Fable 禁令撤销 + Dr. Dana Suskind 谈 AI 育儿 + 预测市场风波

So, Casey, we are back in the office after a vacation, and I was going through my mail this morning, which is usually just junk and—

Mm-hmm.

…books that people send me.

Letters from prisoners.

Letters from prisoners. But today I got a letter that was neither a book nor from a prisoner. It came from a 7th-grader who wrote to us. He said, “Dear Casey and Kevin, I am Nathan Herrick, and I’m a 7th-grader at St. Peter’s School in Cambridge.” He says some very nice things about the show.

Well, let’s hear them in full.

Okay. He says, “I wanted to inform you how helpful your podcast is for me. Before listening to your podcast, I always felt behind in tech. All my friends were talking about things in tech that I didn’t understand.”

Aw.

“It made me feel isolated and alone from them, but your podcast helped me rejoin my friend group with more confidence about technology.” He said, “Thank you for creating Hard Fork, or as my sister calls it, Hard Dork.”

Brutal.

So, Nathan—

And by the way, that is such a sister comment.

It is. Nathan, don’t listen to your sister.

Don’t. Do not listen to her.

But do listen to Hard Fork.

Absolutely.

We’re so glad to have you. I’ll just say it: If you’re a 7th-grader and you’re writing paper letters and sending them in the mail to podcasts, you’re going to make it in life.

The idea that you were so far behind in tech news that it was harming your relationship with other 12-year-olds—

What are they talking about?

What are these nuclear physicists he’s hanging out with on the playground?

They’re running agentic loops and—

Yeah, exactly. It’s a—

Doing their social studies homework.

…a 7th-grader with a Claude swarm is a—terrifying thing to consider.

I’m Kevin Roose, a tech columnist at The New York Times.

I’m Casey Newton from Platformer.

And this is Hard Fork.

This week, the Trump administration lifts its restrictions on Anthropic’s most powerful models. What have we learned about how the government wants to regulate AI? Then pediatric surgeon Dana Suskind stops by to discuss her new book on parenting in the age of AI. And finally, it’s our new segment about prediction markets, Against All Odds. And the odds that you’ll like it, they’re pretty high.

1. The AI Licensing Crisis

Well, we’ve had some big news out of Washington in the past few weeks. We are in the middle of something truly insane and unprecedented that is happening between the frontier AI companies, including OpenAI and Anthropic, and the federal government, which has decided all of a sudden, with no rhyme or reason, that it wants to start picking and choosing which customers are allowed to use the frontier models.

Yeah. When this happened, many listeners reached out to us and asked us if we would consider doing an emergency episode, and I’m happy to say that the emergency lasted so long that we’re able to get to this now on our regular schedule.

Yeah, it’s a rolling emergency.

Mm-hmm.

We’re in a permanent crisis, if you will.

Exactly.

So before we get into Fable, since it does involve AI, we should make our disclosures. I work for The New York Times, which is doing OpenAI, Microsoft, and Perplexity.

And my boy—

Oh.

Oh, ho.

Oh.

Downgrade. And my fiancé works for Anthropic.

Let’s explain what’s been happening, and then we can talk a little bit about what we think about it. The first salvo in this dispute was on June 12, which was the day that we officially left for our break—very good timing by the U.S. government—when the Commerce Department issued an export-control directive shutting down access to Fable 5 and Mythos 5, the most powerful versions of Anthropic’s models. Fable, in particular, had just been released a couple of days earlier, and the government basically said, “No one is allowed to use this if you’re a foreign national, including an employee of Anthropic inside or outside the U.S. You are not allowed to use this model.” That basically shut down access for all customers everywhere.

Anthropic basically had to pull the model because it had no way of doing user-by-user citizenship verification.

Right.

And as its reason for doing this, the federal government said, “Well, we received some notifications from a trusted partner that there was this jailbreak issue on Fable specifically.”

And who was this trusted partner, Kevin?

We believe it was Amazon. Based on some reporting from The Information, Andy Jassy, the CEO of Amazon, had gotten in touch with administration officials after researchers at Amazon had discovered this jailbreak—a way of getting supposedly restricted information that could be used in cyberattacks. Based on some reporting from The Wall Street Journal, we believe that Andy Jassy told the Commerce Department about this. They freaked out and said, “We’ve got to get this model off the market.”

Let’s zoom in as far as we can here, because I think it’s really important to get clear on what was the thing that this model did that gave the Trump administration the heebie-jeebies. My understanding, Kevin, is that when other security researchers reviewed what we believe the issue was here, it was a fairly standard back-and-forth and was the sort of interaction that you would expect a cyberdefender to have with a model to try to get it to identify and fix a bug. What do we know exactly about what was going on here?

We don’t know exactly what Amazon’s researchers discovered that made them so concerned. One cybersecurity expert who reviewed Amazon’s work said in a blog post that these concerns were basically not a big deal, that this was something that defenders who are trying to patch systems from cyberattacks need—basically asking AI to find and fix the bugs in a given file.

I guess Amazon’s researchers determined that that same technique could be used to exploit a piece of software or find vulnerabilities, which, of course, is the whole advertised purpose of giving Mythos and similar frontier models to cybersecurity defenders ahead of the general public.

Right. Anthropic had launched this program called Project Glasswing, where it had shared Mythos with a relatively small number of partners in an effort to harden the defenses of a lot of the critical infrastructure in this country and in allied countries before similar capabilities were found in other models. As soon as the government stepped in and said, effectively, “We’re placing export controls on this technology,” all that stopped.

Yeah. It’s now been a couple of weeks since these export controls went into effect. Last Friday, Anthropic and the Trump administration reached a tentative deal allowing Anthropic to restore access to Mythos for some clients. Then, on Tuesday night, news broke that the Commerce Department was lifting those export controls and restrictions on Anthropic’s models.

Commerce Secretary Howard Lutnick sent a letter to Anthropic saying that the company had taken steps to address the risks associated with these models. On Wednesday, Anthropic announced that Fable 5 is back, that it is again available to users around the world. Mythos 5 is still not publicly available, but is going to be restored to a set of U.S. organizations, and they will coordinate with the government to expand that more broadly. Casey, what did you make of the government’s reversal here?

