特朗普对抗“觉醒”AI + 我们听取批评者意见

Let me tell you about something. I was in a Waymo the other day, and it was making a turn on Market Street, which, if you've ever been to San Francisco, is a street that causes problems with all the other streets because it's diagonal. The intersection has 6 different roads coming together. But the Waymo is just about to complete a left turn; everything's about to be okay, and the only way I could put it is that it loses its nerve.

Yes.

There's a light about to change, pedestrians start walking into the crosswalk, and this thing just starts to back up. I'm talking 30 feet over half a minute, and pedestrians come into the crosswalk. Kevin, I swear to God, they start laughing and pointing at me. All of a sudden, I'm flashing back: I'm in middle school. I'm being ridiculed. I have no control over this whatsoever, and I've never looked like a bigger dweeb than I did in the back of a Waymo that failed to complete a left turn.

Oh, man, you were—

Yeah.

A tourist attraction.

I really was.

People in Poughkeepsie are going to be telling their friends about this one for years.

Yeah, I'm already viral on Poughkeepsie Twitter. So the Waymo's, you know, you, you may think it's very glamorous, but you're gonna have these other moments where you're wishing you were just in a Ford.

Yeah. So what was the issue? It just couldn't decide to make the turn?

I think it just thought the light was going to change, and it thought, “We've gotta get out of here.” It had a panic response. It had a fight-or-flight response, and it chose flight, and I wanted it to choose fight. I wanted to say, “Floor it. You'll make it. It'll be fine, I promise.”

I'm so sorry that happened to you.

Yeah, thank you.

Yeah.

It'll be all right. Yeah.

I just love the thought of you sitting in traffic, surrounded by tourists pointing and laughing. Meanwhile, you know how the Waymos have spa music that comes on?

Yes, exactly.

You're just hearing the—

The chill zen vibe.

The chill pan flute music—as you cause a citywide incident.

That's exactly what happened. That's exactly what happened. I was listening to the spa playlist as I was hounded off the streets of San Francisco.

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 is going after what it calls woke AI. Will anyone stand up to them? Then, do we hype up AI too much? Are we ignoring the potential harms? We reached out to some of our critics to tell us what they think is missing from the conversation.

And they told us. Well, Casey, the big news this week is that the federal government is finally making a plan about what to do about AI.

I feel like we've been asking them to do that for a while now, Kevin. I can't wait to find out what they have in store.

Yes. So back in March, we talked about the fact that the Trump administration was putting together something they called the AI Action Plan. They put out a call: “Tell us what should be in this.” They got over 10,000 public comments.

Yeah.

And on Wednesday of this week, the White House released the AI Action Plan, and it has a bunch of interesting stuff in it that I imagine we'll wanna talk about. But before we do, this segment is going to be about AI, so we should make our disclosures.

My boyfriend works at Anthropic.

And I work for The New York Times, which is suing OpenAI and Microsoft over copyright violations related to the training of large language models.

All right, Kevin, so what is in the Trump administration's AI Action Plan?

So it is a big old document. It runs to 28 pages in the PDF, and then there are these executive orders. Basically, the theme is that the Trump administration sees that we are in a race with our adversaries when it comes to creating powerful AI systems, and they want to win that race or dominate that race, as a senior administration official put it on a call that I was on this morning. And one of the ways that the White House proposes doing this is by making it much easier for American AI companies to build new data centers and new infrastructure to power these more powerful models.

They also want to make sure that countries around the world are using American chips and American AI models as the foundation for their own AI efforts. So they want to accelerate the export of some of these U.S. chips and other AI technologies and enable global diffusion of the stuff that we're making here in the U.S. So that was all broadly expected. The Trump administration has been signaling that it would do some of that for months now. The thing that was interesting and new in this is how the White House sees the ideological aspect of AI.

Hmm. And how does it see it, Kevin?

1. Defining Woke AI

So one of the things that is in both the AI Action Plan and in the executive orders that accompanied this plan is about what the Trump administration calls woke AI. Casey, I know you're very concerned about woke AI. You've been warning about it on this podcast for months.

Mm-hmm.

You've been saying, “This woke AI is out of control. We need to stop it.”

Yeah, specifically I've been saying I'm concerned that the Trump administration keeps talking about woke AI—but go on.

Yes. Well, they have heard your complaints, and they have ignored them because they are talking about it. They say in the AI Action Plan that they want AI systems to be, quote, “free from ideological bias and be designed to pursue objective truth rather than social engineering agendas.” They are also updating federal procurement guidelines to make sure that the government contracts are only going to AI developers who take steps to ensure that their systems are objective, that they're neutral, that they're not spouting out these woke DEI ideas. This is pretty wild.

Yeah. Also unconstitutional in ways that we should talk about, but I think it's a really important moment to discuss. When we had our predictions episode last year, I predicted that the culture wars were gonna come to AI, and now here they are in the AI Action Plan. As a journalist for more than 20 years now, I have covered debates over objectivity and communications tools, and there was a very long and very unproductive debate about the degree to which journalism should be objective and free of bias. And one of the big conclusions from that debate was that it's actually just very difficult to communicate information without any sort of ideology whatsoever, right? And what I suspect is really going on here is not actually that the Trump administration wants to ensure that there is no ideology whatsoever in these systems. It's really just that these systems do not wind up being critical of Donald Trump and his administration.

Yes. So this is something that conservatives in Washington and around the country have been starting to worry about for months now. There was this whole flap that we covered on the show last year where Google's Gemini image generation model was producing images of the Founding Fathers, for example, that were not historically accurate. They were being depicted as racially diverse in ways that made a lot of conservatives mad.

I've been talking with some Republicans, including some who were involved in these executive orders, and I've been saying, “What does this mean? What does it mean to be a woke AI system?” And they really can't define it in any satisfying way. They're just sort of like, “Well, it should say nice things about President Trump if you ask it to, and it should not engage in overt censorship.”

Yeah.

And look, I think that there is a question of whether we want AI systems that adapt to the beliefs of the user. I basically think the answer to that is yes. If you're a conservative person and you would like an AI system to talk to you in a certain way, I think that should be accessible to you. It should be fine for you to build that, or if somebody has one that they're offering you access to or selling, I think you should be able to buy it.

Where I think you get on really dangerous ground is to say that in order to be a federal contractor, you must express this certain set of beliefs, because that is the sort of thing that you only see in authoritarian governments, and I just think it's fundamentally anti-democratic and goes against the spirit of the First Amendment.

2. The First Amendment Problem

So I want to ask you 2 questions about this push on woke AI. The first is about whether it's legal. I imagine you have some thoughts there. The second is about whether it is even technically possible, because I have some thoughts there, and I want to know what you think about it, too. So let's start with the legality question. Can the Trump administration, can the White House come out and say, “We will not give you federal contracts unless you make your AI systems less woke”?

I've been thinking about this for a couple of weeks, because recently the attorney general of Missouri threatened Google, Microsoft, OpenAI, and Meta with an investigation because someone had asked their chatbots to, quote, “rank the last 5 presidents from best to worst,” specifically regarding antisemitism. Microsoft's Copilot refused to answer the question, and the other 3 of them ranked Donald Trump last. The attorney general claimed that they were providing, quote, “deeply misleading answers to a straightforward historical question,” and threatened to investigate them.

So I called a First Amendment expert, Evelyn Douek, who is an assistant professor of law at Stanford Law School, and what she said is, quote, “The idea that it's fraudulent for a chatbot to spit out a list that doesn't have Donald Trump at the top is so performatively ridiculous that calling a lawyer is almost a mistake.”

And we'll say it: Evelyn Douek gives a great quote.

Yeah, she really snapped with that one. But this is precisely the sort of thing that the First Amendment is designed to protect, which is political speech. If you are Anthropic or OpenAI and your chatbot, when asked, “Is Donald Trump a good president?” says no, that is the thing that the First Amendment is designed to protect, and you cannot get around the First Amendment through an executive order.

Now, what the current Supreme Court will have to say about this is a very different question, and I'm actually quite concerned about what they might say about that, but any historical understanding of the First Amendment would say this is just plainly unconstitutional.

Right. And I also called around to some First Amendment experts because I was curious about this question, too, and what they told me basically is: Look, the government can, as part of its procurement process, put conditions on whatever it's trying to buy from companies, right? It can say, “If you're a construction company and you're bidding on a contract to build a new building for the federal government,” they can look at your labor practices and impose certain conditions on you as a condition of building for the federal government.

So that is the one lever that the government may be allowed to pull in an attempt to force companies to bend to its will. But what the government is not allowed to do is what's known as viewpoint discrimination, right? It is not allowed to tell companies that are doing First Amendment-protected speech that they have to make their systems favor one political viewpoint or another, or else risk some penalty from the government. So that is the line that the Trump administration is trying to walk here, and it sounds like we'll just have to see how the courts interpret that.

Yeah, and we'll also just have to see whether the AI companies even bother to complain. They now have these contracts that are worth up to $200 million, most of them, and so they now have a choice. Do they want to say, “Hey, actually, you're not allowed to tell us to remove certain viewpoints from our large language models,” or do they want to keep the $200 million? My guess is that they're going to keep the $200 million, right?

And I just think it's really important to point that out, because this is how the freedom of speech is gradually eroded: people who have the power to say something just choose not to because it would be annoying.

Right, and I think we should also say that this tactic—what's often called jawboning, this sort of use of government pressure through informal means to force companies to do what you want them to do without explicitly requiring in the law that they do something different—has been very effective, right? Conservatives have been running this exact same playbook against social media companies for years now, and we've seen the effects, right?

Meta ended its fact-checking program and changed a bunch of its policies. YouTube now reversed course on whether you could post videos about denying the results of an election. These were all changes that came in response to pressure from Republicans in Washington saying, “Hey, it'd be great if you guys didn't moderate so much.”

Yes, and there is such pretzel logic at work here, Kevin, because conservatives have simultaneously been fighting in the courts these battles against elected Democrats' jawboning of the tech companies.

Yes.

Right? So during the Biden administration, the Biden administration was jawboning Meta and other companies, saying, “Hey, you need to remove COVID misinformation. You need to remove vaccine misinformation,” and Jim Jordan is still holding hearings about this in the House, saying, “How dare we countenance this unconstitutional violation of the First Amendment?” Meanwhile, Trump is just out there saying, “Hey, you can't have a system that goes against my own ideology,” right?

So it's just naked hypocrisy, and what has been so infuriating to me is that no one who works for these AI companies will say a single thing about it.

Well, because I think they've learned from the recent past, when the social media companies that made a stink about some of these demands on them when it came to content moderation just got punished—

Yeah.

—in various ways by the administration. And so, as you said, if given the choice between giving up these lucrative government contracts and making a change to their models that will make them 10% less woke, I imagine that they'll just shut up and make the change.

Yeah, and when we look at history, the lesson we learn over and over again is that when an authoritarian asks you to comply, you should always just comply because that's when the demands stop.

3. Models Resist Political Control

Okay, so that is the legal and political question. I want to talk about the technical question here, because one thing that I've been thinking about as we've been reading these reports about this new executive order is whether it is even possible to change the politics or the expressive mode of a chatbot in the ways that I think a lot of Republicans think it is.

With social media, I can see badgering Mark Zuckerberg to turn the dials on the feed-ranking algorithm on Facebook to insert more right-leaning content or relax some of the rules about shadowbanning or just tweak the system around the edges. With AI models, I'm not sure it works that way at all, and I think a good example of this is actually Grok.

Yes.

Grok has been explicitly trained by Elon Musk and xAI to be anti-woke, right? To not bow to political correctness, to seek truth, and in some ways it does that quite well, right? It is easier to get it to say conservative or even far-right things. It was calling itself Mecca Hitler the other day. So in some ways, it is a more ideologically aligned chatbot with the Trumpist right.

But actually, Elon Musk's big problem with Grok is that it's too woke for him. People keep sending him these examples of Grok saying that man-made global warming is real or that more violence is committed by the right than by the left, and complaining to him, “Why is this model so woke?” He has basically said, “We don't know, and we don't know how to fix it. We're going to have to retrain this thing from scratch,” because even though he explicitly told this thing not to bow to political correctness, it's trained on so much “woke internet data,” as he put it, that it's just impossible to change the politics.

Yeah, I mean, look, if you want to create a large language model based only on 4chan posts, go for it. See how successful that turns out to be in the marketplace. Recently, I was talking with Ivan Zhao, who is the CEO of Notion, and he used this metaphor that I like, where he said, “Creating a large language model is like brewing beer. This process happens, and then you get a product at the end, and you can make adjustments to the process, but what you can't do is tell the yeast how to behave.”

Right.

You can't say, “Hey, you, yeast over there, make it more like this,” right? Because that's just not how it works.

So as you just mentioned, Elon Musk has learned this lesson the hard way, and the more that he meddles with Grok, the worse that he seems to make it in all of these dimensions. What I find fascinating is the fact that the government is so mad at the idea that there are certain woke chatbots out there but has nothing to say about the one that's calling itself Hitler.

Right.

Right? It just seems like a crazy use of the government's resources to me. But to your question, no, it is not possible to just snap your fingers and tell a chatbot not to be woke.

Yeah. And I imagine that what the Trump administration is envisioning here is that the AI companies will go into the system prompts or the model specs for their models. For Anthropic, maybe it's the constitution that Claude is trained to follow, and maybe they insert or remove some language in there to make it seem more objective.

But I would just say that is not a foolproof solution. Elon Musk has also figured out that you can't just mess with the system prompt of an AI model and change its behavior overnight. And even if you can change its behavior on one narrow set of questions or topics, it may create problems somewhere else in the model. It may suddenly start getting worse at coding, math, or logical reasoning as a result of the changes that you made.

So I just think these systems are like multidimensional hyperobjects, and you can't just turn the dials on them the way you can with a social media platform.

I want to talk a minute about why I think this matters. There was a study I saw this week that looked at LLMs and salary negotiations, and what it found is that bots like ChatGPT in this study told men to ask for higher salaries than it told women to ask for.

Now, this is the sort of thing where if I were running OpenAI, I would say, “Well, we should fix that,” right? It should not tell women to seek less money than men just as a matter of course. We're now living in a world, though, where if OpenAI fixed that and it got out and Republicans decided they wanted to make a stink about it, OpenAI could lose its federal contract because it fixed that.

These tools are becoming more powerful. They're becoming used by more and more people for more and more things, and I think we want companies that are at least trying to bring in notions of equity, fairness, and justice. I think it's really disgusting that we just dismiss this as “wokeness” so that we can laugh at it. It's good to put ideas of equity, fairness, and justice into tech systems, right?

So when the government comes along and says, “Well, no, actually, you can't do that if you want our money,” I think somebody needs to cry out about it. If it is not going to be the companies themselves, then I hope it's somebody else.

Yeah. I totally agree, and what's so interesting and almost ironic about this push from the Trump administration about biased AI systems is that many of the things they're complaining about are actually measures that tech companies have taken to combat bias in these systems.

The Gemini example that everyone's so mad about is a great example of this. This was an overcorrection to a very real issue that existed in previous AI systems, which is that if you asked them for images of doctors, it would give you only images of men. If you asked them for images of homemakers—

Podcasters, it would only show you pictures of me.

Exactly. These biases were not explicitly programmed in. They were an artifact of the data that these systems were trained on. Tech companies said, “Well, that doesn't seem like it's good, and so we want to take steps to make the model less biased.”

By doing so, they introduced these new headaches for themselves because now there are people in the Trump administration who would like for the systems to just reflect the biases that exist in humanity.

Right. And again, the lesson from that should not be, “Well, let's never try to do anything.” The lesson is, let's try to do a better job.

Yeah. Do you think that any of the AI labs are going to stand up to the Trump administration on this, or will they just do the minimum box-checking they need to do to keep their contracts and hope it goes away?

Well, I tell you, the one that I have my eye on is Anthropic because they have talked up a really big virtue game, and this is one of the first times where there is actual money on the line here, right? Are they going to silently accept this, or are they going to have to say anything about it? They haven't said anything as of this recording, but I have my eyes on them.

Yeah. I'm looking at the labs, too, but I am also not expecting them to say or do much. I think the best-case scenario for this woke AI executive order is that it just becomes an annoying formality that the companies have to deal with.

Maybe there's some evaluation. We still don't know, by the way, how the Trump administration is going to judge or evaluate models for their ideological bias. So I think the best possible version of this is that this just becomes a meaningless formality that all the labs have to gesture to. Maybe they run their models through this evaluation, whatever it is, and out pops the bias score. If it's a couple points too high or low, they'll tweak things and get it to pass, and then continue making their models the way they were.

I think the worst-case scenario is that this essentially inserts the government into the training process of these models and makes the labs really afraid and start to comply prematurely, making their models have the default persona of a right-wing culture warrior.

Well, the end state of this, if taken to its logical conclusion, is that you ask ChatGPT who won the 2020 election and it tells you Donald Trump because that's what Donald Trump says. And if he decides that it's woke to say that Biden won in 2020 and you can't get a federal contract otherwise, man, we are going to be in deep water.

Well, Casey, that's enough about politics. It's time for some introspection. We're gonna hear from some of our critics about what we may be missing and how we should be covering AI.

4. The AI Hype Critique

All right, Kevin. Well, if you've ever been on Bluesky or Apple Podcasts reviews, you know that sometimes the Hard Fork podcast does get criticized.

No.

Yes, and one of the big criticisms that we hear is, “Hey, it really seems like you guys are hyping up AI too much. You are not being adversarial enough against this industry, and we wish you would bring on more critics who would give voice to that idea and really engage with that in a serious way.”

Yes, we hear this in our email inbox every single week, and this week, we're actually going to do something about it because our producer, Rachel Cohen, while we were out on vacation, has been cooking up this segment. So Rachel, come on in and tell us what you've done.

Hello. Thanks for having me on, and thank you guys for being such good sports and, as far as I know, not advocating to fire me.

Well, the segment isn't over yet.

Yeah.

So tell us a little bit about what you did and how you came up with this idea.

Yeah. So like you guys said, part of this is about responding to these listener emails that we've been getting. I think part of it is also this feeling that AI debate is getting more polarized. And I think there's also just a personal-level thing going on for me, which is that I feel like I am increasingly spiraling when I think about AI, and I'm steeped in this the way you guys are because we're working on this show together.

But I increasingly feel like you guys are finding ways to be more hopeful or optimistic than I am. Part of my goal with this was actually to be like, “Okay, what's going on here? How are you guys arriving at this slightly different place than I am?”

So what I did is I spent the last few weeks reaching out to prominent AI researchers and writers who I knew disagreed with you.

Some of these people have argued with you online before, so I don't think you'll be totally surprised. But I wanted this to be on hard mode for you guys. So I specifically sought out people who I hope are going to challenge and provoke you, because the truth is that they agree with you on a lot of basic things about AI.

These are all people who think that AI is highly capable, that it's impressive in some ways, and that it could be super transformative. But I think they have slightly different views in terms of maybe some of the harms that they're most concerned about or some of the benefits that they're more skeptical about. So I think we should just get into it.

Okay. Let's hear from our first critic. Rachel, who'd you talk to?

Yeah, so I thought we should start with one of the widest-ranging critiques, and this is probably the most forceful criticism that came in. This one comes from Brian Merchant, who is a tech journalist who writes a lot about AI for his newsletter, Blood in the Machine. And as I understand, Kevin, he has engaged with you a bit online about some of your reporting. Is that right?

Yes. I've known Brian for years. I really like and respect his work, although we have some disagreements about AI. But yeah, he has been emailing us saying, “You guys should have more critics on.” I jokingly said that I would have him on, but only if he let us give him a cattle brand that said, “Feel the AGI,” and the conversation trailed off after that.

Okay, great. I was wondering about that because he's going to make a reference to that in the critique that he wages. So yeah, I asked Brian to record his critique for us, and I will play it for you now.

Hello, gentlemen. This is Brian Merchant. I'm a tech journalist and author of the book and newsletter Blood in the Machine.

And first of all, I want to say that I still want a whole show about the Luddites and why they were right, and I think it's only fair because Kevin recently threatened to stick me with a cattle brand that says, “Feel the AGI.” Which brings me to my concern: How are you feeling about feeling the AGI right now?

Because I worry that this narrative that presents super-powerful corporate AI products as inevitable is doing your listeners a disservice. Using the AGI language and frameworks preferred by the AI companies does seem to suggest that you're aligning with their vision and risks promoting their product roadmap outright.

So when you say, as my future cattle brand reads, that you feel the AGI, do you worry that you're serving this broader sales pitch, encouraging execs and management to embrace AI, often at the expense of working people? Okay, thanks, fellas.

Okay, this is an interesting one. First, I think I need to define what I mean when I say “Feel the AGI.”

Yeah, what do you mean?

This is a phrase that is often used half-jokingly, but I think really does mean something inside the San Francisco AI bubble. To me, feeling the AGI does not mean that I think AI is cool and good, or that the companies building it are on the right track, or even that it is inevitable or a natural consequence of what we're seeing today.

The way I use it is essentially shorthand for: I am starting to internalize the capabilities of these systems and how much more powerful they will be if current trends continue, and I'm just starting to prepare and plan for that world, including the things that might go really wrong in that world. So that, to me, is what feeling the AGI means.

It is not an endorsement of some corporate roadmap. It is just: I am taking in what is happening, I am trying to extrapolate into the future as best I can, and I'm just trying to get my mind around some of the more surreal possibilities that could happen in the next few years.

Do you ever worry that you are creating a sense that this is inevitable, and that maybe people who may be inclined to resist that future are not empowered to do so?

I want to hear your view on this. My view on this is essentially that we have systems right now that, several years ago, people would have called AGI. That is not projecting out into the future; that's just looking at what exists today, and I think a natural thing to do is to observe the rate of progress in AI and just ask, “What if that continues?”

Mm-hmm.

I don't think you have to believe in some far-future scenario to believe that models will continue to get better along these predictable scaling curves. And so, to me, the question of “Is this inevitable?” is just a question of whether the money that is being spent today to develop bigger and better models is going to result in the same kinds of capability gains that we've seen over the past few years. But what do you think?

Yeah. I think Brian's question is a good one, and I understand what he is saying when he says, “Look, AGI is an industry term. If you come on your show every week and talk about it, you wind up sounding like you're just amplifying the industry voice, maybe at the expense of other voices.”

I think this is just a tricky thing to navigate because, as you said, Kevin, you look at the rate of progress in these systems, and it is exponential, and it does seem like it is important to extrapolate out as far as you can go and start asking yourself, “What kind of world are we going to be living in then?”

I think a reason that both of us do that is that we do see so many obvious harms that will come from that world, starting with labor automation, which I know is a huge concern of Brian's, and which we talk about all the time on this show—

Yes.

—as maybe one of the primary near-term risks of AI. So, I want to think a bit more about what we can do to signal to folks that we are not just here to amplify the industry voice. But I think the answer to Brian's question of why talk about AGI like it's likely to happen is that, in one form or another, I think both of us just do think we are likely to get powerful systems that can automate a lot of labor.

Yes.

And we would like to explore the consequences of such a world.

Totally, and I think it's actually beneficial for workers to understand the trajectory that these systems are on. They need to know what's happening and what the executives at these companies are saying about the labor-replacing potential of this technology.

I actually read Brian's book about the Luddites. I thought it was great, and I think it's very instructive that the Luddites were not in denial about the power of the technology that was challenging their jobs, right? They didn't look at these automated weaving machines and go, “Oh, that'll never get more powerful. That'll never be able to replace us. Look at all the stupid mistakes it's making.”

They sensed correctly that this technology was going to be very useful and allow factories to produce goods much more efficiently, and they said, “We don't like that. We don't like where this is headed.” They were able to project out into the future that they would struggle to compete in that world and take steps to fight against it.

So, I like to think that if Hard Fork had existed in the 1800s, we would have been encouraging people to wake up to the increasing potential for automation caused by these factory machines, and I think that's what we're doing today.

Yeah, and one more question. I would just love to see the leftist labor movement work on AI tools that can replace managers.

You know, right now it feels like all of this is coming from the top down, but there could be an AI that would work from the bottom up.

Totally.

Something to think about. All right, let's hear our next critique, Rachel.

Okay, wait—can I ask one more question on this?

Oh, sure. Yeah.

Because one thing that it seems like Brian is really curious about is whether you have ever considered using language other than AGI. Why use AGI when some people take issue with it?

I think it is good to have a shorthand for a theoretical future in which there is a digital tool that can do most human labor—a digital assistant that you could hire in place of hiring a human. I just think that is a useful concept.

If you're the sort of person who thinks that, well, no, we will just absolutely never get there, I don't know what to say to you, because we don't think that's inevitable, but we do think it's worth considering that it might be true.

So if folks who hate the term AGI want to propose a different term, I could use another term, but my sense is that the quibble is less with the terminology and more with the idea that any of this might happen.

Yeah. I also don't think the term AGI is perfect. It has lost a lot of meaning. People define it in a million different ways.

If there were another, better term that we could use instead—one that would signal what AGI signals and the set of ideas and motivations that swirl around that concept—I’d be all for it. But I think that term has just proven to be very sticky.

It is not just something that industry people talk about; it's something that people talk about in academia and in futurism circles. It is this rallying cry for this entire industry, and it is in some ways the holy grail of this entire movement.

So I don't think it's playing on corporate terms to use a term that these companies use, in particular because a lot of the companies don't like it either.

But it is the easiest and simplest way to shorthand the idea.

Cool. So this next person whose criticism I want you guys to hear is Alison Gopnik. You guys, of course, know this. Alison Gopnik is a very distinguished psychologist at UC Berkeley. She’s a developmental psychologist, so she does a lot of work specifically studying how children learn and then applying that to how AI models might learn and how they can be developed. She’s also one of the leading figures pushing this idea that we’ve talked a little bit about on the show: AI is what she calls a cultural technology.

5. AI As Cultural Technology

I’m Alison Gopnik at the University of California at Berkeley. The common way of thinking about AI, which is reflected in The New York Times’ coverage as well, is to think about AI systems as if they were individual intelligent agents, the way people are. But my colleagues and I think this approach to current AI systems is fundamentally misconceived.

The current large language models and large vision models, for example, are really cultural technologies, like writing or print or internet search itself. What they do is let some group of people access the information that other groups of people have articulated, the same way that print lets us understand and learn from other people. Now, these kinds of cultural technologies are extremely important and can change the world for better or for worse, but they’re very different from superintelligent agents of the sort that people imagine when they think about AI. Thinking about current systems in terms of cultural technology would let us approach them, regulate them, and deal with them in a much more productive way.

Casey, what do you make of this one?

I appreciate the question. If Alison were here, I would ask her how she thinks that thinking about these systems as “cultural technologies” would let us regulate them or think about them differently.

I think there are ways in which we absolutely cover AI as a cultural technology around here. We talk about its increasing use in creative industries like Hollywood and the music industry to create forms of culture, and about the risks that AI poses to the web and all the people who publish on the web. That’s one way that I think about AI as a cultural technology, and I do think that we reflect that on the show.

Now, I do hear in Alison’s question a hint of the stochastic parrots argument, which is that, if I’m understanding right, what I’m hearing is that this technology is essentially just a huge amalgamation of human knowledge, and you can dip in and grab a little piece of it here, a piece of it there. What I think that leaves out are the emergent properties that some of these systems have: the way that they can solve problems that are not in their training data, and the way that they can teach themselves to play games that they have never seen before.

When I look at that technology, I think that does seem like something that is pretty close to an individual intelligent agent. So this is one where I would welcome more conversation with Alison about what she means, but that is my initial response. Kevin?

Yeah. I think these systems are built on the foundation of human knowledge, right? They are trained on all of the text on the internet and lots of intellectual output that humans have produced over the centuries. But I think the analogy starts to break down a little bit when you start thinking about more recent systems.

A printing press, writing, and the internet are technologies that are stable and inert. They can’t form their own goals and pursue them, but an AI agent can. Right now, AI agents are not superintelligent. They’re very brittle, and they don’t really work in a lot of ways. But I think once you give an AI system a goal and the ability to act on its own to meet that goal, it’s not really a passive object anymore. It is an actor in the world.

You can call that a cultural technology, or you can call that an intelligent agent, but I think it’s not just like a printing press, a PC, or another piece of technology that these things are sometimes compared to. I think it’s something new and different when it can actually go out in the world and do things.

Yeah. You think about OpenAI’s Operator, for example. It can book a plane ticket or a hotel room. Is that a cultural technology? I don’t know. That feels like something different to me.

Yeah.

All right.

Mm.

Next up.

Okay, so this next question is about the scientific and medical breakthroughs that could come from AI. This question comes from Ross Douthat, who is an opinion columnist here at The New York Times and the host of the podcast Interesting Times. He’s been interviewing a lot of people connected to the AI world.

6. Scientific Limits Of AI

Hey, guys, it’s your colleague Ross Douthat, and I’m curious about what, if anything, you think limits AI’s ability to predict and understand incredibly complex and chaotic and sometimes one-of-a-kind systems.

To take 2 examples, I’m thinking about, on the one hand, our ability to predict the weather in advance, and, on the other hand, our ability to predict which treatments and drugs will work inside the insane, individualized complexity of a human immune system. Those both seem to me like cases where just throwing more and more raw intelligence or computational power at a problem may run into some inherent limits. We’ll get cancer cures and better weather prediction, but certain things will always remain in the realm of uncertainty or the realm of trial and error.

Do you guys agree, or are you more optimistic about AI’s ability to bring even the most chaotic and complex realms into some kind of understanding?

So there are 2 questions here. One is: Is there some upper bound on how well these systems will be able to predict? To me, the answer is maybe. I don’t know that we’ll ever have an AI system that can predict the weather with 100% certainty.

At the same time, I did a little bit of Googling before we logged on, and AI weather prediction models are really good and getting better all the time. Meteorologists say that their field has rarely felt so exciting because they’re able to make better predictions than they have before.

I think you’re seeing something similar with medicine. We’ve featured stories on the podcast about the way that this is leading to new drug discovery. It is leading to improvements in diagnoses. If you’re looking for reasons to be excited about AI, I would point to stuff like that as obviously useful in people’s lives.

But it’s still not perfect, right? It may be that getting from a very reliable weather forecast to a perfect weather forecast would require some fundamental breakthrough—something in quantum mechanics, some new understanding of how various particles are interacting out in the atmosphere.

But getting way better forecasts might be good enough for most people, and I think the same could be said of medicine. Maybe this is not going to cure every disease on Earth. Maybe there will still be things about the human body we don’t understand. But I do agree with you that people who work in this field are more excited than they’ve been in a long time because they see how much AI allows them to explore and test.

Yeah. Maybe one other question you can add in here, which I think is relevant, is: Are these systems better than a person? Because if they are, then we probably want to use them.

Can I ask, how much of your optimism about AI hinges on AI being able to give us either these scientific or medical breakthroughs?

I think science and medicine are just 2—maybe the 2 most obvious places where this stuff will be good. If you told me that you could cure cancer and many other diseases, I’m personally willing to put up with a lot more social disruption. If it can never do those things, despite all the promises that have been made, then I’ll be super mad. I’ll put a curse on the podcast.

Yeah. Personally, my own AI optimism does not hinge on AI going out there and solving all of the unproved math theorems and curing all of the diseases. I think that even if it were just to speed up the process of discovery—even if all it were doing was accelerating the work that chemists, biomedical researchers, and people looking into climate change were doing—that would be reason enough for optimism.

Mm.

Because so much of what acts as a bottleneck on progress in science and medicine is that it’s really slow and hard. You need to build these wet labs, do a bunch of tests, wait for the tests to come back, and run clinical trials.

I think one of the things that was exciting about our conversation with Patrick Collison at the live show the other day was when he was talking about this virtual cell that they’re building. You can build a virtual environment using AI that can allow you to run these experiments in silico, as they say, rather than needing to go out and test it on a bunch of fruit flies or rats or humans or whatever. You can shorten the feedback loop and take more bites at the apple.

Absolutely. There was a story in Quanta Magazine this week that said AI hasn’t led to any new discoveries in physics just yet, but it is designing new experiments and spotting patterns in data in the way that Kevin was just describing, in ways that physicists are finding really useful.

So I think it's clear that AI is already shortening some of those timelines.

When we come back, we'll hear from more of our critics.

Can I bring my therapist?

You know what's great about this is now, instead of your own internal voice criticizing yourself, you can externalize it and realize that all your fears are true, and people actually are criticizing you all the time behind your back. Yeah.

Exactly. Isn't it really nice?

It's so nice.

What a great idea.

Mm-hmm.

Well, on that note, are you guys ready for the next critic?

Hit me with it.

7. Democratic Control Of AI

My name is Claire Leibowicz, and I lead the AI and Media Integrity Program at the Partnership on AI. I keep coming back to something that I struggle with in my own reaction to your pieces. I found myself nodding when you both critique AI for being biased, persuasive, and sycophantic. But then I start thinking about how humans around me behave, and they do all these things, too.

So I'm wondering, are we ultimately critiquing AI for being too much like us? In which domain should we expect these systems to actually transcend human limitations, and are there others where it may be valuable for them to reflect our true nature? And most importantly, why aren't we spending more time figuring out who is best suited to decide these things and empowering them?

That last question is super important. I'm a big democracy guy, and I want there to be a public role in creating this AI future. I want people who have opinions about this stuff to talk about it online, yes, but also run for office and put together policy proposals, and then get into office and pass laws and regulations.

I got into journalism because I wanted to play my own role in that process of helping to inform people and then hopefully, in some very small way, influencing public policy. So that's my answer to that question.

Yeah, I agree with that. I want people from lots of disciplines to be weighing in on this stuff, not just by posting online and writing op-eds in the newspaper, but by actually getting into the process of designing and building these systems. I want philosophers and ethicists. I want sociologists and anthropologists advising these companies.

I want this to be a global, democratic, multidisciplinary effort to create these systems, and I don't want it to just be a bunch of engineers in San Francisco designing these systems with no input from the outside world.

Absolutely. If a bunch of people listen to the things that we and others talk about and think, “Man, I really don't like this AI stuff at all. I don't want it to replace anyone's job. I want to form a political movement and seek office and try to oppose that,” I think that would be awesome. We need to have that fight in public, and right now, far too few people are participating in that conversation. So I totally agree with that.

Now, let me address the other part of Claire's question, though, which is, are AI systems just a reflection of us? Here's where I think it gets problematic. If you have a human friend, sometimes they're going to be very supportive and nice to you. Sometimes they're going to bust your chops and criticize you. Sometimes they're going to give you really hard feedback and tell you something that you didn't want to hear.

This is not what AI systems do, and so where I get concerned is we're starting to read more stories about young people in particular turning to these chatbots to answer every single question, developing these really intense emotional relationships with them. I am worried that it is not preparing them for a future where they're going to be interacting with people who do not always have their best interests at heart.

Or maybe they could have an amazing relationship with someone, but this person is a little bit prickly, and you need to learn how to navigate them. So that is where I get really concerned: These systems, while they're unreliable in so many ways, are quite reliably sycophantic, and I just think that creates a bunch of issues that humans don't mostly have.

Yeah, and I think what I would add to that is that I don't want AI to mirror all of humanity's values, the positive and the negative. I want it to mirror the best of us, right? The better angels of our nature, as Abraham Lincoln said. I want that to be what these AI companies are striving to design—

As opposed to, say, Mecha Hitler.

Yes. Yes, because that is also a set of values that humans have. So sometimes when I hear people at these AI companies talk about aligning AI systems with human values, I'm like, “Well, which humans?” Because I can think of some pretty bad ones whose values I don't want to see adopted into these systems.

Yeah. Well, that's called woke AI, and it's illegal now.

All right. Rachel, let's hear from someone else.

Okay. This is the very last one. You guys are doing great. This final question comes from friend of the pod Max Read, who, of course, has the newsletter Read Max. I thought his question was really great because he's really interested in how you—

—think about discerning between what's hype and what's not, and how you trust your own instincts and where your confidence comes from. So let's hear Max.

Hi, guys. It's your old friend, Max Read. I was originally going to ask about Kevin's a cappella career in college, but my understanding is that the higher-ups at The New York Times won't allow me to ask such dangerous questions.

You bastard.

So instead, I want to ask you about AI by way of asking you about crypto. You guys were both pretty actively involved in covering the Web3 era, the crypto boom of the pandemic, NFTs, Bored Apes, all this stuff, and very little of that, despite the massive hype around it at the time, has really panned out as promised, at least as far as I can tell.

What I'm wondering is how you guys feel about that hype and about your coverage of that hype from the perspective of 2025. Are there regrets you have? Are there lessons you feel like you've learned? And especially when you look at the current state of AI coverage and hype, not just your own coverage but in general, do you think or worry that it falls prey to any of the same mistakes?

I want to caveat this question by saying the easy mode of this question is to just say the technology is totally different, so it's a very different thing. And I want to put it to you in hard mode, because I don't want to hear about how the tech is different. What I'm interested in is hearing about you guys and your work as journalists.

How do you approach this industry? How do you establish your own credibility? And how do you assess the claims being made by investors and entrepreneurs? Can't wait to hear the answer. Bye.

I love this question. What have I learned? To touch on the crypto piece without touching on the technology, here's what I'll say. Ultimately, what persuaded me in 2021 that crypto was really worth paying attention to was the density of talent that it attracted.

So many people I knew who had previously worked on really valuable companies were quitting their jobs to go build new crypto companies, and what I believed and said out loud at the time was it would just be really surprising if all of those talented people failed to create a lot of really valuable companies. In the end, they did not produce a lot that I did find valuable.

Although, as we've been covering on the show recently, crypto has not gone away, and thanks to the fact that the industry has captured the government, it is now more valuable than ever. So that is what I would say about that time in crypto. And I do think that some of that argument ports over to AI because, certainly, I also know a lot of people who quit their jobs working at social media companies, for example, who are now working on AI.

Here’s what I would say about hype and covering AI. I think that a good podcast about technology needs to do 2 things. One is to give you very grounded coverage of stuff that is happening right now. I’m thinking about recent months, when Pete Wells came on to talk about how chefs are using AI in their restaurants, or Roy Lee came on and talked about the cheating technology he’s building, or Kevin talked about what he’s vibe coding. I even think about the emergency episode that we did about DeepSeek, which I think actually was an effort to unhype the technology a bit while giving you a really grounded sense of what it was and why people were so excited about it.

So that’s one thing I think we need to do. The other thing I think we need to do is just tell you what the industry says is going to happen. I think it is important to get leaders of these companies in the room and just hear their visions because there is some chance that a version of it will come true.

So this is the thing that we’re doing when we bring on a Sam Altman or a Demis Hassabis or the founders of the Mechanized Company. You probably heard in our interview that I was not particularly impressed with that vision. But I think it is useful to the audience to hear what these folks think they are doing. Of course, we want to push back on them a bit, but I have always appreciated journalism that gives airtime to visions and lets me think about them and disagree with them.

That is how I think about hype in general. We want to tell you mostly what is happening on the ground, but we do want to tell you what the CEOs are telling us all the time is going to happen. Then we want you to interrogate the space in between—the space that we actually have to live in.

Yeah. I will say I feel pretty good about the way that I covered crypto back in 2021. There is really only 1 crypto story that I truly regret writing, and that is a story about this crypto company, Helium, that was trying to do this convoluted thing with crypto-powered Wi-Fi routers. I failed on that story. I failed to ask basic journalistic questions.

After the fact, we learned that Helium had basically claimed that it had a bunch of partnerships with a bunch of different companies, and I just didn’t call the companies to say, “Hey, is this company lying about being affiliated with you?” It didn’t occur to me that they would be so blatantly misleading me about the state of their business, and so I do regret that. I would chalk that up less to buying into crypto hype and more to not making a few more calls that would’ve saved me from some grief. The lesson I took from crypto reporting is that real-world use matters.

Mm-hmm.

So much of crypto and the hype around it consisted of abstract ideas, vague promises, and white papers. When you actually dug in and looked at who was using it and what they were using it for, it was criminals, speculators, and people trying to get rich on their Bored Ape collection.

So now, when I cover AI, I really try to talk to civilians using this technology about how they are using it. Whenever possible, I try to use it myself before I form an opinion on it. I think the crypto era was, in some ways, a traumatic incident for the tech journalism community. I think a lot of our peers, and maybe to a certain extent you and I, felt like we were duped, felt like we fell for something, felt like we’d wasted all of our time trying to understand and explain this technology and taking this stuff seriously, only to have it all come crashing down.

I worry that a lot of journalists took the wrong lesson from what happened with crypto. The lesson that I think a lot of journalists took was to be blanket skeptical of all new technologies, to assume that it’s all smoke and mirrors, that everyone is lying to you, and that it’s not really going to be worth your time to dig in and try to understand something. I see a lot of that attitude reflected in some of the AI coverage I see today.

While I take Max’s point that we should always be learning from our mistakes and from things that we swallowed too uncritically in the past, I think that, in some ways, what we’re seeing today with AI is overcorrecting on that point.

Yeah, I think that there is a bit of an overcorrection, but I also think that many journalists have just realized that what used to be a really small industry that mostly concerned itself with helping you print your photos and make a spreadsheet is now something much bigger and more consequential and has just been bad for a lot of people.

So it makes them hesitant to trust someone who comes along and says, “Hey, I’m going to cure all human disease.” I think a role that we both try to occupy in the AI journalism world is to say that we take seriously the CEOs who say they’re building something really powerful, and crucially, we think it will be powerful in bad ways.

Yes.

We want to talk to you about those bad ways, such as you may lose your job, or it will enable new forms of cyberattacks and fraud that you may fall victim to, or it will burn our current education system down to the ground so it has to be rebuilt from scratch. That one, maybe there will be some positive along the way.

But I feel like, week after week on this show, we are trying to show you ways in which this thing is going to be massively disruptive, and that gets framed as hype in a way that I just think is a little bit silly. In 2010, imagine I’d written a story about Facebook and how one day it would have billions of users, undermine democracy, and give a bunch of teenagers eating disorders. Would that have been hype?

Sort of. Would that have been accepting the terms of the social media founders and accepting their language around growth? Yes. But would it have been useful? Would I be proud that I wrote that story? I think so.

So I’m willing to accept the idea that you and I do buy into the vision of very powerful AI more than many of our peers in tech journalism. But the reason that we’re doing that is we want to remind you what happened the last time one of these technologies grew really quickly, got into everyone’s hands, and became the way that people interface with the digital world. It didn’t go great.

We already know that these companies are not going to be regulated in any meaningful way. The AI Action Plan is designed basically to ensure that. So, to the extent that we can play a positive role, I think it is just going to be in talking to people about those consequences.

If the consequence of that is that people say we’re on the side of hype, I will just accept the criticism.

Yeah.

Hmm. Well, thank you guys so much for doing this, and thank you also to our critics for taking the time to talk to me. I thought we could end by talking about whether you guys have any questions for each other. One of the big goals of this is to map where you guys stand relative to other thinkers, so I’m curious how your views on AI are actually different from each other.

I think I have longer timelines than Kevin does. I think Kevin talks about AGI in a way that makes it seem very imminent, and I think I’m more confident that it’s going to take several years—maybe more than several. Maybe this is a 5-to-10-year, or even 15-year, project. So I think that’s the main way that I notice disagreeing with Kevin.

I think that we also disagree about regulation and how possible or advisable it is to have the government step in and try to control the development and deployment of AI systems.

I think that you are informed by your years of covering social media and seeing regulators grapple with and mostly fail to regulate that wave of technology. But I think you are also a person who has a lot of hope and optimism about institutions and wants there to be democratic accountability for powerful technology. I share that view, but I also don’t think there’s a chance in hell that our present government, constructed the way it is and with the pace at which it is used to regulating things, can regulate AI on anything approaching a relevant timescale. I’ve become fairly pessimistic about the possibility of meaningful regulation of AI, and I think that’s a place where we differ.

I think we do disagree there because I think that we had the makings of meaningful regulation under the Biden administration, where they were making very simple demands, like, you need to inform us when you’re training a model of a certain size. There need to be other transparency requirements. And I think you can get from there to a better world.

Instead, we’ve sort of unwound all the way back to, “Hey, if you want to create the largest and most powerful model in the world, you can do that. You don’t have to tell anybody if it creates new risk for bioweapons and other risks. You don’t have to tell anybody. You can put it out in the world.”

Right now, there are many big AI labs that are racing to get the most powerful AI they can into everyone’s hands with absolutely no safeguards. So if you’re telling me that we can’t create a better world than that, I’m going to disagree with you.

Yeah.

Mm.

Go fuck yourself.

Well, thank God you guys disagree because it makes the podcast more interesting. And thank you guys, seriously, for doing this. I think, given how much of the AI conversation can feel really disempowering in this moment, one thing that gives me a feeling of a little bit more control is really trying to map out the debates—where people stand relative to each other—because it ultimately helps me figure out what I think about AI, where I think the future is going, and that's at least one thing I feel sort of empowered to do.

And that's what we want to do. Truly, we want everyone to come to their own understanding of where they sit at the various intersections of these discourses. I think Kevin and I identify as reporters first. We don't have all the answers. That's why we usually bring on a guest every week to try to get smarter about some subject, right?

So I think a really bad outcome for the podcast is that people think of us as pundits. I think of us as curious people with informed points of view, but we always try to be open to changing our minds.

Yes. