, uh, the legal discovery process. So when you have, when you have these lawsuits moving forward and courts that are willing to, invest and look at the claims, I think I'm starting to realize what a powerful investigative mechanism.
Jai: In a, in a country where we don't have robust, uh, you know, federal or state level investigative bodies, uh, focused on this problem private industry through the plaintiff's bar is providing this, these services. Right. So they're definitely pro, uh, it's a, I mean, they're not doing this because they're charitable, they're doing this.
Deep: Oh yeah. They're gonna get a check.
Jai: They're, they a really
Deep: large check, apparently.
Jai: Yeah. They hope to get a check. But the flip side is that they're performing a huge public service, which would not have happened without, uh, you know, much of the information wouldn't have come to light without their
Deep: Yeah.
Deep: I mean, this is something that always sort of mystifies my European friends is how our system works because we're so regulation light and [00:07:00] so, judicially, you know, heavy in terms of. How our system works. Like we really rely on these lawsuits to fill in all the gaps where our legislators fail.
Deep: And it's, it's definitely not what happens in, in, you know, in many countries, in Europe and many countries.
Jai: Yeah, I always tell people, you know, if you think, uh, you know, toward lawsuits are terrible, try growing up in a country where they don't exist. Uh
Deep: Oh yeah. Or with a dysfunctional court system, or,
Jai: yeah.
Jai: Any, any of those things. So,
Deep: yeah. Yeah. Like, uh, I mean, I have you, you know, you and I are both spent a lot of time in India, so I have a family property dispute case that got settled. I think like five or six years ago, and it took 48 years to settle the property dispute, 48 years, multiple. It went, it went three generations.
Deep: It was like my great grandfather's brother started and filed the case during partition and it like got settled. So it's not [00:08:00] only whether the court system is doing the right or wrong thing. I mean, it's, is it funded well? Is it efficient? And I feel like that plays in here because we might be getting remedies from the, from the judicial side, but they're coming, I don't know, 10 years late, you know.
Jai: Yeah, no, they're definitely, it's not the fastest potential way of produ, you know, proceeding. And I think that's where we feel like the fact that ai, uh, that social media harms, happened over DA over a decade or more. I got my Facebook account in 2007. So, you know, it is been around for almost 20 years at this point.
Jai: And, uh, you know, there I think the what's been interesting is that we are able to use that data point. This legislators all over the country are saying, oh wow, we didn't do anything.
Deep: That's right.
Jai: But we were warned about it all along. In some cases we didn't do anything. And now all this, these harms are happening.
Deep: To what extent do you think that that's a partisan [00:09:00] perspective? Like Democrats saying that and Republicans saying, no, let the market figure it out.
Jai: I find that there's two different types of there's typically in, in states that have a lot of technology presence. There's both kinds of Democrats, uh, and there's both kinds of Republicans in states that don't have a large technology presence, many of which are Republican controlled.
Jai: They are actually much more suspicious of big tech and are much more willing to take that. Hmm.
Deep: And
Jai: true like Nebraska, Utah, you know, these are, uh, I mean Utah is not, does have a tech presence, but they are more vigorous in trying to protect children and families in particular. And so they tend to be more vigorous about it.
Jai: Uh, states like Washington and Oregon are definitely in catch up mode. You know, they're doing much better this year than they did in years past in terms of trying to protect people with robust legislation. Uh, California has had a mixed record, but they're also moving forward. So I think in general, but Texas for [00:10:00] example, has a really robust ai, uh, regulation that they passed, uh, in the 2025 session that includes a lot of very.
Jai: Strict restrictions on, uh, things like, uh, non-consensual intimate imagery that affects children, things of that sort. Right. So they're, they've selectively focus, uh, on some topics, but in general it doesn't seem to be a partisan issue. Thankfully at this point. Uh, you know, people keep trying to turn it into it, but it's, uh, at this point it's, we work with a large coalition of you know, groups on both sides of the aisle.
Jai: They're mostly from the nonprofit side of things, but they're, they advocate for all sorts of policies. But when it comes to AI and tech we tend to be pretty well aligned. So,
Deep: and to what extent do you find that just the general knowledge is a problem? You know, going back to the social media trials, I think it was the Facebook trial of when Zuckerberg was, uh, testifying to the senate.
Deep: I can't I think it, [00:11:00] correct me if I'm wrong, I thought it was Mitch McConnell who was asking a whole line of questioning and, and Zuckerberg, you know, like, I don't understand how you make money. He is like, well, Senator, we run ads and all of us in tech were like, you know, palm on forehead. Like how can this guy with this level of knowledge be the one that's supposed to do or not do something?
Deep: And I'm wondering, do you have that problem where you just have extreme like tech, non savviness trying to decide and legislate?
Jai: Yeah, I think yes and yes. Right. So, like in, typically in our work, we work primarily at the state level. There are change agents who are tech savvy.
Jai: Uh, there are, uh, frequently, you know, mothers of young children who are also lawmakers. Uh, there are people who can see that this social media movie is gonna play out again. Uh, so they're interested in taking action. One, one other function critical function that we perform is we educate other lawmakers who are not the change agents, right?
Jai: So the change agents need help convincing their colleagues to [00:12:00] vote for bills. So I, when I went to Salem a few weeks ago to help pass the Oregon law, uh, you know, we were, I had to go, I went door to door, met with, you know, key, uh, lawmakers, and just answered none. It didn't even have anything to do with the bill.
Jai: They just had basic questions. What is this? Why should I care? Uh, what, how does it actually work? Most people have not used AI to the, they still call it the ai, you know, so it's like, uh, it's, they, they you know, there's certainly a lack of understanding and they
Deep: really, you run across people who have not used.
Deep: LLM?
Jai: No, never. Yeah, never used an LLM. Uh,
Deep: the numbers are like, I think a billion plus now.
Jai: They would admit it freely in a hearing. You know, I have never used it. I don't know what it is. I've heard of it. And they're in a very key decision making positions, and so you have to get across to people.
Jai: So we resort to, the other thing we're always trying to look for is analogies. So we're looking for analogies to things they may understand. [00:13:00] Uh, so we frequently will talk about, cars breaks, you know, things like, so we try to make sure people understand what is this similar tool that you've already seen and understood.
Jai: Yeah, you don't have to understand ai, but you have to understand what it's similar to, uh, you know, um, and if you think you know, it's okay to regulate how brakes operate, then you should be willing to. You know, as an example. Right. You know, so we're always, or if you're willing to regulate, the food that is served in a school kitchen.