The reversal seemed inevitable to me. It still concerns me how little was said about the process behind this. I was probably more interested in Anthropic’s response because they put out a blog post that effectively put the industry on blast and said that GPT-5.5, Kimi K2.7, and several other models could find the same vulnerabilities that had resulted in Fable being banned.

So, effectively, Anthropic said, “Hey, we added another safeguard, but this is really an industry thing. Don’t point fingers at us about this.”

Yes. And speaking of other models, let’s talk about what is going on with OpenAI and GPT-5.6. Last week, as OpenAI prepared to release its newest GPT-5.6 models, Sam Altman told the staff that they would not be rolling them out to the public right away. They would release them first to a limited group of partners that would be approved by the government, and the Trump administration had asked the company not to release the model more broadly than that.

He didn’t sound happy about it. This was clearly not the way that OpenAI and Sam Altman had wanted to release GPT-5.6, but the way that he phrased it made it seem like this was pressure that had been applied to him by the Trump administration. So this new business of the government telling AI companies, “You can’t release this model,” or, “You can only release it to a list of customers that we approve beforehand,” seems to be applied to at least 2 companies now, and this seems to be coalescing into something more like a broader strategy, not just a single targeted act of retribution.

Yes, and what we have said before on this subject is that this is the exact state of affairs that the people who are now running AI policy in the Trump administration were apoplectic about—

during the Biden administration, right? They were warning, “They’re going to create a licensing scheme, and they’re going to start picking winners, and only a handful of the administration’s favored labs are going to be able to release any models, and that’s why we’ve got to get rid of this Biden executive order.” Fast-forward to today: They have implemented a de facto licensing regime, and because it’s the Trump administration, there are no known rules. There’s no transparency whatsoever.

We have no idea what makes a model considered safe to ship or unsafe to ship. We’re truly just in this authoritarian limbo where, until a few people decide that Fable or GPT 5.6 are safe to use, we’re out of luck.

Yes. AI is now being regulated by vibes.

Yeah.

And that is—

Yeah.

—a state of affairs that I think both you and I were worried about, which is why, several years ago, when these kinds of pre-deployment testing regimes were proposed as part of these AI safety bills, we said, “Well, this is probably a good idea to have something like this in place so that you don’t just have this kind of slapdash, chaotic approach—

Yes.

—where models with really scary capabilities start coming out, and then the federal government has to scramble and try to figure out what the models are even capable of, and then take some fast action on that basis, rather than having something that is actually thought through and planned from the start.”

Yes, and let me say, I think the government should be able to prevent the release of large language models. I do believe that. I even think that after a model has been released, if it does something really dangerous, I think the government should be able to come in and say, “Hey, we actually need you to place a pause on that.”

I just think that it should be done with a set of clear rules. I think the companies should have due process. I think there should be a process for fixing those models, getting them into public shape, and getting some sort of consensus on when they can be released. Basic good-government stuff, and we are just so far from that world.

I want to return to the point that you made about the hypocrisy of the people, especially on the tech right—

Yeah.

—who made opposition to AI regulation and licensing their entire personalities—

Mm-hmm, mm-hmm.

—several years ago. Marc Andreessen has credited this, I think, nonexistent meeting. I have tried to back up, through reporting, the facts of this alleged meeting that Marc Andreessen says happened with officials from the Biden administration, where they basically said, according to Marc Andreessen, that they were essentially going to pick winners and losers in the AI industry, and that their favored labs would be able to develop and release powerful models, and that no one else, including open-source developers, would be able to.

Again, I have tried to back this up. I have tried to find people who would have been in a meeting like this. No one remembers it happening like this, so this may be a convenient fiction that Marc Andreessen has decided to tell. But this was their whole personality for years.

They said, “We are becoming a socialist, a Soviet-style state where the government is going to tell private companies what they’re allowed to do.” David Sacks, the former AI czar, famously said that the private sector should be allowed to cook. And now, just a couple of years later, we have what you call the de facto licensing regime, and they’re picking winners and losers.

They’re saying, “This company can release its models. This other company can’t.” They seem to have no clear criteria. They’re going on these reports from random companies that are testing these models, and it does not seem like they have any interest in formalizing that or making it any easier for AI companies to navigate.

Yeah, I mean, you say they’re picking winners and losers. Really, they’re just picking losers, right?

Right.

They’re saying, “This model’s too good. You can’t release it. Sorry, loser.” Right? So it’s unclear what the way to win here is.

I mean, the way to win is by becoming a political actor, right? That is sort of the downside here: Now deploying models, if you’re an AI company, is a political process.

Well, look, I would love to ask Greg Brockman over at OpenAI how he feels about his $25 million donation to the Trump people, right? Because that didn’t seem to buy them a lot of goodwill when it came to GPT 5.6. So I don’t know that sucking up to the Trump administration will even get your model out the door right now.

It won’t get it out the door right now, but I think they’re going to, if I had to predict, have an easier time over at OpenAI releasing these models in the future. I think the big shift that all the labs are trying to process is that they just moved from a default-yes environment, where the assumption was that if you trained a really good model, you could release it, and now I think they’re in a default-no environment.

If you build a model that is more capable than Claude Mythos or GPT 5.6, your assumption has to be that the government of the United States is not going to let you release it—

Yeah.

—at least not right away, and at least not to everyone.

Yeah.

So I think that is a huge shift in the structure of the AI industry that has happened more or less overnight.

Yeah, and this is just another moment where I wish that we had a Congress that was taking this seriously. This is a clear moment where we need a legislative framework for how these companies should be regulated that does govern how and when models can be released and governs how and when they are taken off the market over safety concerns, right? We know what to do here. We just have a government that is not doing it.

Yes. And Dean Ball, former Hard Fork guest—

Yeah.

—former Trump administration AI adviser, soon-to-be OpenAI employee, had a great post about all of this just a few days ago where he made the point, which I thought was reasonable, that the government should have some role. It’s reasonable to expect that in the future, when these models are very powerful and have very scary capabilities, the government will want to take an active role in deciding what can and can’t be sold. That seems reasonable.

Yeah, yeah, and I think there’s a good argument that that moment is now. The models are now capable enough and scary enough that the government has an interest in addressing it.

Yes.

Well, Kevin, I have some great news, because on Monday, none other than Marc Andreessen himself joined the U.S. Defense Policy Board—which will put him in regular conversation with Pete Hegseth and other government officials. And so now Marc can finally get to the bottom of this and demand that the government drop this de facto licensing regime.

