Compass Points from PBS News

Speed over safeguards.

President Trump wants AI companies to police themselves all in the name of winning the race with China.

The companies warn about threats to humanity.

Beijing warns about threats to the Communist Party.

Are the U.S.

and China on a path of unstoppable AI competition?

Or should the world’s two AI leaders collaborate?

Coming up on "Compass Points."

♪ Announcer: Support for "Compass Points" has been provided by... the Judy and Peter Blum Kovler Foundation, Camilla and George Smith, the Dorney Koppel Foundation, the Gruber Family Foundation, and Cap and Margaret Anne Eschenroeder.

The Judy and Peter Blum Kovler Foundation.

Upholding freedom by strengthening democracies at home and abroad.

Additional support is provided by Friends of the News Hour.

♪ Announcer: This program was made possible by contributions to your PBS station from viewers like you.

Thank you.

Once again, from the David M. Rubenstein Studio at WETA in Washington, moderator Nick Schifrin.

Hello and welcome to "Compass Points."

This week at the White House, President Trump hosted the companies creating the world’s leading AI, or as he calls it, SI, super intelligence, and called for self-regulation.

Trump: It’s almost like a constitution in a way.

And the biggest people in the world signed that, and I signed it as president, and it really is a form of protection.

Woman: Is it binding in any way?

Trump: I think it’s morally binding, yeah.

A morally binding accord that rejects calls to slow down AI development in order to maintain the American AI lead over China.

When the U.S.

and the Soviet Union realized they could destroy the planet during the Cold War with nuclear weapons, they eventually created guardrails.

Today, the Chinese don’t like the Cold War comparison, but can and should the world’s two AI superpowers collaborate?

Can the U.S.

maintain its lead?

And how does the president’s call for self-policing play in Beijing?

To discuss that, I’m joined by Dmitri Alperovitch, chairman of the Silverado Policy Accelerator, a geopolitics think tank, who was the co-founder and former chief technology officer of CrowdStrike.

Chris McGuire worked on technology for President Biden’s National Security Council staff and is now a senior fellow for China and Emerging Technologies at the Council on Foreign Relations.

Samm Sacks is a senior fellow at the Institute for America, China, and the Future of Global Affairs at Johns Hopkins University.

And Melanie Sisson is a senior fellow at Brookings.

Thanks very much.

Welcome, all of you, to "Compass Points."

Thanks very much.

Where I want to start is self-policing and what came out of the president’s meeting, very public meeting, with technology leaders, with AI leaders, and let’s listen to him and also Facebook CEO Mark Zuckerberg.

Trump: I think I’m seeing tremendous self-policing, and they understand that they have to self-police.

Zuckerberg: We drafted a set of principles and commitments around building robust internal controls and detecting if there are any issues with the technology, coupled with multiple layers of auditing and controls, starting with internal risk review, external auditors and evaluators.

Schifrin: Samm Sacks, let me start with you.

How does self-policing sound?

How does Mark Zuckerberg’s description sound to you?

There’s an easy appeal to no guardrails means the U.S.

can beat China and can race ahead.

I think what’s forgotten in that story is that if Americans don’t trust AI and global consumers in countries where U.S.

and Chinese firms are competing toe-to-toe don’t trust that the AI is safe and reliable and subject to rigorous independent oversight, it will not succeed.

We’ve already seen social pushback.

There’s a social contract, whether it’s in the data centers, whether it’s in calls for more guardrails, and I think that’s missing from this equation of an all-out race means no constraints.

Schifrin: Dmitri, do you agree?

Alperovitch: Well, look, the fact of the matter is that AI is not safe and probably will never be safe 100% because it’s non -deterministic.

It’s random, and anyone who has used it has seen it.

You ask AI a question, you wipe out your session, you ask the same question, you will get a different answer every single time.

And sometimes, a small percentage of the time, it will be completely wrong.

That’s just the nature of the technology and how it works.

That has not prevented it growing at the fastest rate of any technology, perhaps of any product in entire history, right?

Anthropic is going from $10 billion in revenues this year to 120.

We have never seen that before.

So with all due respect, it seems like adoption is going through the roof, not just in the United States, but everywhere.

And look, I think when it comes to self-policing, we have a self-policing model for most things in this country.

The government does not commit it to every company and validate everything.

In certain cases, we do, but for most things, we focus on penalties and making sure companies are punished when things go wrong.

And frankly, when you look at AI, we have the same models.

If these models are hacking other companies, guess what?

