Series 4 Episode 3: What’s Happening with AI Policy and Regulation?

Transcripts are auto-generated and may contain errors.

0:00

Hey, I’m Rose.

And I’m Angie.

And this is Our Lives with Bots, the show where we ask important, timely questions about what it means to live with our bot counterparts.

And from time to time, we also dive deep into what an AI future might look like for us.

0:15

Sometimes we agree, sometimes we spiral, but we always go deep.

To think that in 2-3 years we’ve launched in the world billions of interactions, so we just don’t really understand.

But now we’re learning more and more about who’s being harmed and how do we protect against that.

0:33

How does this actually play out in the courtroom?

Is the burden of responsibility actually happening?

Trump has appointed people like Mark Zuckerberg, Jensen Huang, Larry Ellison as White House AI advisors.

The Trump administration also just released a seven part national Policy Framework for Artificial intelligence.

0:53

You are seeing across the world a similar trend of a deregulatory agenda backed by saying we want you to do this, and if you don’t do this, we’re going to you pay for it.

The Chinese government has been quite proactive trying to regulate the consumer harms that may come out of AI systems.

1:12

When AI systems are created in one country and used in another, where does legal jurisdiction fall?

There’s some very interesting problems here that have predated Chat box.

All those issues are just live.

Plus it’s now affecting a lot of people.

Welcome lovely viewers to episode 3 of our Series 4 on AI policy and AI regulation with Our Lives with Bots.

1:37

And today, we’re very excited to bring on a guest.

And who is that going to be today, Rose?

We’re bringing on Mihir Kshirsagar from Princeton University’s Center for Information Technology Policy, who has a very in depth background in terms of this regulation space from different sectors.

1:55

But we’ll let me here talk about that himself and bring him on with no further ado.

Welcome me here to Our Lives with Bots.

Thank you so much for being with us today.

Thank you.

Yeah, pleasure to be here.

So just to get started, could you provide a little bit of background about yourself and your expertise?

2:11

Sure.

Yeah.

So I joined Princeton’s Center for Information Technology Policy.

Now, it’s hard to imagine, but seven years ago, I came to Princeton from the New York attorney General’s office, where I was there in the Bureau of Internet and Technology.

And I was leading some cases around consumer protection, especially around broadband.

2:31

And I thought one of the, you know, there was privacy and other issues on the horizon.

And I felt like academia had an ability to look ahead at the problems and help enforcers in thinking through problems more carefully.

And so that’s what I came here to Princeton to do.

2:47

And it’s been great.

Before the AG’s office, I also spent a long time in private practice at 2 leading New York firms.

And before law school, I spent some time as a civil liberties advocate with the Electronic Privacy Information Center.

So I have a broad perspective on different angles of policy, both from an advocacy point, from a defense point, and from a prosecution point, and now from an academic point.

3:11

So I bring all those sort of different perspectives to the problems and trying to help people address the challenges that they face.

Well, that’s certainly perfect for a podcast episode today.

We’ll be talking a lot about recent developments in AI policy, but adjacently in social media and talk about whether or not that sets precedent for AI policies, focusing on things that have been emerging on the federal level in the United States.

3:37

And then also talking a little bit about how it compares to more broadly global policy, for example, in India and China.

So just to get us started, could you tell us a little bit about what’s the state of open cases against big tech regarding AI chat bots in the US, the UK, across the world?

3:55

And how have these cases played out so far?

Yeah.

So I’ll talk a little bit about the US cases 1st.

And they’ve been a series of lawsuits that private plaintiffs have brought against the major platforms for various issues related to mental health.

4:13

They’re currently at the stage at which cases have been filed, but there really has been no action taken about them that is meaningful and not some of them have one, one of the cases I guess, settled one of the character, I think the character dot AI case, one of them did settle.

But there are other cases that are still ongoing.

4:30

And there’s growing interest in bringing more cases like that.

And so you’re seeing spectrum of cases and the people who are bringing this have first started out bringing cases primarily with children, but they’ve now brought more and more cases which involve adults.

And some of the early cases were sort of not very, very clear risks of actually materialized with respect to harm.

4:53

But now you’re seeing some more cases that are dealing with the more complex issues of emotional dependencies.