Do you think he’ll be running any of the safety evaluations himself?

I’m terrified to find out.

2. China Tests The Frontier

So there’s one other twist to this story, and it’s something I’m really interested to get your take on, which is the whole China angle.

Mm-hmm.

Because the other backdrop of these discussions between the U.S. government and the U.S. AI labs is that these Chinese models are getting pretty good. So recently, we saw a new model out from the Chinese company ZAI called GLM 5.2. People are saying that this is one of the best open-source models out there.

There’s been a little bit of what I would consider overheated reporting that suggests that it’s as good as or better than some of the frontier models from the AI companies in the U.S. I don’t think that is probably true, but it is at least in the same tier of consideration.

And so there are people who are saying, “Well, this is going to be a big problem,” because at the same time the U.S. government is pulling American models off the market that are very powerful, we now have these Chinese open-source models that customers and big businesses can use for much cheaper, that are just as capable, and that are not in danger of being clawed back by the U.S. government. What do you make of this?

I think this is basically BS. I do. I think it feels like a lobbying tactic from people who want to get the American models back into production. I have seen no credible reporting that the Chinese models have kept up.

Even when I’ve read the reporting that we’ve seen over the past week out of The Wall Street Journal and others, there aren’t numbers in there that would lend credence to it. What there are is a handful of people saying these models are really good.

I’m sure these models are pretty capable. I’m sure they’re more capable than the open-source and Chinese models of a year ago. But we learned during the DeepSeek freak-out that a reliable way to get attention in AI discourse is to wave your hands and say, “The Chinese are catching up.”

The Chinese are catching up,” and everyone will pay attention to you. But when you look at how the Chinese are building these models, the theory just doesn’t hold, because you have to keep in mind what these Chinese companies are doing: They’re distilling American models.

Yes.

They’re getting in there, and they’re scraping the responses of these chatbots and using those to make their own chatbots. So the best Chinese models are kind of like American models once removed.

Yes.

And that is why they’re always going to be at least a little bit behind the frontier. And let me say one more thing, Kevin, while I’m all worked up.

Go off.

I’m increasingly coming to think about the whole A.I. market as 2 systems. There’s the frontier, and there’s everything else, okay? And if you’re truly at the frontier, we will know because you can see it in the benchmarks. You can see it in the revenue figures of the companies that are selling this stuff. That is where the actual value is.

And then there’s everything else, which is the million little open-source models, the distilled models, the Chinese models, and they’re just a step behind. And look at the revenues of those companies. There’s an insane amount of competition, and it’s not as good as the best. And so, in my view, these Chinese companies that we’re talking about right now are firmly in the everything-else camp.

That’s not to say they aren’t getting better over time, because they are. But if you’re going to tell me right now that this is as good as a Mythos, I’m going to need to see the damn numbers.

Right. I think that’s right, but I’m also not quite as skeptical of this fear of China catching up as you are, in part because I think these distillation tactics appear to work. You get a model that is almost as good as the original. If the Chinese companies are able to keep doing that, I think there’s a plausible case that they will shrink the gap between the American frontier and the open-source frontier. Maybe it’s 9 months now; maybe in the future that becomes 6 months or 3 months.

I think that sounds plausible to me. I also think that there are a lot of businesses right now that are really freaked out about what is happening with Fable and with GPT 5.6.

Sure.

And they are saying to themselves, “Well, if we were to go with the latest American models at all times, not only are we paying a premium for that, but it could be yanked away by the government with no explanation and no due process.”

Yeah.

And it would totally screw up all of our workflows and all the software that we’re building on top of those tools. So I did actually talk to someone recently who works at an American tech company—not one of the A.I. labs, but a solid, big American tech company—who was saying that they’re actually spending a lot more time with these open-source Chinese models, not because they’re as good as the frontier models, but because you can download them, you can run them on your own hardware, and they’re pretty good for a lot of tasks that don’t need the absolute capability frontier that you would get from a Mythos or a GPT 5.6.

So, in this moment where there’s so much uncertainty coming out of Washington, I think a lot of companies that rely on these tools are going to take a hard look at the Chinese models and see whether they might be good enough for some percentage of what they’re doing.

Yeah, that makes sense. The American economy has been so strong for so long, in part because the government has mostly been safe, boring, and predictable, and the Trump government is not those things, right? We saw this during tariffs, where American businesses were having similar problems because they couldn’t predict how expensive their products were going to be on any given day based on what the tariff of the day was, right? And now we have the L.L.M. restriction of the day, and it’s causing similar havoc in the economy.

So I think one obvious link between the actions of the U.S. government and the China subplot here is that maybe there’s a possibility that the U.S. government restricting access to these frontier models from American companies actually doesn’t allow China to catch up. Maybe it slows China down because, if they are so reliant on distilling from Claude and the GPT series and other leading American models, maybe not having access to those models will hurt them.

So is there any sense in which you think this action by the Trump administration could actually widen the gap between the U.S. capabilities frontier and the Chinese open-source frontier?

I mean, maybe, but that is an extremely ham-fisted tactic for accomplishing that goal. The United States government has many levers it can pull if it feels like Chinese companies are attacking its companies. For example, they could sanction companies that they believe have been caught distilling. They can prevent them from selling their wares in America. They can get allies to do the same thing.

We saw Anthropic the other day send a letter complaining that Alibaba has apparently been doing wide-scale distilling of Claude, and the hope there is that the government will intervene. So that is how these things are traditionally done.

While it is true, I suppose, that this is going to make it harder for Chinese companies to distill the next models, there are still all of these other reasons why, at least for me, that’s the worst way to go about it.

Right. That makes sense to me. I mean, the argument that I would make against that point is that what matters is not just the capabilities frontier of the internal models that the companies in each country have. What matters is the deployment frontier, and for a society to really harness the benefits of A.I., you have to be able to use the models.

You have to be able to deploy them inside big companies. You have to be able to deploy them inside consumer products. People have to be able to use them. It is not just which lab has in their basement the most capable model that nobody else can use.

Right, and all of this happened at a time when companies and government institutions around the world were using these very powerful models to improve their cybersecurity defenses. That’s one of the reasons why this issue feels really urgent: There was an ongoing project to protect the critical infrastructure of the United States and its allies, and that was yanked away, while at the same time we’re selling advanced chips to the Chinese to help them catch up.