Jai: You know, hey, you should be willing to put equal amount of thought into like what software goes into the kids' laptop or how three things. Yeah. Right. You know, so I think it's, it's important to get to use analogies, uh, to effectively communicate, because many of these people may never become, you know, yeah.
Jai: Users, they may never, and then you'll meet people who are like, you know, building
Deep: all kinds of their
Jai: own, their own agents inside a So I've met people like that too. So there are there [00:14:00] are, I remember one lawmaker sent me his, uh, GitHub commit chart as a way to open the conversations. I'm like, oh, that's
Deep: great.
Deep: Yeah.
Jai: Hey, you have that. I don't, I don't, I haven't done a GitHub commit in many years. So, but it's, it's an example of, you know, it, it's all over the place, but it's certainly great to see more and more people in some states. Uh, there's a lot of really informed people. California is very lucky that way because they have a full-time legislator, legislature, and the a lot of people are, you know, former tech people who go into service.
Deep: Yeah. Let's kind of jump into the actual problems of ai. I think maybe a year ago the whole doom story was big. You know, AI's gonna jump outta the box and kill humanity. That seems to have subsided. Uh, I don't like it is not getting the kind of attention that it was. Now we hear, you know, we hear about all kinds of things.
Deep: We hear about the Homer Simpsonafication of people [00:15:00] just getting dumber from using ai. We hear about plagiarism and cheating. We hear about, you know, the, um, I think it was the Trump administration when they came up with that incoherent tariff policy. Somebody all, I'll get a bunch of count.
Deep: Economists were baffled trying to figure out what actually happened. And they eventually tracked it back to an intern that just was using GPT to like, come up with the policy. It's the only way they could explain it. So I guess the question would be like, what are the problem, like the core maybe categories of problems.
Deep: Maybe we start there and then let's try to get to like, what's your process for prioritization across those?
Jai: You know, I think happy to do that and I'll, I'll hit on some that are not on our priority list as well, just so to be try to be complete. And you know, we were always sort of skeptical of those DOR bills.
Jai: Uh, they occupied a lot of attention and in some cases they were actually helpful to us because
Deep: yeah, they just feel like a giant red herring. I mean, there's so many practical, [00:16:00] anytime you like overhaul, societal interaction, there's gonna be all kinds of practical problems and they're just focusing on these like fantastical ones.
Jai: Yeah. No, and I think in some cases it was. Maybe I'm cynical, but I thought it was a deliberate strategy to distract people with those
Deep: You are cynical more than I'm
Jai: with, with P Doom you know, conversations, but
Deep: I
Deep: thought I just attributed it to like excessive, geekiness.
Jai: I, I'd say sort of that today you'll see, uh, they fall into some broad categories.
Jai: One is there's still people trying to do what I call foundational privacy bills. So they, because they, they understand that data is the, main fodder for ai. So there's, uh, there's. Still continues to be effort to do, bills regulating what data can be, uh, used, what consent is required.
Jai: They're not really AI specific, but predominant use cases for many of these data streams is AI of some [00:17:00] sort. So that's one, one area another.
Deep: And so here you're not only talking about the data from a training and like efficacy impact vantage, but from,
Jai: and user data,
Deep: but from like a rights vantage.
Deep: Like who, like do they have the rights to use this or that?
Jai: Do they have the right can, you know, what can they use? What can't they use? Uh, how can they use the data? What permissions do they need? So that's, that's, uh, that's one one work stream. There's another work stream, uh, which, and I'm gonna hit on the ones that we don't necessarily focus on first and then talk about the ones we do.
Jai: The other one that we see a lot of now this year, and I think increasingly there's a focus on something called surveillance pricing. Surveillance pricing is a practice of, algorithmically altering the price of a product individually. The best way to think about it is it's not happy hour.
Jai: It's you walk into a, uh, the bar and they're like,
Deep: I know Jai drives that [00:18:00] car and
Jai: Yeah, exactly. So I'm gonna charge him $10 for this beer instead of six for, you know, somebody else. Right. You know? And so that uh, this is particularly cropped up, uh, because of the rise of delivery services, right? So delivery services, when the, uh, when the, uh, you know, price of, you might place an order for something and when they.
Jai: Tell you how much it costs you. That price may change based on your ability to what they think is your ability to pay. So that, is that
Deep: a real problem? Like, are companies actually doing this kind of pricing?
Jai: There's been testing that's proven them, that's happened and so there's a, uh, burgeoning stream, especially because of concerns around affordability, right?
Jai: People are very concerned about that. So
Deep: are there existing laws that already address, it seems like kind of a.
Jai: Nothing on the books yet. So this year you'll see a bunch of states. I mean,
Deep: it feels like, uh, it just falls underneath maybe more general discrimination or something,
Jai: or unfair or udap, right?
Jai: There's something called a unfair and deceptive trade [00:19:00] practice. So,
Deep: yeah. Yeah.
Jai: So, so yeah, that it's possible that there's existing laws on the books that can be used to but but you know, certainly that's, it's a, it's also politically very often you also realize what policies get pushed forward and don't.
Jai: That
Deep: one makes, like, it seems like it would be easy to push because whoever has money is gonna want that law.
Jai: Exactly. Yeah, exactly. No, it's a interesting, uh, line of work. Then there's a set of laws that are now kind of almost being deprioritized just because it's been so hard to get them done, which is, uh, around anti-discrimination.
Jai: So algorithmic discrimination. And the way those were typically done is that it was focused on something called algorithmic discrimination in consequential decision making. So consequential decisions are like healthcare, employment, housing, uh, those are viewed as consequential decisions. And so, that's kind of a third body of work and that, uh, if you've heard of the Colorado famous Colorado AI Act, uh, which was vetoed and [00:20:00] you know, was fought on and that was an example of that.
Deep: Yeah, no, there, I, I used to, you know, we used to build a a lot of models for the city of Seattle and one of their really top concerns was, and the use of any demographic information for for example, crime forecasting, like predicting where crimes are gonna happen. And so then, you know, like most of the companies that build those predictive policing apps, they ended up just settling in on simple lat long coordinate.
Deep: Timestamp crime type, and that was the only data they would train on. But nonetheless, they would still exhibit bias because the bias is like inherent.
Jai: They were inherent.
Jai: Yeah. The, yeah,
Deep: it's inherent to
Deep: the
Jai: Yeah. Where, where people live, you know?
Deep: Yeah, yeah, yeah. All those things. Yeah. So, so I think it's, it's, and it's a very non-obvious solution there.
Deep: I mean, if there is one, I don't know what it would be.