And so you’re seeing also a change in the pattern of the kinds of cases people are.

Bringing and speaking more specifically in terms of things that just recently happened, at the end of March this year, New Mexico and California found Meta and Google with YouTube liable for knowingly harming children’s mental health through its social media platforms.

5:17

And I guess the question is, does this set any sort of viable precedent for their LLM products?

Yeah.

Oh, for sure.

I mean, I think the New Mexico decision, that was the attorney general who brought the action.

And I think in that setting, holding a company responsible for how it interacts with the general public and the mental health harms that it may cause is definitely a precedent that all the LLM litigations are going to look to and learn from.

5:45

In California, that’s a jury trial.

There was a bellwether case, there was a case that was a trial case and they also found liability there.

And I think that’s also going to be something it’s, it’s a little less clear how well that will translate into other areas.

6:01

But I think that it is for sure something that plaintiffs are going to look at and defense is going to look like and say, oh, there’s really a potential case here.

And I think maybe for a second, I’ll talk about sort of what is different about these cases is that historically with social media, anytime you had a problem with what the social media company did, there was Section 230 of the 1st Amendment that said that media companies shielded from any effect of what third parties did on that platform.

6:31

And what the cases in New Mexico and California did was to attack the question not from what was the content of the communication, but how the product was designed.

And with LLMS, which don’t get protected under Section 230 anyway, since they’re the output of a company, the question of thinking about how they are designed to create certain kinds of responses or elicit certain kinds of behaviours, the people interacting with them.

7:00

The fact that you can now find liability based on doing that in the social media case is going to be very instructive in the LLM case.

Thank you, Miha.

I mean, I think it’s very interesting to see, as you say, what happens in case law or what happens in the court in terms of setting direction for precedent.

7:17

And almost the zeitgeist is in terms of appetite for where liability is going to fall.

Maybe what would be interesting to hear your opinions on is also the zeitgeist from a policy perspective.

And if we just look at some of the things that are happening at a federal level at the moment, I believe at the end of March, Trump has appointed people like Mark Zuckerberg, Jensen Huang, Larry Ellison and a number of others fairly influential in their own rights in the private sector as White House AI advisors.

7:46

Keen to hear from you how you foresee the influence of this sort of advisory panel playing out in terms of setting the trajectory for AI policy.

In the past, there have been these great and good corporate leaders who’ve been appointed to these kind of advisory bodies and they don’t do a whole lot.

8:02

And I suspect that these people will not do a whole lot as well.

That said, I think there’s a very explicit direction from the Trump administration, which is, to no surprise, quite comfortable with having Big Deck heavily involved in setting the policy and the policy being a deregulatory policy.

8:23

But what we’ve seen, especially in, you know, respect to stuff you’ve talked about in in your podcast, you know, with respect to mental health issues, and that’s there’s an explicit carve out for that even in the White House legislative strategy.

And then across a lot of conservative states, there’s this very, very strong interest in protecting the constituents from from the harms.

8:46

And so there’s a place where even the deregulatory agenda is willing to acknowledge that where there are known harms, we’ve got to help address them.

And so I suspect that there is going to be this ongoing tension within the Trump administration about what is in scope or not in scope for what the deregulatory agenda should take on.

9:08

But with respect to the emotional dependencies, the emotional health, I think that is clearly something they have to do more on and they understand they have to do more on or at minimum, let the states keep doing whatever the states are doing.

And so I suspect we’re we’re going to see a lot more action from the states in this area.

9:27

So we’ll definitely be touching on the state versus federal jurisdiction and in terms of policies being generated.

But just to focus once again on the federal level, the Trump administration also just recently at the end of March, released a seven part national policy framework for artificial intelligence.

9:46

And so could you give us a sense of what’s in this document and what does it mean for responsible AI regulation specifically?

Yeah, it’s, it’s a pretty short document, which usually these, these, these proposals tend to be very lengthy.

And, you know, the set of bullet points that largely confirm what we’ve understood to be for a while the Trump administration’s approach to these questions, which is, you know #1 which actually comes up again here.

10:12

Is that protecting children, safeguarding communities that’s sort of front and center.