I think there’s one silver lining to this story for me. For years now, you and I have been hoping against hope that the U.S. government would stop just talking about A.I. as this unalloyed good that was just a normal technology that was going to supercharge the economy, and that they would actually start to pay some attention to the risks of these powerful A.I. systems.

That is happening now, for better or for worse. The government is now firmly aware that these models do pose risks. They are thinking about cybersecurity, but I imagine they are also thinking about biorisk and other types of risk.

So I think there is a way to spin this whole saga as a clunky, ill-advised first step on a good path, which is that the government is waking up. These models are very powerful. They can be very dangerous. They’re going to get more powerful and more potentially dangerous as time goes on.

And so I think what we’re seeing now is a fumbling attempt to do something, and I think they will come up with better ways of doing this over time. This is not the permanent structure of the A.I. licensing regime in the United States. But at least they are not just pretending that this is a normal technology anymore.

That is true, and I’m glad that it’s the case. But at the same time, this same administration has been pushing to allow the export of more and more advanced chips to China, which will allow China to train more powerful models over time, and they’re doing that at the same time as they are preventing good models from being used by our allies like Britain and other countries.

So there is still an extreme level of incoherence in the administration’s position, in my opinion.

Well, you can’t win them all. That’s true. When we come back, a conversation about A.I. and parenting.

Do you know where your children are?

I have children?

Well, Casey, I’m very excited for our guest today, who is an expert on the intersection of parenting and A.I. This is a topic that I’ve been wanting to explore on the show for a while now, in part because it’s just become a big part of my social life: talking with other parents about how we are or aren’t exposing our kids to technology, including A.I. and chatbots.

So today we have on the show our guest, Dr. Dana Suskind. She’s a pediatric surgeon and childhood development expert. She’s a professor at the University of Chicago and the founder and co-director of the TMW Center for Early Learning and Public Health.

She’s also an author, and she has a book coming out soon called “Human Raised: Nurturing Connection, Curiosity, and Lifelong Learning in the Age of A.I.” And when I heard about this book, I thought, “This is just the person we need to tell me how we should or shouldn’t be using A.I. with our children.”

Absolutely, because in addition to all of that, Kevin, Dr. Suskind is a parent to 8 children, who are now all grown up. So this is somebody who has a lot of expertise in this subject.

Yes. So I think this is an important conversation today, but I think it's about to become much more important because there are just a lot of AI products and tools aimed at young kids that are making their way to the market and are going to become, I predict, more popular in the next few years.

So already, we've seen Miko and Luna, which are 2 AI-powered robots that are designed for kids, and we've got all these child-specific companions and chatbots with names like Aski, Elo, and Hey Otto. We're just starting to see a lot more products using AI that are aimed at children. And I think for a lot of parents, that is creating a lot of anxiety because how the heck are we supposed to know which of these things are going to help our kids grow and develop, and which are going to turn them into glorified slop cannons?

Yeah, and what you don't want is a slop cannon child.

Yeah. Heaven forbid.

But here's what I'll say: If your child is a slop cannon, you need to read this book.

It's true. Let's bring her in. Dr. Dana Suskind, welcome to Hard Fork.

Dana Suskind

Thanks so much for having me.

3. Human Connection Comes First

So you've got this new book coming out, Human Raised, which is about parenting and child development in the age of AI. I was so excited when I heard that this book was coming out because, as the parent of a now 4-year-old, I see this issue coming for me in a way that makes me quite nervous.

I was so happy when I heard that you were starting to do the research about what we, as parents, should think about our children potentially using AI, which I think is going to be a bigger issue for me as my kid gets a little older but is already an issue for a lot of parents I know. First of all, what made you want to pursue this topic as someone who's long been interested in childhood development?

Dana Suskind

Yeah. So I'm a cochlear implant surgeon, and you may say, “Well, why exactly are you on this or even thinking about this issue?” But early in my career, I started noticing profound differences in the outcomes of my own patients.

I would do the same surgery, use the same technology, and have the same parents loving them, and the outcomes of the children were so profoundly different, with some of them being able to develop language and others not. So that experience actually brought me into this incredible world of brain development and neuroscience that shows that human connection is not just a nice-to-have; it is the foundation of how we learn and how we learn to be human.

My whole career has been focused on that and supporting parents and caregivers in that important role. And then suddenly AI comes onto the scene. For the first time in human history, we have technology that can actually mimic that interaction that has traditionally wired our children's brains.

Now I'm like, “Oh, we need to probably step back and really think about what is happening in our children's world,” because the choices that we make at this moment are going to determine what our species looks like.

Dr. Suskind, can I run some scenarios by you from my own parenting journey, and you tell me whether I'm being a bad father or not? The 2 primary ways that I'm using AI with my 4-year-old right now: One is to answer random questions that he asks. We're in a big question phase right now.

The other day, he asked me, “Where does wind come from?” And I realized I actually don't know the answer.

Yeah, I have to say, that's a hard question.

Yeah.

Yeah.

So I pull up an AI model, and I ask, and I get back this answer, and it's about air pressure. And I'm like, “Okay, this is for a 4-year-old. Could you dumb it down a little bit?” And it gives me a great answer, and we tell him the answer, and he has the answer to his question. Now he gets to repeat it to all his friends. That's option 1.

By the way, on his playground, the kids are obsessed with where wind comes from. It's a hot topic over there.

It's made him quite popular at preschool.

Yeah, yeah.

So the other way is just creating stories for him. One of the things that these models are quite good at is you give it some guidance: “I want a story about a kid this age who has a best friend who's a dinosaur, and they go off and play soccer together,” or whatever it is.

I would do one where his best friend is the wind. I think that could really hit home with him right now. Are either of those examples of AI use in parenting a young child going to harm my son?

Absolutely not.

Okay. Phew.

So first of all, those are great examples. In this book, I tried to give grounded and practical frameworks. One of the frameworks that I give parents to think about AI in general is an evergreen. I call it HOPE.

H is human connection is irreplaceable. It is what we need to protect and double down on. O is owning your imperfections. As I said, kids don't grow through perfection. They grow through imperfection. I don't know about you, but I was a very imperfect parent. I am a very imperfect parent, but now I know it's really good for my kids.