Jai: Yeah. And frequently they also, uh, the remedies are not clear. So this is one of the reasons we don't focus on it. The remedies aren't clear what is asked for. [00:21:00] We, I don't understand, like for me, I need to have a theory of change. Yeah. You know, this is gonna be the lever, this is the action, the model developer or the model deployer must take.
Jai: And this is how it's gonna actually reduce the instance of the problem you're looking for. And I haven't been able to connect the dots in my head. So we, we, we. It's also politically very complex because every stakeholder, so you can imagine all the banks, all the hospitals, all the insurers, everybody comes after you to try to influence the legislation.
Jai: So we're bigger fans now of, uh, what we call through use case specific. So if you want to do one that just focuses on, like, one of the things that people are trying now is prior authorization. So the ga, the granting of prior authorization before medical procedure. It's a very clean use case. It's very important.
Jai: It's life altering or, or threatening for some people, you know, because of,
Deep: oh, you're talking about the insurance companies, whether or not they authorize the procedure,
Jai: denial or acceptance. Yeah. It's a pretty vibrant category [00:22:00] that could benefit from additional logging.
Deep: I mean, it feels too surgical though, because like whatever you solve there only applies in that scenario. That's right. But it, but maybe you learn something in generalize. Is that the theory? That's right.
Jai: That's, that's right. And so that brings us to kind of one of our new areas of focus. It's, uh, which we call, um, uh, because there's always been debates about whether artificial intelligence, technology, how to define it in law.
Jai: And then the question is, you know, uh, because social media, uh, benefited greatly from something called Section two 30 of the Communications Decency Act. So that essentially was meant to prevent ISPs or internet service providers from discriminating against content
Deep: that's the law that says, as long as you don't look inside the stuff, you're not responsible for what runs around on your platform.
Jai: That's right. Yeah, that's right. And so, there have been our, there, I mean, it hasn't been tested in court, but there's been a feeling that, you know, social media, like social media companies were resorted to those defenses to avoid. [00:23:00] Absolve themselves of liability that people in the the AI space would also do the same thing.
Jai: So, so we are pushing for a new stream of legislation, which aims to classify AI as a product, which it is, it's not a service that mediate speech between two people. It's a, it's all, uh, it could
Deep: be though.
Jai: It could be, but, but, uh, but those two could be different, right.
Jai: You know, you'd have to regulate I
Deep: see. So that you're trying to like somehow anchor it outside of two 30. Yeah.
Jai: Yeah. So, because it, if, if it's machine generated outputs, those machine generated outputs are not human speech. They're not generated by anybody. They were, you know, they're potentially attachment hacked or engagement hacked in a way that will make you you know, engage the platform more or some for some other reason.
Jai: Or it could be a, you know, it could be an algorithmic decision making, uh, bill or could be any form of AI technology. And so we're actually pushing for those types of legislations is a more sort of future-proof way of regulating the [00:24:00] space. So you would end up with, uh, you could do rulemaking and laws that are more, prescriptive, but you could also do like this backstop, which is a, a bill that essentially we call a product liability bill, which is AI is a product and you should hold AI developers accountable the same way you hold other product developers accountable.
Deep: Oh, I mean, is that was ever in question?
Jai: There, there are, uh,
Deep: this is like the whole argument.
Deep: Like a a, a Tesla or a, you know, a self-driving car whack, somebody that the manufacturer's responsible regardless of.
Jai: That's
Jai: right. That's right. Yeah. Regardless of, so the, uh,
Deep: but has that ever been in question?
Jai: I mean, yeah, that was a defense that was presented. Like one of the, one of the cases, the first cases to settle AI lawsuits to settle the original defense was that it's not a product, it's a service.
Jai: You can't sue us. Uh, and then the court first ruled that it was a product, and then over a period of time there was discovery and then, you know, [00:25:00] Google and somebody else. Uh, I mean, you're liable for services too, aren't you? I mean, uh, it depends. That, that, I don't know. Um, like if you applies you somewhere, like if you post, if you post hate speech, I mean, there's always been an argument that if you post hate speech on.
Jai: On, um, like what happened in Myanmar, you know?
Deep: Yeah, yeah.
Jai: That Facebook allowed hate speech to, against, uh, Rohingyas to, you know,
Deep: oh yeah. Millions of people were affected, who died,
Jai: were killed in, or, or
Deep: yeah. And Zuckerberg didn't care at all.
Jai: Yeah. So I think, so there's, there's that type of argument, but then the, in the, in the US we are, we're, you know, I think thankfully we care a lot about freedom of speech and so there's a certain, uh, reluctance to get in the way of that type of speech.
Jai: So. Uh, or, or speech of any sort. Right. You know? So, uh, as long as you are not shouting fire in a crowded theater, and that is not allowed. But you know, if you're actually doing things that are as [00:26:00] where you exercise your own free speech. So in this case some of these cases, the first cases that are settled in the, was settled in the AI space.
Jai: That determination had to be made first. So the court had to say, yes, I've seen the evidence. This is now service. This is a product under, uh mm-hmm. The official definition. And so you, you have to be treated like a product and then, you know, discovery, et cetera, et cetera. So those things happen. And so we're, so that's kind of our broad brush approach to ai.
Jai: And then we have a, a set of bills that are. Emerging, which are all the rage this year. And that's around regulating what's called a companion chat bot. So these are chat bots, or these are all chat bots essentially can be used as companions today. Right. So you can go to chat GPT and say,
Deep: yeah,
Jai: you know, tell me what I should do about this or that problem, or, uh, and or answer this question as if you are my needy, boyfriend or whatever.
Jai: Right. You know, you could, you could [00:27:00] ask the chat bot.
Deep: Yeah. Yeah. This happens millions of times. Probably a second or exactly at least a minute, you know?
Jai: Yeah. So, and then it's, you know, children are shown to be disproportionately affected by some of these, uh, manipulative design practices. And so there are laws that are being prac, uh, we helped pass the first one in the country in California last year. That was a very basic law that basically said, Hey, if you're, uh, talking to a chat bot, the chat bot must tell you that it's not a human, it's actually a chat bot. The other one was, if you ask for, if you express, uh, suicidal ideation, then you know it needs to direct you to crisis line resources or humans or, talk to someone.
Deep: How, how do you deal with the, efficacy rate of any kind of detection? Let's take this suicidal, let's just take all the mental health kind of questions, right? I mean, I have have so many questions in this area from a regulation vantage, but part of me, I'm just gonna give you a little blurb analogy, react to whatever makes sense.
Deep: Part of me is sort of mystified that the FDA has [00:28:00] not gotten involved because you and I have both spent long time building healthcare products and the biggest concern is always like, how are we gonna get this through an FDA process? And we would never have imagined, just like issuing a generic tool that happens to like have critical health conversations with everybody in the country pretty much.
Deep: And the FDA has nothing to say about it. That's just mystifying to me. I don't understand it. I'm not saying that they should go and just, you know, like stop all those conversations, but it's mystifying to me. The other kind of question that is sort of strange to me is once you, you know, like I think this was maybe two or three months ago, somebody published a paper about, about delusion like embracing delusions of mental, mentally, um, like mental health problems. So you have a patient that's like bipolar or something, it's talking to Jet GBT or to Gemini or whatever, and the thing just says, oh, you think you're Jesus Christ. Oh, well, how's that? You like, just embraces it, which it turns out is the [00:29:00] absolute thing you're not supposed to do.
Deep: So once you pass a law that says, Hey, like how do you think about that law? Because you have to like a, detect it with some level of efficacy, and then I, ideally you want them to put resources in to try to increase that efficacy, and then once you do, yeah. How do, how do you think about all that?
Jai: I, I, I think the, the harsh reality is that, you know, the vast majority, we, we we're used to thinking that compute resources in ai, that the majority of it is used for training increasingly.
Jai: It's inference, right? Yeah. Inference is expensive. And in a world where there aren't enough people paying 20 bucks a month, uh, you know that this is, um, these companies are burning money, right? So, for example, one of the most efficacious, at least from, my reading of the literature libraries for detecting some of these natural language processing has, uh, you know, been around forever, right?
Jai: And [00:30:00] so the real question becomes,
Deep: some of us have a very large bank of useless knowledge. Now it's all
Jai: but the all been
Deep: replaced by these.
Jai: But the reality is when you're, you know, going down some of these weird rabbit holes that un unencumbered LLMs can go, go through. There's no, there is a tremendous amount of expense required if you want to try to filter for these different types of, so these companies are making an economic decision to prioritize
Deep: Absolutely.
Jai: Their cost per transaction over user safety, right? And so, so this is where laws can come in to try to change that calculus. Say, Hey, you need to, and where we are, you know, where we tend to be a bit more practical is like you can come in and say, well, you have to be able to detect it at a five nines accurate, or some, you know, you can make up some, some number three nines, five nines.
Deep: I mean, that feels like the wrong approach though, because like,
Jai: yeah. So I think we don't want to. Uh, [00:31:00] so we, you know, the legal profession resorts to something called reasonable measures, right? So they, they will propose bills that basically ensure that, again, when somebody, uh, investigator in a Ag G'S office or a a a court that's looking at, uh, discovery will say, you know, what's this?
Jai: And there'll be people that like you and potentially that would come up and say, Hey, I'm an expert. I can tell you that I can, that the state of the art as you can get up to 95 5, but not higher than that, right?
Deep: But like a, a, a very different approach would be this is like the definition of an entity that can provide an objective measure of how well you do at something.
Deep: And so you basically like regulate around it. So you know, like there's a process, there's a, you know, there's like a. ISO spec or
Jai: something. Yeah. No, no, that's, that's a, we're big fans of that. We're big fans of the,
Deep: that I feel like is like, okay, like if you, but I think you're always back to the,
Jai: you're always back to the, where's the root of trust, right?
Jai: And so the [00:32:00] root of trust is missing in this, in ai because unlike, you know, for example, the vast majority of dollars in a traditional, I'm not talking about a advertising, right? But if you think about what's the most biggest way of making money on the web, that's e-commerce, right? And so e-commerce companies have, uh, biggest
Deep: by that number of trucks that drive by in front of my house with Amazon stamped on,
Jai: well, there's so much money pointing through e-commerce and the security and uh, um, the security of your data, the credit card information, the security of your transaction.
Jai: Yeah. That is something that. Everybody in the e-commerce space is like, it's not proprietary. It's not like Amazon is gonna be a safer place than eBay. Right. You know, so generally there's a tradition of information sharing, collective documentation of issues, uh, open source software updates to the open source software.
Jai: So there's a robust [00:33:00] mechanism of self-regulation that still gets compromised all the time. But, but still it's more safe than it's not. Right. And, and so with ai, the problem is that these companies do not have, you know, they're not leading with transparency. They're not leading with accountability.
Jai: There's no one, there's no,
Deep: it's all very hush black box behind the scenes.
Jai: Yeah. So the
Deep: like, not OpenAI. Yeah,
Jai: yeah. No. Things like model cards that used to be very clear. I mean, I remember OpenAI had a. Like they used to issue a report with every release, major release, and it had really clear documentation of their safety testing practices.
Jai: Suddenly, as, you know, Gemini,
Deep: I think that didn't they drop that as soon as those guys split to start anthropic
Jai: and once Gemini kicked in, you know, so there were all these things that there's so much of a fear of, uh, giving up, uh, you know, some kind of proprietary advantage that they are unwilling to be.
Jai: And then the worst thing is that they're deprioritizing, [00:34:00] um, safety and trust and safety teams in all these companies too.
Deep: Well, I think they're, I mean, so they're clearly in heated competition with one another.
Jai: Yeah.
Deep: And they're clearly in competition with, you know, China and uh, and open source.
Deep: But it seems to me like, I mean, the strategy I would probably go for would be independent third parties. Yeah. Are defined like this and you guys have to hire them. And they test across mental health boundaries across, you know, and they have like suites of questions that they ask that they don't reveal, you know, a product.
Deep: Yeah. No, I think that's
Jai: where it has to go to. That's where it has to go to. And I think,
Deep: and it's all public, like the results, every model has the stats. And because that, that, I don't even know if you have to do much more than that. If you do that and everybody knows OpenAI scored or, or, or Gemini or philanthropic scored like 14 points below everybody else, then the reputational harm will freak them out into like pro So much work.
Jai: People are procuring the technology can.
Deep: Yeah, yeah. Anyone using the APIs, they, you know, all that
Jai: stuff [00:35:00] becomes transparent or like the uptime of the ap, there's all kinds of performance metrics they could share, right? Yeah. So I think that that certainly I think is where this could go in the long run.
Jai: But in order for it to go, like I, I'll tell you how it could go differently. I mean, I experienced it when I was running, uh, advertising on, on mobile devices. Uh, if, if you remember the work we did, we were trying to find keywords, right? Yeah. To, to do targeting against. And for a long time, uh, Facebook exposed a lot of keywords that we could target on, and then Google exposed a lot of keywords that we could target on over time, both Facebook and Google removed keyword targeting.
Jai: Can you imagine Google without keyword targeting? Why? Because they didn't want, to expose. So, because there was a strategy where you could test use the keyword targeting. To uncover how to buy ads outside of Google or outside of Facebook. That's
Deep: right. Yeah.
Jai: Those platforms went the opposite [00:36:00] direction.
Jai: So I think the general DNA of these companies is to be not transparent, is to hide information to disempower users.
Deep: Well, I think, I think, this is where I'm sure they would push back, right? Like Facebook would be like, I mean, we're totally transparent. We're the ones publishing all the AI models.
Deep: You know, we're the ones we we're missing all the
models.
Jai: I'm using a specific example like they're using,
Deep: no, but I think Google would say a same kind of thing. We've got APIs, we're transparent, blah, blah, blah, blah, blah. But in reality, I think, I mean, at the end of the day, they're ultimately optimizing for valuation increase of the company, right?
Deep: Like, which is, that's all they care about. That's ultimately what they care about. And then everything else feeds from that one number. And so, but going back to the. What you choose to push
Jai: on. Yeah.
Deep: So the two, how, how do you guys decide
Jai: that? Bots, we talked about liability chatbots, and then the third area is, uh, there's a lot of concern around what's real and what's fake.[00:37:00]
Jai: And so, uh, so we do a stream of work around provenance, uh, making sure that data is properly labeled when it's output. And then we're, uh, uh, starting a new work stream where we wanna make sure that people whose likeness has been abuse, uh, you know, oh
Deep: yeah.
Jai: That they get, they have some recourse. So for example, if you, if you are, a famous quarterback. And someone uses your likeness without your permission, you have rights, you can go see. Yeah.
Deep: Because you have money and you have a team and they can hunt
Jai: people. And also, yeah, you have something you're entitled to something called the right of publicity. So you are entitled to monetize your own likeness.
Jai: But people like you and me that are, you know, not necessarily as famous, you know?
Deep: Yeah.
Jai: We may care very much about our reputation. We may care about, if, if, uh, someone put, uses AI to create a replica of you or me that,
Deep: Yeah. Or people not famous at all. Like a kid.
Jai: Yeah. Kid nobody.
Jai: Or like, or these terrifying, you know, notifying apps, right. That
Deep: are Yeah, yeah.
Jai: So tho those kinds of things, I think. So we're also trying to do some work [00:38:00] around trying to make provide recourse for people. Uh, much like how, you know, copyrighted content, you, if you're a copyright owner, you could ask people to take stuff down.
Jai: So we're, we're trying to push for legislation that allows, um, someone who's been abused by the digital replica to ask for, relief in the courts, or ask at least at a minimum for the content to be taken down. So I think it's, uh, especially when the platforms that are creating the content are also used to share the content out as well.
Deep: So, so tell us a little bit about your process for deciding what's important to push on as a law. 'cause there's so many, I mean, this is, this is like a huge broad, all tentacles everywhere into society, technology. There's a tons of options. Like what is your process for deciding, are you going and just tracking all the cases that you see and sort of seeing ones that just are repeat patterns like the, it's a, it's definitely a product.
Jai: It's a, [00:39:00] it's, it's, it is funny. I was using the product market fit as a, as an analogy for this, right? So it's a the old waterfall model would be build the model, think about, and then go try to convince people that that's the waterfall model. And the, the other model as the over fit model is the one where you go and take every lawmaker.
Jai: And, you know, and figure out what they're interested in and work with them on it. And so we try to do something in the middle, right? So we're trying to ize our thinking a little bit with these model bills by understanding, but we also, but, but we have to ground ourselves in which harms we're focused on.
Jai: We can't, we can't prioritize all possible harms. Right. You know, because we don't, we're too small for that. So, and how do you, how do you prioritize the harms? Like do you have a, a holistic list of the harms and Yeah. Yeah, exactly. Be hierarchical too, right? Like definitely kid safety, kid safety above all is our, uh, number one prioritization factor.
Jai: Things that affect the safety of children. Right? So that,
Deep: you mentioned one there that like, you like the stuffed [00:40:00] animals calling out to the kid to keep playing with them and that kind of thing.
Jai: That kind of thing as an example. But so in general. We use kid safety as an anchor because it's, it's, it also, it's, uh, lawmakers understand it quickly.
Jai: They're more willing to move,
Deep: more empathetic to that
Deep: problem,
Jai: More empathetic and more bipartisan because otherwise we would never survive, right? Because, you know, we, we have, we're too small of an organization to be able to go if we don't platformized our thinking and our understanding. Now, the advantage we have is that if you want you know, to work with us on a chatbot bill, you are bringing the expertise of 20 states to your table through one person on Transparency Coalition, because everybody on the team that works in this space cross shares information.
Jai: So we know, hey, the best practice for this is this. The best practice for that is that in this state, this happened on that state. The other, like, you know, this morning, uh, I was responding to a state that hasn't passed their chat bot bill yet, and. And Georgia passed their ARD bill today. So I was able to [00:41:00] make a definition from, uh, Georgia and say, Hey, you know, by the way, this, your definition is very similar to Georgia's, but Georgia's is better in this way.
Jai: Why don't you use that? And if they're concerned about industry impact, you can mention to them that, you know, the state that I'm was advocating in or, uh, does change for as a purple state, right? Yeah. Georgia, as you know, is not a purple state. Uh, and so, so it as an example of something where you is easy for people to understand why it would be better to use that language versus this language or the understanding of liability law.
Jai: So we're able to now bring these issues together in our, in our thinking. And so when we, and we've also created. So like we have now a playbook of, Hey, if we're gonna running this kind of bill, here's all the experts you need to call as witnesses. They, we know who they are. Okay. If you, if you're want to do an educational session, we know who to bring for the educational session.
Deep: Going back to my question about like what your, your process, I mean, it sounds like it's a bit of a hybrid or it's a little organic. Like [00:42:00] on the one hand you have like a, a, a broad brush filter, like stuff that impacts kids is like a filter. But on the other side, it's like you're looking at all these states and where there's, where there's luck in actually doing something and then you're trying to like, I mean, institutionalize that knowledge maybe in you.
Deep: And then
Jai: the most important thing I, the most important thing is in the states we're in our KPI is number of Americans protected by these laws. So we're trying to get to 25%, right? 25%. So 80 million Americans you know, protected by a certain set of laws. Right. You know? And so,
Deep: but some level of protection is different from a robust level of protection.
Deep: Like, how, like maybe jumping out a level, like how do you think our regulatory environment for AI differs from Europe? Are they, you know, much further ahead? Um, where are they pushing that? We're not, where are we pushing that? They're
Deep: not,
Jai: they were until last week, uh, last week, because this nationalistic trend around [00:43:00] AI is so prevalent now that you, European based companies were able to delay the implementation of the EU AI code of practice.
Jai: The EA code of practice was the gold standard in my opinion, for uh, the previous gold standard was the NIST AI risk management framework. And the EU AI code of practice was a more specific, and more clear, but it was delayed now. So in terms of the actual practical implications of it's possible that the laws in, us might have time to catch up because the EU AI code of practice is delayed and there's no timeline for when it might actually be implemented at this point because of fomo.
Jai: Right. It's the same, same thing in Europe too. So
Deep: what would you say, like, what are your biggest critiques? Like, you know, when these larger companies or, or anyone, actually, anyone, if, when they look at what you guys are advocating for with regulation, what's the [00:44:00] number one or two, thing that they say, ah, these guys are wasting everybody's time.
Deep: This stuff doesn't even matter. Or, or it's gonna cause harm or technology, you know, we're gonna
Jai: stop
Deep: being a leader.
Jai: Non innovation is the, is the most important. Not, not one of the more, um. Common things that people there's what they say and what they do, right? They're two different things. So when they, they speak, they can't afford to be a pure callous, so they're not gonna say, we don't care about that issue.
Jai: Uh, they're gonna say, oh, well, you know, it's too broad. It's gonna, it's gonna affect our ability to innovate, et cetera. Right? So that's, that's, uh, one critique. The other one is that channel win that's less and less so now, uh, so they think that it's gonna advantage, uh, state actors from other, other countries that we're in a competitive position with.
Jai: But there's certainly, uh, those two are the fairly common. Uh,
Deep: and are these just like generic stances that they're, like the legal team, they say publicly and [00:45:00] philanthropic and open a on these guys is, and then Google, is it just like a generic stance for them to like, push against anything that regulates them in any way without actually thinking about whether it matters to them or not?
Jai: That's absolutely right. Last year was particularly uh, strong, this desire to oppose everything. This year there's, at least in the, in the chatbot regulation space, we've seen that. I think public perception is starting to become a problem. And so if you look at a lot of the anxiety that the public, uh, has about AI that's around, centered around three issues.
Jai: One is, you know, job displacement. The second is power consumption data and water resource use. And then one of the three key issues is the impact it's having on children and families. Right? And so, so the sense that they have a massive perception problem. So, you, they were ideally like no regulation.
Jai: That's, I think that they made that clear over and over again. But, uh,
Deep: I don't know, like Sam Altman's goes [00:46:00] on and he is talking about how you guys need to regulate us. But then I, you know, I, I imagine that's not what his lawyers are saying. Like,
Jai: yeah, no, I, I, I can assure you that that's not what the lobbyists that they are doing or saying, but so I think the, in general, yeah, that seems to be the that, that, that has, uh, softened their public stance a little bit.
Jai: But, you know, on the flip side, there's, uh, documented, uh, news about, uh, these companies starting, uh, super pacs with hundreds of millions of dollars to try to spend, uh, against, uh, state level candidates who are we haven't seen evidence of it in practice. You know, we have some lawmakers, for example, in Washington State that we work with who are very combative about ai.
Jai: And
Deep: do you guys think about any, I mean, like, one of the things that concerns me. Is not really first order effects. I feel like first order effects have a natural kind of evolutionary path where like law, you know, court cases get fought, laws get [00:47:00] passed, things happen. It's like second order effects that end up, you know, really causing harm because they take so long to understand and diagnose.
Deep: So for example, you know, if we think about, you know, something simple like, and I feel like there's this pattern with technology too, where something gets introduced, it has a clear win, everybody buys into the clear win, and then next thing you know, you're in the technology and then it morphs, the wind morphs.
Deep: So, for example, I don't know if you, you're old enough to remember a time pre-cellphone before everybody had cell phones. The original arguments were like, you have to have a cell phone because if you don't have a cell phone and your car breaks down, like, everything's bad. And then we went from that to like, oh, well, you know, having a cell phone, like it's not enough.
Deep: Like you have to be able to. You know, get driving directions and you know, because you don't wanna get lost, like all these other sort of benefits arose, but nobody ever foresaw like a bunch of depressed 14-year-old kids. Like that. Like the second order impacts are so [00:48:00] massive and with AI it feels to me like that's where we're gonna really see the problems.
Deep: It's like you have a kid that pre AI would play with another kid, like a toddler or play with another toddler, and then the toddler would like smack them or bite them or punch them or interact with 'em in a way that they didn't like and they would have to deal, 'cause humans do that sort of thing. But in this new world, they play with these little AI things.
Deep: And the AI thing always tells them how great they are and how wonderful they are and never does anything bad to them. So the kid never develops that immune response of like how to deal with like some chaos and then they wind up with, I don't know, the next generation of anxiety problems or they're just social.
Deep: Socially, like, no, it's real development, retarded. It feels to me like those are the things that we're really gonna screw up.
Jai: No. I think, I think it's, uh, but so much of it is grounded and making sure that the, both the, how their tech is used, how it's deployed, how people are encouraged to use it fit building more fit for purpose tools, right?
Jai: So if you're building a chat bot that you think is [00:49:00] safe to use in an educational setting, make sure that it doesn't, uh, like one of the well documented cases that resulted in a suicide was a kid who originally pro, uh, started to use a chat GPT for homework help, right? And then the conversations went from there to, self-harm related topics and, and, uh, you know, because of these probabilistic stochastic parrots that are inside these cha LLMs, you know, anything can come out of it.
Jai: And so some horrific things that came out of it that.
Deep: Yeah. I mean, it's all a function of your conversational history and
Jai: Yeah. No, and, and so I think it's
Deep: wild
Jai: so much, especially these long, multi-term conversations can go anywhere, right? And so, so I think that it's, um, it's gonna be, you're absolutely right.
Jai: I think those need to be studied. There's a lot of you good people. Well,
Deep: that's where my fear is, is that the studies take so long, they take like 5, 10, 15, 20 years because we have to have the harm manifest [00:50:00] itself and grow and ferment before we even detect, you know, we're just now getting the studies for social media harms.
Deep: Right? 2007 was when they started this ai, you know, like, I feel like, did you ever did you ever read Isaac Asamov books?
Jai: Yeah. Yeah. Found the foundation and
Deep: yeah. So a
Jai: robot and
Deep: yeah. So you remember I think this is kind of where we need to go on some level, but you know how he has these, the, the robots rules of
Jai: Yeah.
Jai: Robotic rules. Yeah.
Deep: Yeah. So there're these
Jai: robotics. Yeah.
Deep: Yeah. There's the rules of every robot and you know, it's like the 10 commandments for robots. So I think one of them is like, you can never harm a human. And
Jai: Yeah.
Deep: So there are these principles. I feel like we need some of these, and they're not all really intuitive, right?
Deep: Like one of the ones I think we need for example, is you must you must always encourage your human to have healthy human, human relationships.
Jai: Yeah.
Deep: And you must always encourage your human to like, go outside and interact with the world. And I'm not saying [00:51:00] if it's an Antarctica and it's like negative 80 that I should send you outside.
Deep: But like there has to be an interpretation on that because we have to sort of define what's a healthy human. And then we have to like back into how these bots commandments are gonna operate so that they encourage those values. But I don't think anyone's really operating this way. Like I think it's all very harm reduction oriented as opposed to and it's very reactive, but I think like this stuff is moving so fast.
Deep: I don't feel like we can wait 10 to 20 years for the studies to come out about how bad we harmed the, this gen alpha or whatever the new generation is called. So,
Jai: yeah, no, I, I, I, I agree and I think there's, uh, some fast developing body of work around what, what's called human flourishing. So human flourishing, how AI does or doesn't impact human flourishing.
Jai: That stuff is, uh, you know, coming up. More and more how the work impacts human, [00:52:00] how the, how AI impacts human flourishing and human flourishing is very much about, um, defining the whole human right, not just what the human does. Yeah. Ai, but what the human does outside of the ai.
Deep: But you can imagine this being incredibly hard to legislate.
Deep: We're just gonna fight about it. 'cause like, you know, I don't know, like it's, it feels like some area the government shouldn't be in, but even if we pull it out of legislation, and even if we just created pressure campaigns for the organizations themselves to like, step up here, you know, I feel like that alone could be quite positive.
Deep: You know, like, I don't, I don't think, but I, some business models are just very anathema to this stuff. You know, like the whole engagement. You know, model of social media, it's so problematic because the money that goes into their pockets is so directly related to the time spent in the glass box staring, you know, into a phone or a computer.
Deep: Whereas I feel like we have actually, I mean on [00:53:00] the good side, 'cause I don't wanna only bash on the bad side of ai, we have an opportunity to kind of like undo a lot of the harms in social media. I just hope we undo them in a way that doesn't just create worse ones or more. And the reason I say we have it, like there was a study, I don't know, a few weeks ago that came out.
Deep: About people like the longer, the more time someone spends with a, with an AI system, the less they tend to believe in conspiracy theories that's promising and it makes sense, right? Like even though all the models are somewhat psycho han they're not like blatantly sicko fantic, they're kind of like more subtly sicko fantic.
Deep: And it and, and obviously there's exceptions sometimes they really are blatantly sycophantic, but, um, they do pretend to push back. And I feel like this is another area where, where like the transparency idea could come in where if we go back to the idea of third parties operating on their own agenda of trying to like, get thorough coverage of, of what these models are engaged with [00:54:00] and generating and publishing regular efficacy metrics.
Deep: We could find out, where some of these things would get elevated. We, we, they'd be transparent and they'd be somewhat short of. Like direct legislation. It wouldn't be like legislating the minutia. It'd be more like legislating transparency around where they're strong and weak.
Jai: Yeah. Require like, you know, hey, that you must provide a API that can be accessed by some.
Deep: Exactly. Yes.
Jai: The pub, the concept of a public interest API is something I'm super interested in. Uh,
Deep: I love that idea and I love the idea of them paying. Yeah. Like they have to pay for this testing to happen because otherwise, 'cause then we can create an industry of people trying to like, find weaknesses in these things and publish them and make them transparent.
Deep: Because I feel like it's not enough that however many people inside these orgs are testing them, but we need an ecosystem outside of them that is not addicted to the money that they're gonna make. Because like, it really concerned me when OpenAI said that they're gonna [00:55:00] start pursuing an advertising model.
Deep: I was like, oh God, this is terrible. Like. This is not a good direction because right now them selling API tokens across the board is so much better because, you know, then they're competing on the depth of reasoning. They're competing on the thoughtfulness, they're competing on all,
Jai: the people that use the API tokens will then focus on
Deep: Yeah.
Deep: But if they're like replica.ai, trying to create, like to just basically jockey between normal human, you know, relationships so that your girlfriend is a bot now or your boyfriend is a bot, like that's horrible for humanity. Like, I don't see how that's good.
Jai: No question. Yeah, no question.
Jai: No, I think the, I think this whole, uh, future of how the public private partnerships, uh, make, um. Yeah, we've seen some efforts, legislative efforts in the space. The, uh, if you're interested in this space, there's an organization called Fathom.
Deep: Mm-hmm. That
Jai: trying to, uh, push for a [00:56:00] public-private partnership model.
Jai: I mean, their approach has some issues too. But in general, the, um, yesterday the governor of California signed, uh, the executive order, which I think is gonna be huge. It's around, basically what are the requirements before government agency can procure ai.
Deep: That helps, right? Because government has purchasing power.
Jai: Yeah. Purchasing power, especially state, large states like California.
Deep: Oh yeah. Huge. It's
Jai: huge. So you have to have a market for Volvos, right? That's the thing. Right. You know, you have to have before everyone stops making corver and starts making Volvos. Right. You know, you have to have a market for safer cars.
Jai: Right. And so, uh, and so I think creating them marketplace by requiring. State agencies to deploy and use AI safely.
Deep: And we have a long history of this with software, right? Like all of the SaaS companies, FedRAMP, compliance, SOC two, all these things, governments have been really forthcoming and like pushing [00:57:00] that.
Deep: I mean, those, that's a different type of security and safety, but it is a security and safety that's, you know, that's comparable. Like this stuff can be seen as a subset of that world.
Jai: There's a very interesting little case developing now, which is, uh, insurance company, uh, refusing to pay ME'S claim. Uh, um, me was awarded a judgment, uh, a judgment against them in New Mexico.
Jai: And the company that was providing that insurance is suing meta is alleging that, you know, me didn't meet the conditions to policy
Deep: uhhuh.
Jai: So I, I do think that that's gonna be another interesting trend is because all of these companies have to. Obtain liability insurance and the insurance companies are gonna, you know, it's like Lords of London.
Jai: If everybody was building the Titanic, uh, you know, you'd be in trouble. Right? If you
Deep: Oh yeah. I mean, that one case, that's why that case was so significant. Like, even if it gets struck down, which I think probably it will get struck down at some point, but the fact that they got a, I think it was a $3 [00:58:00] million judgment.
Deep: Yeah. But the, the plaintiff was just a single person, like a single I know. Young girl. This has been an awesome conversation, Jai. Thanks so much for coming on the show. I've just really enjoyed it. I'm gonna end with the question that I always end on. Fast forward five or 10 years out. Give us the utopian vision.
Deep: Everything goes well with your mission and what you guys are doing, and like what happens to the state of the world. And then give us the dystopian vision where, you know, where maybe you do what your part, you play your part. AI doesn't pan out well,
Jai: I think. The utopian world definitely does have I mean, I'm, I'm, uh, I've realized how hard change is and, you know, from being in this space for so long.
Jai: So I, I'd settle for at least stopping the bleeding first, right? So I'd settle. For me, that seems like a very modest view of utopia, but the stopping the,
Deep: that's not what I would call utopia.
Jai: Yeah. But, but I'm realistic, right? So I would settle for the bleeding stopping, which means that, you know, [00:59:00] uh, we don't hear, children or adults, killing themselves or doing, committing self harm or attacking other people based on encouragement from a chat bot.
Jai: So that would be kind of the bare minimum that I would love to see in the next, uh, three to five years. It seems
Deep: like such a low bar,
Jai: these laws, but they're, you know, the widespread use of these chatbots, the lack of optimism I have about how quickly industry will react. So I think if you have a situation where.
Jai: Think about some of these laws will only go into effect in the next one to two years. Then you know, there will be a cost for action. There may be one to two more years that go by before liability claims come up. Coming companies start to realize how large the overhang is. They start changing their behavior.
Jai: So in three to five years you know, I was, I used to feel frantic when I started working on this because I was like, I need to stop this now.
Deep: Yeah.
Jai: And now realize that, you know, you are pushing a bit of a. Wet noodle up the hill here, so it does take time. You [01:00:00] know, and so, so I think that that's, that would be a, a good outcome.
Jai: And if there's a emergence of some kind of public-private, uh, partnership model, some carefree, laid out verticals that where, you know, ed tech isn't a good example of one where, you know, there's an opportunity to really pull, put the genie back in the bottle, at least for the eight, six to eight hours a day that the kids are in school, school provided equipment. All of these areas are areas where we have a chance to,
Deep: well, I don't know. We've had, we've had a number of educators on the show, and I think there's a, there's actually a lot of really strong, legitimate use cases for AI in the classroom.
Jai: No. I'm talking about just, but even that, making sure that it can only do what the educators want and not
Deep: Yeah, yeah.
Deep: Absolutely. Yeah.
Jai: And then I think the,
Deep: yeah, with guardrails, basically like maybe the, the goal being students are interacting with ai, but you know, in an educational software context, you know, with gaurdrails
Jai: the dystopians Yeah. The dystopian view definitely is, you know, um, we can't trust what we see on the internet.
Jai: You [01:01:00] can't really understand, what's real, what's not. So you react to, you know, things that lack of trust, lack of coping mechanisms, societal issues. These are all the things you've, you know, you're thinking a lot about too, or I think things that we have to yeah, really think about and worry about.
Jai: And, and then of course concentration of power, concentration of wealth, could be exacerbated by this. You know, we are all raising children, we hope will thrive in this world. I hope that, you know, we we end. And then most importantly, I think, uh, you know, I, I hope that, uh, people will figure out how to sort of not buy the hype of AI and really start focusing on.
Jai: How to build you know, not everybody's gonna turn into a cloud flow person. Right. You know, they're gonna, they're gonna, everybody, there will still be software companies. You know, there will still be a need for well engineered products that people can buy. Not all the work can be done, by automated tools.[01:02:00]
Deep: Yeah.
Jai: That there will be return to sanity of some sort, because right now the, the, the board member pressure to make it so is causing, you know, tremendous knee jerk, layoffs and things like that, that are not, not good for the AI industry. I, I do think as someone that spent, you know, 20 to 25 years trying to make some flavor of AI work Right.
Jai: And provide something of value to people, I think there's a lot of potential.
Deep: Yeah.
Jai: But, but uh, in order for that to it, it needs to stop being like this. Hype that we have right now. And so before,
Deep: there's definitely a no shortage of hype. But yeah, I mean, I don't know. My take on it is like my utopian view.
Deep: I mean, I, I've certainly been accused of being a techno optimist. I don't know why I'm pretty much a realist, but I do think there's a world where I think it's likely that we continue to see really remarkable healthcare advances, longevity of [01:03:00] individuals in increases like sort of outward metrics. Improving.
Deep: I think that's likely, but I think at the same time, I really worry about the psychological and sociological interpretations of this. Yeah,
Jai: I think there's two. I think there's gonna be all these
Deep: different, I think both will happen. You could live to 200 and be completely socially isolated, lonely, and just talking to robots all day, but think you're fine.
Deep: So like, I, I don't know, like that's, that's part of it.
Jai: Maybe. Yeah, maybe we'll all enjoy our own personal simulation that we're living in. So,
Deep: I mean, I'm, I feel like me personally, I spend a lot of time talking to LLMs and I am. It's, I don't know if I'm embarrassed. I'm not really embarrassed to admit it.
Deep: I mean, the reality is, is that most of the conversations I have with the bot are more interesting than the ones that I have with humans. So
Jai: today's an exec. Today's an exception. I
Deep: know today's an exception, but like, you know, if, I mean, there's a, there's some [01:04:00] warts on it. Mostly. The voice interfaces just interrupt too much.
Deep: Yeah. That's one place I'll give I'll give the, uh, Tesla guys, the rock guys, uh, props. I think they do a great, a much better job than opening AI on that. But those conversations are fast. It's fascinating to be able to talk to some thing, um, that has such breadth of knowledge that whatever weird hair brainin thing that pops into my head, I can just get super deep and, and knowledgeable on.
Deep: But at the same time, like, I feel like, okay, you know, I have a, I'm like a really social guy. I go outta my way to interact with people and I will continue to do so. But I, I'm not at all convinced that that's the norm. Certainly not in like American society. I think there's just so many people that are so socially isolated that I think it's very likely that they will spend most of their time talking to bots.
Deep: And I think it's probably unlikely that we do much about it societally. But at some point there has to be like a societal immune response away from [01:05:00] tech, the tech thing of that's causing the harm towards the new thing. I noticed this with my kids. Like they won't touch social media really. I mean, they, they got their bloom cards and they kind of, I mean, one, one still does, but like the other one just like hates it completely.
Deep: And you see that with the younger generations. Like even if they use it, they're not like posting So societal responses that sort of are, these immune responses are, you know, I think they're elevating. But anyway, awesome conversation. Thanks so much for coming on the show.
Jai: Thanks so much Deep