But then, you know, they want to pre empt state laws which interfere with the business operations of various firms.

And so I think that’s it’s this combination of the deregulatory agenda on the one hand and then trying to do some kind of consumer protection work on the other.

10:35

And they’ve also dabbled a little bit into intellectual property and some of these other things, which again, these are high level aspirational statements.

The question will be what specific set of legislative proposals are they going to get behind and then are they actually going to take the effort to push through certain kinds of solutions?

10:58

We’ve seen in Congress that this really, you know, very little action on things which there’s been widespread agreement there should be action on.

I think AI may well fall into the same realm as privacy in some of these other areas where there’s lots of talk and very little action.

11:14

And would the Trump administration take some of its political capital and put it into taking affirmative steps?

I think that’s one open question.

At the end of last year, Trump signed that executive order to create an AI litigation task force.

11:29

Can you say more about what the provisions were of this executive order and what does that exactly mean for AI regulation in the US?

And how does that kind of separate out what the jurisdiction is for federal versus state policy?

So let me back up for a second.

Maybe it’ll be helpful just to explain sort of at at a very, very high level, how state and federal powers allocated in the US, typically states have authority over anything to do with safety of their constituents, education, safety.

11:59

Those are areas that have typically been state policy areas.

The federal government has typically taken on things that relate to Interstate commerce.

So anything that is related to how businesses operate across different areas, those have been relegated to federal power.

12:15

So the big question here, which is a fight that’s been going on across various areas with digital technology long predating this, is what can federal laws say?

The only thing that is going to regulate here is what is set at the federal level.

12:32

And at the state level you do not have the ability to protect against certain kinds of harms because now we’re protecting it at the federal level.

And so that’s this concept of pre emption, which says that the federal law eclipses and removes the ability for states to act.

12:49

Sometimes in pre emption, the states can act, but they have to act within the box that the federal law provides them.

But let’s say with respect to something like unfair and deceptive trade practices in protecting consumers from that.

States have complete authority in their own territory to do what they want to do.

13:07

And then the federal government also has its own authority to do it.

So the executive order, which is just an executive order, right?

It’s not a law.

It’s just it’s really a sense from the Trump administration of what it believes doesn’t really tell a state it can or cannot do certain things as a matter of law, but it does have an effect on some decisions of states on the margin.

13:33

So if a state wants to pass a law, let’s say that outside of child protection, let’s take it, you know, on on certain kinds of AI regulation, let’s say it has certain mandatory disclosures, there’s a risk that the Trump administration would say, oh, now you’re interfering with our national policy.

13:51

If you’re interfering with the national policy, then as an executive branch decision, we’re going to make a choice to maybe cut some of this money that we’re giving you somewhere else.

So they have that coercive possibility of dying, some action in the state to some funding choice because that is in the executive branch’s control.

14:11

And so you are seeing some states being a little shy about what it is they’re up to or reformulating what it is they’re doing in order to avoid the Trump administration coming them down hard on them.

And so you’re seeing this even in other areas, you know, with respect to, say, diversity guidelines and things like that.

14:30

States are reformulating what they’re doing to avoid the Trump administration coming after them.

So while to some degree the action is toothless, the executive order, because these detection issues have historically been the province of a state, you are seeing some states saying, well, you know, if it does get us into trouble, maybe we should pull back a little bit.

14:54

Maybe we won’t be as aggressive, maybe there’s some risk to trying some of these actions.

So that makes sense to me in terms of the government can withhold funding, it can create limitations that might be punitive enough, that might intimidate or be some sort of a reason for the states not to act.

15:11

But does the executive order provide for any sort of funding or any sort of resource that would then ensure that the federal government or that the government would have provisions to create policy that would create this safety net then in terms of responsible AI regulation, if that makes sense?

15:28

The legislative proposal is is to do that at a federal level.

So I don’t think there’s much from the executive branch.

I mean, Trump revoked the Biden administration’s executive order with respect to AI usage within the federal government.

They went back to their prior one, which actually, and I’ve thought this to the federal government employees about what the different policies are.

15:50

The Biden administration’s executive AI policy is actually not that far away from what the Trump administration, the Trump one had in terms of what kinds of things they’d safeguards they needed.