P is protecting the early years. Those first years, when 85% of the physical brain is being built, we've got to be extra careful about what we let in. But lastly is E: If you're going to use it, use it to enhance, not replace.

You're really using this technology to enhance your relationship and fill in the gaps of knowledge that you may or may not have, and that's a great way. This book, in my view, is not at all anti-tech. I'm an implant surgeon. I build AI tech. I'm all about it. But how do we use it to deepen our relationships and support us in this important role?

Hmm.

Yeah, that makes a lot of sense to me, and it is compatible with my own vague optimism about AI in the lives of kids.

I see a lot of parents right now with a lot of anxiety about AI that seems like it is borrowed from this earlier era of the screen-time debates or the debates about addictive social media platforms for kids. And to me, it just feels fundamentally different: what AI is as a technology and what it is capable of doing for children.

I just think back to my own childhood. I was a nerdy kid who loved to learn stuff, and if I had had a chatbot that could have taught me about the Bernoulli effect in a way that was enjoyable and entertaining to me, I would've done that, and I would've been happier than going to my World Book Encyclopedia and looking up that fact.

So I think for the curious kids out there, my sense is—and please tell me if you agree with this—that this could be an incredibly enriching technology for them if it is designed well and if parents are in the loop about how their kids are using it.

I mean, 100% I agree with you. People go to social media as the best analogy, and I get it. But I think a better analogy is actually processed food.

Processed food is a continuum, right? There's whole-wheat bread, which is processed, but it's still nourishing, and then you've got the ultraprocessed variety, the Doritos and Twinkies, which are good occasionally. But the truth is that, in the same way, AI is a huge spectrum.

Not only can it provide answers as a chatbot, but it can lift invisible labor off parents and teachers, all the way to diagnostics. I come from the world of medicine, where it's transformative. I think people are really focused on the AI companions, which I get, and which I am concerned about in terms of crowding out that necessary human connection, but that's a small part of what AI is.

You can have socially assistive robots that help children with autism learn to read social cues to better connect with other humans. That's an amazing example. So I think, to your point, it is a spectrum, and it's not a monolith.

4. Children Need AI Guardrails

I'm curious, though: Where, if anywhere, are you drawing the bright lines? Where are you saying, “Do this, don't do that”?

Yeah. That's pretty easy right now. We are at the very beginning of this tech revolution, and AI companions and AI toys that claim to be better alternatives to screen time feel like a pretty hard no.

I think the thing that we can learn from social media is that we need to take a precautionary principle. You've got to show me it's safe before I'm going to let it into the developmental sanctuaries of our kids, because the stakes are just way too high.

Even higher than social media, in my opinion.

Mm-hmm.

What about smart devices that are not meant to be plush toys but, like, my kid loves talking to Alexa and asking for various children's songs that are intolerable to adults but that he loves to play on repeat?

Is that the same thing, or is that different because it's not pretending to be a companion or a stuffed animal?

I think it's how much. It's a slippery slope, but I think we can all agree that Alexa playing songs when you're around or even if you're not around is probably a fine thing. At this point, it's more just the crowding out of the human interaction that I'm very concerned about.

Yeah. My kid would otherwise be asking me to play the Blippi Garbage Truck song 400 times, so I'm somewhat sympathetic to that.

You've mentioned that we're pretty early in many ways in the development of this technology, but I also know that you're really concerned with some of the things that you're seeing. Is this a moment where you feel like the government has a role to play in regulating the technology, and if so, what do you think that role should be?

Absolutely. I think we know in Norway they've actually set out a law that there's no generative AI in the school system in the early years. In the same way, I think we're seeing that in the U.S. In some ways, this really mirrors the industrial food revolution. When the food revolution happened, initially there were no guardrails, and eventually we started the Pure Food and Drug Act, and then eventually we got nutrition labels.

Nutrition facts.

Nutrition labels.

Yeah, yeah.

The nutrition labels. These are the things that are going to happen. We're just at the very beginning. But hopefully we'll start seeing it.

But what do you think of the Norway approach? Does it seem wise to say, “Hey, until we've done a little bit more research, we're going to keep kids under 16 from using ChatGPT in schools for any reason”? Does that seem wise?

I think it's taking a scientific approach. I don't think they're saying, “No, never.” I think they're saying, “Wow, this is happening so fast. Let's take a step back. We know what works.” Look, if we didn't know how to educate humans, I would get it, but I think they're just taking a prudent approach.

I'm a scientist who is about innovation. I think about what happens in medicine, and we're much more thoughtful when we have a new drug on the market. It's not like we have to say, “Ooh, I wonder what's going to happen to my child if I give them this new medication.” I'm not sure why this new technology, which is so incredibly powerful—it's amazing, right?—but we have to acknowledge that it's a powerful technology, that we need to understand what happens to us.

I think the story of AI is as much about technology as it is about understanding what it does to humans, and I think that's what Norway is saying.

5. Parents Need A Safety Checklist

Well, let's try to give parents some practical advice. You suggest that when parents are trying to figure out whether a particular AI tool or product is suitable for them, they should ask themselves 6 questions. What are the questions that parents should be asking?

Because I'm a surgeon and I love acronyms, it's called DETECT. So what is the DETECT method? D is design. What is this tool designed for? Is it interacting directly with your child, or is it helping you? Do you really need it?

E is, was it ethically trained? T is, were there any troubles with this technology in children? E is, what is the evidence? Does it really do what it's supposed to do? C is confidentiality. What happens to your child's data? And lastly is T: What is it teaching from a value standpoint?

Hmm.

That DETECT method is used so that parents can quickly and easily figure out whether this is something that they want in their child's life.

All right, so I want to see if we can apply your DETECT framework to an actual AI thing that is on the market right now. There is something called Cradlewise, which I learned about from a tweet by Sam Altman, the CEO of OpenAI. He has a baby, and he tweeted that they had bought, quote, “a lot of silly baby things that we haven't needed, but definitely I recommend a Cradlewise crib and a lot more burp rags than you think you could possibly need.”

As for the Cradlewise, though, this is apparently a smart crib that uses AI to detect early signs of a baby waking up and automatically starts a gentle bouncing motion and soothing track to get them back to sleep. So if we were to use your framework, doctor, how would we rate Cradlewise?

Dana Suskind

All I can say is I wish I had had one when I had my children.

Yeah.

Dana Suskind

What is it designed to do? In some ways, it's designed to support the parents by helping the child sleep. How is it trained? I think it's probably ethically trained. Is there any evidence that kids have had trouble?

Not that I'm aware of, but I haven't so much as Googled it yet. But that would be a thing to do: investigate whether there have been any problems with the device.

Dana Suskind

Problems.

Yeah.

Dana Suskind

And what the evidence is. Probably some researchers have looked to see: Does it keep kids asleep? Do they have any problems? And I think the long and short of it is that it feels very aligned with using technology to enhance, not replace, Sam Altman's parenting—to allow him to have a good night's sleep and let the baby sleep.

Yeah.

Dana Suskind

Just hearing it for the first time, it sounds like a good thing.

It seems okay. I'd take a bed that would actually gently bounce me and rock me back to sleep when I wake up in the middle of the night, because whatever I'm doing now, it's not working for me.

You just have to update the firmware on your Eight Sleep.

Yeah.

I'm Googling this right now, and it seems like there have not been serious problems with the Cradlewise smart bassinet, although it does say that it has proprietary linens. You cannot buy cheap third-party crib sheets because the mattress is a nonstandard shape.

Cradlewise was one of my favorite songs in The Sound of Music.

That was “Edelweiss.”

Oh, yeah.

Dana Suskind

Hey, listen, it's either that or my late husband and I had 3 kids in our bed at 1 time, and I don't know, if I had had Cradlewise, we might have had more sleep.

Wow.

Yeah, that's wild.

That's heroic.

Yeah, seriously.

Dana Suskind

Yeah, I know.

I think this falls, for me, into the category of: It makes you a better parent. It does actually make you a better parent to get 6 hours of sleep in a night.

It does. Here's what I'm going to say: Do not overclock your Cradlewise so it starts shaking incredibly quickly. Even if it seems like it would be kind of funny, don't do it.

I do think eventually we will need some kind of TV rating system for AI products. I'm curious whether you've heard anyone suggest that, or whether you have any ideas about how that could be implemented.

Dana Suskind

100%. Actually, I talk about the Good Housekeeping Seal of Approval, which came out of the industrial food revolution, and suggest that we actually need something similar. I've heard different groups talking about doing it. Common Sense Media, I think, is thinking about it, and others. Maybe you all should be doing it.

The Hard Fork Seal of Approval.

Yeah.

Dana Suskind

Yes.

I think one of the hardest things is that these models are generally useful. They can do amazing things that are very good for kids. They can teach them science, they can teach them to read, they can give them creative outlets, and they can addict them or become harmful companions or shut out the rest of their human friends because it's just so fun to talk to the AI companion.

I want the AI companies thinking more about parental controls and visibility, because I think parents should know what their kids are talking about. But, as with screen time, I think there's this dangerous tendency to flatten every product into a single verdict, which is: It really matters what your kid is watching on their screen.

Yeah. It also really matters what your kid is talking to AI about, and I think the more that parents can be aware of and respond to that rather than just saying, “This product is okay, and this product is bad,” the better off kids will be.

Well, Kevin, that actually brings up something interesting, which is that earlier during this interview you brought up the Bernoulli effect.

Yeah.

I thought that I could imagine a lot of children listening to this and racing to their AI to find out what it is. So I thought I would take this opportunity, as a human who wants human connection with children, to share what the Bernoulli effect is. You guys already know this, but it's...

Honestly, Casey, it's more commonly called the Bernoulli principle, so I don't want to embarrass you, but it really is known as the principle. It states that for a fluid such as air or water flowing smoothly, an increase in the fluid speed is accompanied by a decrease in its pressure. So children, I hope that you heard that, internalized that, and are not going to ask AI about it.

It's what allows airplanes to fly.

Exactly.

Didn't you learn that in middle school science?

Yeah.

Okay, let's get back to Dana Suskind here.

I just had to say that.

We just like to do our little bits and force our guests to suffer through them. Okay.

Dana Suskind

Well, I'll tell you an interesting story. Do you want me to tell you an interesting story?

Yes.

Yes.

Dana Suskind

Related to knowledge?

Yes.

Dana Suskind

In our center, we've built this really cool computer-adaptive tool that measures what parents know about child development in all domains. It's very predictive of what parents do and child outcomes, so we use it in lots of research. Recently—we haven't published it yet—we had Claude take the, it's called the SPEAK, to see how Claude would do.

So I'm going to ask you all: Do you think Claude did really well, mediocre, or not so great on its knowledge of child development?

Really well.

I would say well, yeah.

Dana Suskind

You're right.

It aced it.

Dana Suskind

Claude did incredibly well. Claude aced it. Because we didn't want the story to be, “Oh, you know what? Now Claude can parent,” because that's not the message. The message is that Claude can obviously be a resource. When parents have a question about child development, they can confidently ask Claude, and Claude will give a right answer.

That tracks 100% with my own experience of these tools, which is that they are often more seasoned and more confident about parenting advice than I am, and they've been a very good resource for me when I'm feeling stuck or trying to deal with some behavioral thing that I don't know how to deal with, or I just need a little backup. I can just go to Claude, Gemini, or ChatGPT and ask the question, and usually the answer it gives back is pretty good.

Dr. Suskind, I want to ask you about the title of your book, which is essentially a provocation, as I read it, that says that as AI tools become more normalized, this idea of being human-raised is going to become rare and maybe something of a luxury. If you are born into a family with a certain level of security and privilege, you will get the human experience of having human parents, and if you are not, you will be raised by the machines. Am I reading that right? Is that what you think is likely to happen?

Dana Suskind

I'm writing this book so that hopefully it will not happen. But yes, in the same way that 100-plus years ago you never questioned whether the food on your table was organic or farm-raised, and then, fast-forward with the rise of ultra-processed food, suddenly organic was for children with privilege, in the same way, my concern is that artificial alternatives will become the cheap calories of brain nutrition, and human-raised and human connection will become a luxury item. We just can't allow that to happen.

Every child deserves to have parents and caregivers who love them raising them. That's not to say that AI can't support those parents and support some of the development, but I think we need to be very careful because we've seen this story before.

Well, Dr. Suskind, thank you so much for coming on and for your great book, Human Raised. I think this is a topic that a lot of parents are really interested in right now, and so I predict it'll do pretty well.

Dana Suskind

Thank you so much.

Thanks so much.

Thanks.

Thanks for coming.

Dana Suskind

So great to see you all.

When we come back, scandals and controversies abound in the world of prediction markets. We'll recap it all on our new segment, Against All Odds.

6. Prediction Markets Sell The Dream

Well, Casey, one of the other stories that keeps on giving is prediction markets. This has been a source of nonstop entertainment and scandal this year, and so we thought we would introduce a new segment where we round up all the latest news and scandals about prediction markets. We're calling that segment Against All Odds.

Okay, well, Casey, I think we should start this segment by looking at a recently released ad from the prediction market company Polymarket. I'm just going to play this for you.

If you could ask one question—so we start with Rick Rubin—what would you ask?

Dana Suskind

Can the U.S. win it all?

Est-ce que la France retrouverait la gloire?

Dana Suskind

Que dirias, football o soccer? Nda Senegal muna shoke aadu nabi?

Pueden las tradiciones sobrevivir? Werden die Deutschen gewinnen? Would the world be better off 4 years from now?

Sort of a Benetton ad of gambling. A series of questions appears on screens: Will borders matter in 100 years? Do aliens exist? Will Messi win again? Will tradition evolve? There's a Kanye West song in the background, which is a choice.

The people who never stop asking are the ones who find answers. Now, Casey, did that ad make you want to open up an account on an offshore, crypto-denominated prediction market and start gambling with your savings?

I find this ad so offensive. The way that it presents prediction markets as, first of all, grounded in concepts of multiculturalism and togetherness. It's like it has this almost mystical feel of, if we could just all get in the same room and gamble together, maybe wars wouldn't happen anymore. That's the feeling.

Yeah.

I truly don't get it at all.

It's very sad that they didn't feature the actual customers of Polymarket, who are troops on active duty in war zones.

Or middle schoolers.

Yeah. People in Bad Bunny's Super Bowl entourage.

My half-joke for the official marketing tagline of any prediction market should just be “Betray your friends,” because that's the way to make money out of a prediction market. You found out something you're not supposed to know? Get on a prediction market.

Yeah. Go make some money.

That's the ad I want to see: Screw over someone close to you.

Yeah. Well, my second question about this is: How much do you think they paid Rick Rubin, and do you think they paid him in Polymarket credits or in dollars?

I'm going to guess it was in the millions, Kevin.

Yeah.

Yeah.

You don't think Rick Rubin just did this one for free as a testament to his commitment to craft and creativity and the artist's dream?

No. Rick Rubin produced the Beastie Boys. He has not had to work in a very long time.

These ads from prediction markets are just inescapable. I was watching the World Cup on TV over the break, and every other ad is now for some kind of prediction market or gambling site. It is wild.

7. The Donk Truth Machine

But there's also been more than advertising happening in prediction markets. I want to talk about this incredible story that David Segal at The New York Times recently wrote called “The Donking of a Truth Machine.”

Yes.

It's about a very silly but serious and high-stakes controversy on Polymarket, the very company whose ad we just saw. It all boils down to this seemingly crazy question: Did a guy in a video say the word “donk”?

Mm-hmm.

The backstory here is that there is a professional gamer named Daniel Kryszkevicz, whose nickname is donk. Back in April, there was an active betting market on Polymarket about whether, during the broadcast of this video game tournament, someone would say the name donk.

Crucially, donk was not participating in this video game tournament.

Because if he was participating, it would seem very likely that at some point someone would say the word “donk.”

Yes. This is the kind of silly market that often happens. Will this word be uttered during a CEO's earnings call? Or will the White House press conference say some word? People bet on this. It's a time-tested form of gambling on these websites.

But this particular market, the donk market, went sideways because there was this very long, 7-hour-long gaming tournament stream. It appeared that donk was not mentioned, but then somebody on this market was like, “Wait a minute. At one point, the commentator doing the play-by-play on this gaming tournament appeared to stumble over the word ‘don't,’ and it sounded like ‘donk’ to this person.” And they basically—

No. Here's what I'd like you to do. I'd like you to say a sentence with the word “don't,” but make it sound like “donk.”

“Donk look back in anger?”

Okay. Very good. Thank you. You can go on now.

Okay. This started a very fierce argument about whether this commentator had actually said “donk” or whether this was just a slip, and whether an accidental donk is still a donk for the purposes of resolving this, as this big fight occurs.

And David goes into this fight in very amusing detail in his story. But to boil it down, basically, whenever there's a contested market where the resolution isn't clear on Polymarket, there is a system called the optimistic oracle. Basically, the people who hold this special crypto token called UMA can vote on disputed outcomes.

At the end of the sentence, I wanted to jump out a window. I just have to say that. Learning all of that information, I truly want to forget I've ever learned that information.

None of those words are in the Bible.

Truly.

So basically, the way this works is the more of these UMA tokens you have, the more voting power you have in these disputed markets. And basically, this is their way of trying to inject some sort of fair resolution criterion into these markets where it's not exactly clear who is supposed to win and who is supposed to lose.

And the fair criterion is that whoever amasses the most UMA tokens is the arbiter of truth.

Yes.

What a great system.

Yes, it's a crazy way to run a market. You're essentially giving the power to resolve these markets and determine who gets paid out and who doesn't to the richest people on the platform, or at least the people who have spent the most money on this meaningless crypto token.

Anyway, this one resolved in predictable fashion, which was that a company called UMA Rocks amassed a bunch of these tokens and announced that it was going to vote no on this question of whether “donk” was uttered on this gaming stream, which led to a bunch of follow-on no votes. So the nos won out in the end, but it was sort of a rigged resolution.

So there's even a form of gambling in how this thing gets resolved—

Yes.

—is what it sounds like.

Yes.

Yeah.

You can gamble on the gambling mechanism. It sort of turtles all the way down.

When I was very young, I just decided that gambling was not going to be for me. And to watch the way that the world has transformed over the past 20 years to turn absolutely everything into gambling, I continuously feel like I'm losing my mind.

Because here's the thing: You will lose money. Statistically, you will lose money. This is not going to turn out well for you. And yet now some meaningful portion of our economy is just people who have decided to lose money that way.

Yeah.

Why don't you lose money on journalism? Buy a subscription to a publication. At least you'll get something in return other than suffering.

But you won't get the satisfaction of amassing UMA tokens in order to resolve the market on whether “donk” was uttered during a video game broadcast, and that's what really counts in today's day and age.