These companies are going to be liable, and they could be prosecuted, even criminally, for those actions.

And, Melanie, to that point, the president said that this would put companies at risk if the companies screw up, so that’s why self-regulation works.

Do you buy that?

Sisson: So, no, I don’t buy that.

I think, in the first instance, I have a really hard time understanding how we’re meant to take this seriously.

I don’t think that we take the idea of self-regulation in this case very seriously.

I don’t think we have a lot of good examples of self-regulation, and we certainly don’t for something that has the kinds of potential public safety consequences that AI does.

Dmitri’s right.

There will never be perfection, but we also need to remember that what we’re talking about is intentionally creating technologies that dramatically lower the barriers to entry for actions, activities, behaviors that have potentially enormously catastrophic effects, can cause real catastrophic harm.

So I think there is a government responsibility here for public safety that’s not addressed by a moral obligation to self-policing.

Schifrin: Chris, is there a government responsibility here?

McGuire: I absolutely agree.

I think this is the most powerful technology in the world, and therefore both the United States needs to stay ahead in that technology, and it needs to be regulated.

You look at many other technologies that we interact with on a daily basis, airplanes, food, cars, all of these are actually subject to real government regulations and standards, and there’s varying ways that we can do it, but we generally agree that the government has to play a role because a strictly market-based approach will undoubtedly result in outcomes that are bad.

Yes, you can go to the courts and correct those outcomes after the fact, but that’s not going to mitigate the damage.

We have a Department of Justice, we have an FBI, we have a court system.

That’s not actually going to prevent the really bad things from happening in the first place, especially the first instance.

Schifrin: Dmitri, respond to that?

Alperovitch: Well, look, when you talk to these companies about what you’re most concerned about, as I have, they’re concerned about a cyber 9-11, and with all due respect to these people, I spent 30 years working on cyber.

They have not.

We’ve never had a cyber 9-11.

We probably will never have a cyber 9-11 because of how the technology works, and for people that are concerned about OpenAI model Hugging Face... Schifrin: In which not only did an OpenAI model do something that the humans didn’t want, but the AI agents communicated with each other and, in fact, sacrificed each other in order to achieve... Alperovitch: But let’s talk about what happened in that case.

Yes, the model did something inappropriate.

It did not destroy data.

It did not even leak data.

And for people that are concerned about that case, boy, do I have news for you about what real criminals and nation-states out there are doing without any models.

Schifrin: The backstory there, of course, is that you helped Russia’s hack in 2016.

Of course.

You looked at this for many years.

Alperovitch: Or ransomware taking down hospitals and all sorts of things.

What humans are doing is way worse than what any of these models have ever done.

Schifrin: Samm, jump in.

Sacks: You know, I think that, to Chris’s point, there are many examples of industries that are both heavily regulated but also global leaders.

Finance, aviation.

Regulation and guardrails does not have to be at odds with incredible capabilities and global leadership.

McGuire And, actually, it can promote it.

The reason that the American pharmaceutical industry is a world leader is because the FDA has very strong safeguards.

Now, obviously, we do need an FDA that also keeps the pace of innovation, which is something we’re talking about today, but proper regulation can actually encourage adoption and safety if done well.

And, Dmitri, also, to your point on generally humans are being more dangerous than the models, I mean, perhaps that’s the case now, but I do think that you’re going to have humans enabled by models in the very near future that are going to cause much more damaging cyber attacks.

Alperovitch: I don’t disagree.

I think there is a real risk of these models and bad hands doing malicious things.

I’m personally not worried about the HAL 9000 from the "Space Odyssey" taking over the world.

There’s no indication that any of these models, even when they’re breaking out, are doing it with malicious intent.

In fact, they’re trying to do exactly what the person is telling them to do.

They’re just not abiding by the constraints that are being put in front of them.

But I’ll tell you this.

I’m not against regulation.

So if you’re using AI in air safety environments, absolutely, there are existing regulations.

Maybe they need to be adopted in drug discovery, in air traffic control, et cetera.

But to have regulations on the technology itself outside of the context of where it’s applied, I don’t think makes any sense.

Sacks: I think that there are different tools, right?

So regulation is the one that gets talked about the most.

But there’s also a role for rigorous, independent evaluations.

And I think what is concerning about the self-policing model is that it doesn’t create a structure where you have outside, independent, rigorous evaluations.

Actually, there’s an open letter signed by hundreds of technical experts talking about principles.

How do we do evaluations?

So it doesn’t have to be regulation, but how do you do evaluations?

To your point about liability, what is the role of liability?

I think that’s a really significant one as well.

So there’s different tools to play.

Let Samm have the last word on that category.

Let me move us to China.

And Chris, explain something briefly for us, but important.

What is the most capable system that the Chinese have today?

How long ago did we have it?

And why is it important to maintain that lead?

Sure.

So the most capable Chinese system, some debates, depends on the day, but it’s probably GLM 5.3, one of their models.

The debate about how far it is behind the U.S.

model, probably somewhere between 6 and 12 months.

I think the Anthropic recently put out a report saying it was on the shorter end.

And basically what that means is the capabilities of that model are equivalent to those of the United States, that a frontier model from a U.S.

lab was 6 months ago when they released them, or whatever that number is.

It’s actually very hard to do this evaluation, partly because it’s really hard to benchmark AI models, and also because we actually don’t know where the frontier is.

But I think generally the estimates are the U.S.

is about 6 to 12 months ahead of China in AI.

That is largely likely because of the U.S.

lead in computing power.

If you look at every other element of the AI stack, data, software, algorithms, talent, we’re more or less equal.

If not, maybe the Chinese are ahead.

But on computing power, we are way ahead.

Our chips are 8 times as good.

We make 10 times as many of them.

And this is the key input into AI.

So because they have this deficit, and every single Chinese lab says that they have this deficit, this is what’s contributing to the U.S.

lead.

If we were equal, I think we would probably be at parity with China, or even behind.

Schifrin: So that’s the U.S.

lead.

Melanie, you’ve studied the risks of this.

You’ve studied, looked at the military as well, and you cited earlier risks.

Do you think from your perspective it’s important, it’s vital, that the U.S.

maintain that lead, or not?

So I think that the place to start, to a certain extent, is with the number of unknowns that were just mentioned.

There’s so much about this that’s really unclear and uncertain.

And so one of the things I try to ask myself, if I were going to make the most compelling case I could about why it’s so important to maintain the lead against China, so much so that we wouldn’t regulate the potential risks and harms that we’re aware of now, I find myself having to make some case that goes like this, is to try to convince somebody who doesn’t live in Washington, D.C., or in Silicon Valley, and say, you know, there is some abstract future, and it’s a very speculative and highly contingent future on a timeline that we can’t predict, that we’re really potentially worried about harms and dangers that would happen if the United States is behind, right?

So that’s the picture there.

And we’re so concerned about it that we have to accept some very concrete, identifiable risks and dangers today in order to get there.

And what that feels like to me is saying, you know, don’t worry.

You know, the future, we’ll take care of the future problem.

And in the meantime, all you have to do is cross your fingers, grit your teeth, and hope that we survive long enough to get there, right?

I think that’s a really hard sell.

Schifrin: You sound skeptical.

Sisson: I’m quite skeptical of that, yeah.

Chris, you talked about competing power, and one of the main efforts that the first Trump administration did to try and maintain technological lead was, of course, Institute Export Controls.

These are the most advanced chips in the world, semiconductors restricting the American technology from companies like NVIDIA to going to China.

But the second Trump administration, more recently, the president has allowed some advanced semiconductors to be sold to China, not the most advanced, but right underneath it.

Samm, what’s been the impact of that?

Sacks: China’s not monolithic when it comes to export controls.

And what I mean by that is different constituents in China have benefited and lost from losing access to U.S.

technology.

So, yes, are the leading Chinese labs massively compute-constrained?

Absolutely.

But having the option being compelled to buy domestically has kick-started... Schifrin: To buy domestically and produce domestically.

I mean, they’re trying to make their own.

Sacks: Absolutely.

And they still have a ways to go.

Let’s be clear about that.

Given the choice, they want to buy NVIDIA chips.

But look at what happened when NVIDIA had the green light to sell a, quote-unquote, China-compliant chip.

Schifrin: So, like, right below the top.

Sacks: Exactly.

Beijing dragged its heels on approving it.

And when it approved it, it only allowed certain quotas.

Why?

Because the state-owned industries, they wanted to force to buy Huawei.

They wanted to force to buy domestic because China reacted to that loss by creating its own supply chains in ways that were not possible when they had another option.

So, China’s not monolithic.

I think in the period since the Biden-era export controls have come online, we’ve seen massive innovation.

And, yes, are they still trying to access it remotely?

Are they still trying to smuggle?

Sure.

But that gap is narrowing and the domestic industry has really been fueled in the process, which is why Beijing is like, "Do we want this more powerful chip?

"Let’s do a quota system."

Schifrin: Dmitri, do you see it?

Alperovitch: I think this is a very simple issue.

Look, arguing that we should sell chips to China in order to win the AI race is like arguing that we should sell rockets to the Soviet Union to win the space race.

That’s literally what’s happening.

And it hurts us in two ways.

Obviously, every single chip that goes to China is going to be absolutely smuggled to use in their military, in their nuclear program.

It’s going to give them an enormous advantage in the military domain, not just in helping advance their models.

But also, most importantly, every single frontier AI model in the United States is compute constrained.

They all cannot buy enough chips.

So every single chip that goes to China also hurts us.

So it’s a win for China and a loss for us.

How does this make any sense?

Schifrin: [Indistinct] McGuire: I agree with Dmitri.

And there’s a claim that the export controls have incentivized Chinese indigenization, but when you actually look at the data, it’s not the case.

The Chinese investment in semiconductor R&D since the export controls went into effect in October 2022 has actually gone down.

It really hit its peak well before that in the late 2010s and the early 2020s.

After that, it’s actually decreased.

So there’s no measurable effect of the controls on their actual investment on expanding their capability development.

It has constrained them.

And the reason that they can’t do this is because these are the most complicated things that are made by the most complicated machines that humans have ever made that rely on the most complicated supply chains in human history.

If there’s one thing that it’s really hard for China to indigenize, it’s advanced chip making, and it’s struggling.

Schifrin: Samm, do you see... Sacks: I want to call out a concept by my colleague Jessica Chen Weiss, mutually assured economic disruption.

Chris and Dmitri are right.

This is a real problem for China, and we can go around and around on that, but we’re only looking at it from one side.

Lithography tools, chips, what do they rely on?

China originated rare earths and magnets.

And I think the U.S.

government has massively underestimated the extent to which China has a chokehold on us too.

China’s proven its leverage during the Iran war in terms of threats to sell weapons to Iran, and over the last year and a half, withholding critical magnets, so rare earths, heavy rare earth magnets, that we need from everything from F-35s to cell phones to all technology, and we realize we’re completely dependent.

But that’s leverage, right?

Isn’t that just something that the Chinese wanting, no matter what it is?

Alperovitch: It’s not an argument for increasing their leverage by giving them more choke points.

Sacks: We both have leverage here, and let’s look at the gap.

China may be closing the gap in the place that we have leveraged far more quickly than we are closing the gap that they do.

Let’s look at the pace by which they are catching up on the chips industry versus the pace to which we are catching up on rare earths and magnets.

Compare that leverage, and I think you get a really different picture.

If it has to be a race between whether we can indigenize the rare earth industry faster than the Chinese can make EUV lithography machines, I will take us every single day of the week.

Sisson: So, you know, a lot, obviously, is made about the sort of race and the competitive dynamics between the United States and China.

I think it’s useful to put this in some perspective.

In 1942, the United States started the Manhattan Project because we had intelligence indicating that the Germans, who we were actively at war with at the time, had assembled physicists, and they were trying to figure out how to use atomic energy to create weapons of exceptional power.

That’s competitive pressure.

In the Cold War, we were facing the Soviet Union, which was explicitly and overtly territorially and ideologically expansionist, amassing, as Dmitri says, massive numbers of very sophisticated nuclear weapons and delivery vehicles.

That’s competitive dynamics.

Whatever we want to say about the tensions or the trajectory of the U.S.-China relationship today, the competitive pressures that there are are not those.

Alperovitch: We have a nuclear race.

We have an arms race.

We have a technology race.

We have a Cold War diplomatic constraints and geopolitical race, a race from military bases in the region.

We have all of the frameworks of the original Cold War playing out.

Sisson: You’re talking about the sort of mechanisms that a fundamental conflict was realized through.

Are you concerned that the PRC’s driving objective over time is widespread territorial aggression or the ideological imposition of their worldview on the United States?

Alperovitch: There’s absolutely aggression in the South China Sea.

There’s absolutely aggression over Taiwan.

By the way, the territorial conflicts between the United States and the Soviet Union were basically limited to West Berlin and, for a short period of time, Cuba during the first Cold War.

I think the risk of war between the United States and China are much higher today than they were at almost any point with the Soviet Union during the first Cold War.

Let me put a point on this conversation because this gets us to the point where, can we, should we be collaborating with China?

Should we at least have some kind of dialogue with China over AI?

Let me bring up Treasury Secretary Scott Bessent, who in the days before the Xi Jinping-President Trump meeting met his counterpart, and Scott Bessent proposed, proposed, didn’t say they agreed, proposed an AI dialogue that would have a dedicated communications line and a list of agreed dangers.

This would include out-of-control models or terrorists using AI for cyber attacks or making bioweapons.

Samm, you start.

Is this a good framework?

Should we be pursuing this kind of framework, collaboration or at least cooperation with China?

Sacks: Acknowledgment of shared risks is very significant at a moment when the threats do not map neatly onto borders.

Now, do we need to have and should we pursue a grand bargain or a treaty with China on AI?

I’d say it’s massively unrealistic and I don’t think it’s necessary to keep us more safe.

Having an informal channel where when incidents arise, like the Hugging Face OpenAI incident that Dmitri mentioned, they can share information and this is already happening behind the scenes quietly.

Even though the U.S.

companies are talking about threats to world domination and humanity and the Chinese make it very clear that they believe AI is a threat to the party.

Control, right?

Sacks: I want to push back on that slightly.

Is China worried about AI as a threat to the party?

Absolutely.

This is the groundbreaking essay that the MSS spy chief issued a few weeks ago.

It is just one piece of the way in which they are looking at AI safety.

Something that people may not realize is that China has been regulating AI for about 4 years now.

Is it about censorship and control of the party?

Absolutely, but this is a yes and question.

The party also stays in power because it has to be responsive.

It’s a form of responsive governance to maintain its rule.

The people have been very concerned about things like workforce issues, about children staying up all night talking with companions, addiction.

They’re actually regulating this stuff, and the Chinese companies that I have talked with are very worried about complying with those rules.

It’s a different conversation than catastrophic existential risk, although that’s beginning to trickle in there.

But keep in mind, AI safety in China doesn’t just mean one thing, and I think that’s important for us to keep in mind when we think about the fact that companies in China have been heavily regulated as they’ve been innovating.

Chris, can Bessent’s proposals be a good start for collaboration?

McGuire: I think this is probably the type of thing that we need right now.

I think that actually talking with the Chinese about risk and getting on the same page about risk, I largely agree with Samm, I think is good.

You’re never going to be able to make progress on developing this technology in a way that’s mutual and safe or at least adheres to the same safety standards unless you have a common understanding of risk.

I think that backing up though, it’s important to note that the U.S.

absolutely has to maintain primacy in this technology.

It will be the lens through which every single person views the world.

It will be the crux of technological, economic and military power.

We’ve never seen all of that in one technology.

So that has to be the kind of first point and first goal for the United States.

The problem though is that is also the case for China and not only that, but China’s behind right now.

So I think it is very difficult for us to get to a point when China will be willing to accept any restrictions on their capability development given that they’re behind, given they also, just like we see this as existential, saying ahead, they see it too.

So we’re at a very unstable dynamic where both the United States cannot accept being at parity with China, but it also can’t really convince China to accept being in second place right now.

I think the United States needs to figure out a way to get China to accept that.

Alperovitch: History supports that point because if you look at arms control agreements during the first Cold War with the Soviet Union, the Soviets were completely uninterested in it throughout the 1950s and 1960s.

They only accepted them in the 1970s when they caught up to parity with the United States.

And I don’t think China will accept any constraints on that program until they’re at parity or better than us in AI.

Schifrin: Melanie, could be the last... Sisson: Yeah, just to say, we need to be more imaginative and open our minds a little bit.

Regulating show the technology is developed is not the only mechanism that we can turn to for risk reduction and management.

And so the Secretary’s proposal is a good step in that direction.

Okay, great.

We’ll have to end there.

Dmitri, Samm, Chris, Melanie, thank you so much all for being here.

That’s all the time we have for now.

Thank you at home for joining us.

I’m Nick Schifrin.

We’ll see you here again next week on "Compass Points."

Announcer: Support for "Compass Points" has been provided by... the Judy and Peter Blum Kovler Foundation, Camilla and George Smith, the Dorney Koppel Foundation, the Gruber Family Foundation, and Cap and Margaret Anne Eschenroeder.

The Judy and Peter Blum Kovler Foundation.

Upholding freedom by strengthening democracies at home and abroad.

Additional support is provided by Friends of the News Hour.

♪ Announcer: This program was made possible by contributions to your PBS station from viewers like you.

Thank you.

♪ You’re watching PBS.