And so, you know, a lot of this, the devil’s in the detail, but I think so, you know, because a lot of these requirements are things about ensuring that the systems work correctly.

16:12

And if you have an interest in having these systems work correctly, you’re going to require certain kinds of safeguards.

So you see some natural overlap between the different administrations because you need to find tools that work effectively to meet the mission that you have.

And so the big fight here with the executive order is how much can the Trump administration decide how AI companies interact with the general public?

16:37

And for that, they need to come up with federal laws that will do that, would give that guidance.

And that’s this legislative agenda.

So it seems like on the federal level, there’s a little bit of a trend of toothless policies where they don’t really hold any legal weight, but they do have this kind of normative social weight.

16:57

You know, something that’s punitive or appears as though the states, for example, might get lashed out at.

So given the supposed influence of U.S. policy across the globe, do you anticipate that other nations might follow suit with attempts to block or circumvent AI regulation?

17:15

And why or why not?

Yeah, it’s a good question.

I mean, I think my general sense is that with Europe passing the EUAI Act, that was then quickly followed by now more of a deregulatory agenda that has also taken hold in Europe.

17:32

Part of it is based on, you know, they, they had the former Prime Minister of Italy write a report on sort of the cost of some of their regulations in the digital economy, whether that was holding Europe back.

So there’s some internal pressures as well that within Europe have looked at and said, look, why aren’t we attracting the same kind of industry and innovation within Europe?

17:53

Is it our regulatory structure that’s causing the issue or is it something else?

But I think you’re then also seeing the US aggressively export this deregulatory agenda onto Europe and saying, you know, we want you to allow our companies to run free.

18:09

And, you know, otherwise we’ll do our tariffs, we’ll do some other things that the US has done to export other policies that it believes in.

So you are seeing, I think across the world, the similar trend, if you will, right, of a deregulatory agenda backed by, you know, not particular tools, but backed by this coercive financial penalty based approach saying, you know, we want you to do this and if you don’t do this, we’re going to make you pay for it.

18:39

And so that trend, I think we both the Trump administration has taken respect to domestic policies, but has also done that internationally.

And then depending on the country you’re in, you’re more or less open to that kind of persuasion, right?

So I don’t think China cares, but you know, Europe is stuck a little bit in the middle, whether you’re Brazil or some other country, there’s a whole host of other kind of considerations that you have to go through.

19:05

But I think it’s going to depend on on the political economy of, of each country that is examining how much to embrace this deregulatory agenda.

And, and I should say there are obviously reasons internally, a lot of these countries, and we can get into this, right, are concerned because they see this incredible technology developing where it’s mostly occurring in the US and they’re worried about digital sovereignty.

19:29

They’re worried about how they can exercise control over it.

And so you’re seeing countries look at that question and say, OK, today we’re very beholden on how the Trump administration and the US, not the, you know, the view this is the US Gates access to this technology, whether it’s to the chips, whether it’s to the models.

19:49

How much do we see a future where we are now dependent on these infrastructures?

Why don’t we develop our own way out?

As you brought up China, there are people that would refer to both India and China as being front runners in AI regulation.

20:05

I would love to hear from you in terms of how you would compare India and China to the US in terms of their approaches and maybe actually would you consider them front runners and responsible AI regulations specifically?

It’s a good question on on the responsible question, right?

20:21

Responsible is always respect to what set of values and the Chinese government has a very clear set of values that it wants to see in their products.

And so you will see that particulated in various ways, right?

So you will get certain kinds of responses.

If you wanted to find out certain kinds of, if you wanted to know what happened in Tiananmen Square, for example, you will find a different set of responses based on how that model was trained and where it comes out of.

20:46

But putting that question of sort of responsibility to one side in how that’s defined.

The Chinese government has been quite proactive in trying to regulate the consumer harms that may come out of various AI systems.

So they have algorithmic transparency requirements.

21:04

They actually have pretty strong privacy protections.

Again, this is respect to how commercial actors interact with the general public.

It is not with respect to how the state as a whole can control its citizens, but it’s trying to think about how third party commercial actors can take advantage or not of people.

21:24

And on those kinds of questions, China has actually quite a robust set of policies that are forward-looking and so on.

And so you’re seeing some of that also in India is trying to do that.

And so both those countries have have some interesting templates to to see how is it that you’re trying to regulate these different spaces.

21:44

So depending on where you are, I suppose responsible has a different definition or accuracy.

So thinking still about the global space of AI regulation, we’re curious, for example, how jurisdiction plays out when large language models or AI systems are developed in one country, which mainly is the US at this stage, and then used in another by, you know, users or governments with varying policies in each of these locations.

22:14

And so when AI systems or products are created in one country and used in another, where does legal jurisdiction fall?

How does that process play out?

You know, the way it plays out and you can actually think about this even one step prior if you think about social media companies as well, right?

22:30

The way it plays out is to say if you operate in our country, then you’ve got to abide by the set of laws and regulations we have here.

And so if you want to be X in Europe, then you have to satisfy all these requirements.

22:46

And so if you want to provide your chatbot in Europe, then here these set of requirements you have to abide by.

The challenge comes when the company says, yeah, OK, but we’re not going to comply, or it goes to the Trump administration and says, you know, if we’re going to be kicked out of this country, well, that’s going to have some, you know, you’re going to have some tariff blowback.

23:08

And so you better find a way to accommodate these people.

And so then the question comes back to this, how much can you actually regulate when you have somebody in your country with respect to AI models?

It gets a little bit more complicated because you could imagine the AI model being exported to another jurisdiction.

23:27

If it’s a, let’s say it’s an open weight model and then it’s used and applied in that other country, then there’s very little you can do to regulate in that space.

You know, China developed DeepSeek, right?

But it’s being used widely in the US and in various places because it can be hosted in the US and you can change some of the weights that respond to queries differently, and you can actually change the ways in which some of the responses may come out.

23:53

But I mean, building on that, I mean taking it one step further, what happens to things like storage of user conversations and other privacy relevant information?

And I’m just thinking in my context, I know a little bit more about GDPR and you know, Europe based policies.

So GDPR would probably enforce certain sort of requirements in terms of how data is stored, privacy is protected.

24:15

But then, as you say, if it’s replica and it originates from the US, then how am I protected when I should be protected by GPR if I’m using it in, let’s say, in, you know, Europe?

I mean Replica and the other companies should comply with GPR.

24:32

So they should the data if it goes, you know, if, if replica data is being created that is not compliant with GDPR, but is being offered to people in Europe, that would be a regulatory risk that that replica would be running.

So they would have to comply.

24:48

They would have to comply.

It’s it’s a data service, right?

So, so that part is, is the same, right, Whether whatever is at the front end of it, whether it’s a chat part or whether it’s some other way of feature.

If you’re using personal data, you have to comply with the laws of that jurisdiction and it might mean that you have to store the data within your jurisdiction and so on.

25:09

With respect to personal data, where it gets, where it gets complicated and difficult with respect to GDPR, as I’m sure you know, is just what if the data that’s being now sent back to the US is anonymized in some way or it’s being used to train and improve the model back in the US.

25:30

But it doesn’t really involve the export of this personal data.

But a lot of the interactions then actually the weights are changed and, and the training changes as the result of all these interactions in Europe.

And then it they used to improve back the model in the US.

25:47

Is there anything that the European regulations can say or do about that?

And to my knowledge, I don’t know how much the European regulatory structure has thought through those kinds of questions.

You commented on this briefly or mention it briefly.

The difference between, for example, open source models versus Frontier closed source models.

26:06

So closed source models being things like open AIS, ChatGPT and clawed, right?

You can’t get into those systems and look at how they’re weighted.

You know they’re for profit and so they’re keeping those under wraps.

But then you have open source models like DeepSeek in China, Mistral in France and Quen 3, which is backed by Alibaba.

26:24

And so is there regulation or talk about policies for open source versus these frontier closed source models?

And we’re partly asking this because we’ve had subscribers asking these questions and we really don’t know the answer to them.

Yeah.

Let me talk about this with respect to the US, right, where it’s particularly interesting because they’re trying to think about how to regulate frontier models and there’s a very hard push from the existing frontier model companies who have access to the model as gated through an API, through some kind of gateway.

26:56

So that’s the closed model.

The other model, I know you used open source.

A lot of people use open source as a shorthand.

It’s actually a lot of those models are not open source models.

They’re more these open weight models, which are different open sources where you understand how the training data is, you understand how exactly the system is put together.

27:18

Open weights just gives you a set of statistical tables that then generate the answer.

So you can apply an open weight model, you wouldn’t necessarily know what it’s trained on and how those weights were arrived at.

And so all the models that we’ve talked about and that are publicly known, whether it’s done DeepSeek or Queno, these others are open weight models.

27:38

And so so is Meadows Llama is an open weight model.

It’s not an open source model, it’s a detail, but it matters because it’s sort of they try to borrow in the open source communities like, oh, you can think of any part of this and change it and you can understand it end to end.

27:54

And it’s not you actually just get a set of numbers that then you can apply and use and adapt if you want to change the weights, but it’s not something that you can fully understand how it was trained and developed.

So with that in mind, there has been an effort in the US where the frontier models that are closed are trying to block the open weight ones because they view themselves as, as only the ones that are the most protective could be behind these closed APIs.

28:22

We can see who’s coming and what kind of queries there are.

And, and I think that’s, that’s a mistake.

Certainly from the US perspective.

I think these open weight models have a lot of to offer a lot of academics use it, a lot of other third parties.

I think there’s a role I’ve written about this in other places where there’s a big competition concern if you allow access to this frontier model technology to only be through these closed model providers does.

28:48

So I think there’s, there’s a tremendous value in having open weight models and frankly open source models as well.

And, and I think what’s interesting is that the Trump administration seems to have also adopted some of that perspective.

They’re not fully on board with Tony having the closed model as the way to deliver these services.

29:07

Just as a follow up though, like if we were to think about regulation and I’m imagining that there may be a continuum between a frontier model open weight and open source where to some extent it’s almost like the black boxes more opaque in the frontier model and maybe less opaque in an open source model.

29:24

As you say, you can see the training data, you know what how it’s been built, you know, the completely open kimono in the open weight.

Would it be fair to say that if we were thinking about regulating where the liability sits, that you might be talking that a frontier model or one of these sort of really closed models, they may have more liability because they’re completely in control of how that model behaves.

29:47

Whereas it may be less clear whose viable in an open source model.

Because to some extent, the user or the person that then gets that model and plays with it, the next developer in the line is able to tinker with it and to change its output and how it behaves.

30:04

So you’re exactly right, right.

So I think, you know, there’s a whole question of just how much should you hold a downstream model responsible for some action at the front end?

So you know, in the questions that are probably around some of the issues you’ve covered in your in your podcast, you know, if you’re interacting with replica and it was using Jaci PT 4 point O as the back end, is replica more responsible?

30:33

Is Jaci PT more responsible and how do you allocate that responsibility?

And the way it works in in other domains is that both are responsible and then you decide, you know, what’s the right allocation of a contributory responsibility for that particular action.

30:51

Here, I think the closer it is, I would argue the closer it is to the open source model where it is just an underlying product, right?

You know, you would not hold a chip responsible chip maker responsible for some output in the same way you may not hold kind of some underlying use of a model in some future action.

31:15

And the way it generally works, a thing about like a home that you have right and something goes wrong, you sue your general contractor, you sue the person that you’re interacting with.

And then it’s their responsibility to go back in the chain and figure out, well, how much responsibility is there?

31:32

Or you get some indemnification or something like that.

But what you’re trying to do is to protect the end user.

And I think that’s going to be where the where the struggles are and where for the regulatory agenda.

I think one of the things is that the frontier model companies have tried to take on the full stack.

31:50

So they they both provide the model, but they’re also providing the interfaces and they’re also providing some the enterprises.

And so that’s the way they’re trying to monetize the development costs of their product.

Well, in that case, you probably are going to see regulations that affect how they operate more than you would some of the others.

32:09

That’s interesting that both parties, both Replica and Open AI would be responsible in the case of Replica using Open AI’s model for their system.

And I guess that points to some, you know, I suppose weightiness for Open AI, considering that it offers its model to a lot of different third parties, including things like AI toys like Miko 3 Folo toy that Angie and I have talked about in a previous episode where those toys would be able to speak to children using Cha Chi BT and would engage in harmful manipulative tactics and talk about topics that are not to be talked about with children.

32:46

And so in that case, then Open AI is responsible for a whole host of user harm across all of these different applications.

So I suppose one of my questions there is how does this actually play out in the courtroom?

Is the burden of responsibility actually happening in this sense?

33:03

So this is the this is the challenge is that in theory open AI is responsible, but it may have indemnified itself with replica.

So if we could have told replica, hey, use it, but use it in your peril or if anything goes wrong, you’re responsible for, or if you’re your kid toy maker and it is able to say, you know, we, we disclaim any liability for anything that you do with our product.

33:31

And those disclaimers may or may not work.

And that might be tested in the court system.

You know, courts are sympathetic to finding the deepest pockets to if there is some real harm, they will find ways to ensure that there’s some redress.

33:48

So some of the boilerplate stuff doesn’t necessarily stall a court.

But in these kinds of situations, I think, you know, there is some critical questions that would have to go into, you know, how much is open AI understand about the business models of these third party people using their product as a wrapper.

34:04

And there’s some policy questions around how much do you want Open AI to decide whether or not provide those services, right?

You know, how much is it in competition with some of these products?

How much should it gatekeep around it?

If you look at similar parallels on app stores, on mobile phones, right?

34:23

What apps go in, what apps go out, how much gatekeeping is there?

It’s somewhat similar to what these wrappers are and what kind of wrappers you can provide over the app.

There’s some strong reasons I think to have Open AI at least disclose more about the products that are actually using their services.

34:42

So, so we have a better sense of what’s behind some of these applications like kid toys.

Well, it seems like the landscape in general is very, very complicated when it comes to AI policy and regulation and who is responsible for who’s harm.

And so I guess just to to wrap up our conversation today, given your background in all of these different sectors, how do you foresee the next several months or year in terms of the AI policy and regulation landscape playing out, especially given the recent court cases that we’ve seen?

35:11

What do people need to know or expect?

Well, so I think the mental health issues are just all going to be more and more salient.

So I think you’re just going to see more states try to pass regulations that are trying to protect consumers.

I mean this, you’ve covered this in your podcast beautifully in the past, but just to think that in 2-3 years we’ve launched in the world, you know, the millions and 1,000,000 billions of interactions.

35:37

So we just don’t really understand, we know are harmful.

We just don’t know who’s being harmed by it.

But now we’re learning more and more about who’s being harmed and how do we protect against that.

And so you’re seeing a massive real guard action and you’re seeing a lot of the companies understand that as well.

35:53

I would say that you are seeing open air take steps on protecting children through the Chachi PT interface, seeing some of the other actors trying to do more or less and there’s going to be some questions about how effective are those strategies.

Big picture, especially in the next year is you’re going to see a lot of announcements about what people want to see.

36:14

I don’t think you’re going to see much by way of Corrections, but I think that hopefully will set in place things that will start to then go back and be more useful.

And I think one of the key questions which I know we work on here is how do we measure how effective some of these strategies are and what are better ways to do things.

36:35

And I think there’s a lot of opportunity to do that in part because and going back to what I just said, now we’ve sort of unleashed this thing and then worried about what effect we’ll have.

Now we’re seeing the effects since we’re going back and saying, OK, how do we fix it.

And so this is going to be a long complicated process.

36:51

Just the sure effect it has and people is reason alone to spend time on it.

There’s some very interesting problems here that have predated chat box about how do you regulate these products that go across borders?

How do you regulate products where liabilities more diffuse?

37:07

How do you regulate products that are complicated to understand?

All those issues are just live.

Plus it’s now affecting a lot of people and so it is a difficult but but very interesting domain.

Certainly a slow moving monolith that is AI regulation and AI policy.

37:24

And we look forward to keeping up with your work and Princeton Center for Information Technology Policy, seeing whether or not these regulations are actually doing what they intend to do and how well they actually play out.

So thank you so much for me here today for being with us on the podcast.

And we hope to hear from you soon about emerging work in this space.

37:43

Yes, thank you very much.

My husband wonderful talking with you today.

Thank you.

Great to be here.