Well, that's true. So what else is going on in—

Okay.

—here in Against the Odds, Kevin?

Well, speaking of Polymarket—

Yeah.

Casey, would you believe we have another Polymarket scandal?

I'd believe it.

8. Polymarket Fakes Its Winners

So a couple of weeks ago, there was this great investigation in The Wall Street Journal about Polymarket, and specifically its strategy for social media marketing. They've done these social media videos for a while now where they'll pay a creator to film a video of themselves making a bet on Polymarket. And the Journal actually went back and looked at those bets to determine, first, whether they were real—whether these bets actually happened on Polymarket—and second, whether, if they had been real, the people making these bets would have won or lost money.

Oh, I can't wait to find out.

So they analyzed more than 1,100 videos from 10 creators, and they found that, although 70% of these videos showed a bet being placed, none of these bets were real. These were all fake bets. 118 of the videos showed the creators winning, and the videos suggested that these people would have won almost $900,000 in total. But in reality, if those bets had really been placed, these same people would have lost more than $166,000.

So in some sense, it's like every ad you've ever seen for a casino, where it's just people celebrating winning.

Right, just making it rain, tossing $100 bills in the air.

Yes, but it turns out that you do not always win on Polymarket, and you are not always happy and tossing money everywhere.

But I love that. First of all, this was a really great story by the Journal. And second of all, what I love about it is that it shows the actual surefire way to make money on Polymarket, which is to make a video as a creator in which everyone is lying about everything. And then you'll really rake it in.

The other guaranteed method for making money on Polymarket: insider trading.

Mm-hmm.

But that one doesn't play as well on camera.

9. Meta Builds A Betting App

Okay, we have one more prediction market story this week. This one was broken by our colleagues at The Times, Mike Isaac and David Yaffe-Bellany. This one is about Meta. The headline is “Mark Zuckerberg Directed Meta to Create a Prediction Markets App.”

Mike and David report that after seeing the success of prediction markets, Mark Zuckerberg got the bright idea to build his own. Two people with knowledge of the matter at Meta said that Zuckerberg had recently dispatched a small team to create a smartphone app similar to Polymarket and Kalshi, where, instead of wagering real money, the app would rely on a kind of video game-like fake-money point system. But the company, they reported, had not ruled out the eventual use of real-money betting. The app is internally called Arena and would be functioning independently from Meta's existing social networking apps. Casey, what did you make of the Meta Arena story?

So this was a great scoop, and I think the truth is I don't know how big of a reaction I should have to this, because Meta does stuff like this all the time. Whenever a hot new thing comes out, they spin up their own version of it. Sometimes it works, like with Instagram Stories or with Reels, which was its answer to TikTok. Many, many other times it just completely fails and flops.

I don't know what is going to happen here because, as we've just been talking about, these prediction markets are becoming extremely popular. And I imagine that the version of it where you can't lose real money will be less popular, weirdly, than the one where you can lose all your money.

But I suspect that if it is successful, one, they will absolutely do real-money wagering. I think they're probably salivating over that prospect. And two, it will come to Meta's social networking apps. And you will see little carousels for this in your Instagram and your Facebook. I just think, yeah—

I love this. Imagine you're on Instagram and you see a post of your friend with his new girlfriend—

Yeah.

—and right below it there's a market for how long this relationship will last?

Totally. Totally announcing your engagement. And then it's like, “What are the two-year odds on this one, gang?” Yeah, no, that's so perfect, and they would absolutely do it.

You know, the ethos at Meta—and I know this from speaking to them—is that they believe that if people are doing it, then we should respect that they're doing it and let them do more of it and encourage them to do more of it. And they almost don't care what that thing is. But if you want to make a bet, if you want to make endless bets, and there's a chance you'll talk about it on Meta properties, where they can show you an ad and maybe even take a cut of your prediction, they're absolutely going to do it.

I just love that this is a company that has been mired in lawsuits about the addictive nature of its platform, especially for young users—

—and is currently mired in dozens of these lawsuits around the country.

It's like, you know what sounds like a really great next focus area for us? Gambling. Gambling. The famously addictive behavior.

Yeah.

I was laughing, though, because there was a story after you guys had the scoop that came out and said that some Meta employees were really concerned. They were really disheartened to learn that their company was exploring a prediction market app—of all things. And I was like, what? Is this somehow going to interrupt you from the important work of sending push notifications to 13-year-olds in the middle of the night telling them to check Instagram? Do we really need to protect that—

Yes.

—from the prediction market?

I'm horrified to learn that there's gambling going on in this establishment.

Yeah, if kids spend too much time on these prediction markets, Kevin, they might not have time to get an eating disorder. So we really have to watch the trade-offs here.

Wait, wait, you are laughing, but that is literally something that the CEO of Kalshi recently said.

Wait, really?

Yes.

Wait, what did he say?

Tarek Matsour, the CEO of Kalshi, recently literally compared gambling on Kalshi and other prediction markets to the cure for Instagram addiction. He said—the way to think about it is, the more that people spend time on Kalshi and prediction markets, the less they're spending time brain-rotting on Instagram.

Wow.

And Mark Zuckerberg heard that, and he said, “Por que no los dos?”

The cure to one addiction, Kevin? Another addiction.

Yes. Anyway, I’m sure this will end well.

Mm-hmm. Yeah, well, everyone involved, I would just start preparing for your testimony to Congress that you’ll be giving eventually. Probably not soon, but eventually.

Ugh.

And that’s something you truly can bet on, Kevin, but I wouldn’t use real money.

All right, that is Against All Odds, our new and possibly final installment of our prediction markets roundup.

Hard Fork is produced by Whitney Jones and Rachel Cohn. We’re edited by Viren Pavich. We’re fact-checked by Caitlin Love. Today’s show was engineered by Elisa Moxley. Original music by Alicia Meitube, Rowan Nemestowe, Elisa Moxley, Dan Powell. Video production by Sawyer Roquet, Jake Nickell, and Chris Schott. You can watch this whole episode on YouTube at youtube.com/hardfork. Special thanks to Paula Schumann, Pui Wing Tam, Brooke Minters, and Dalia Haddad. You can email us at hardfork@nytimes.com on whether you think we’ll say donk on next week’s episode. Donk.

Donk. Jinx.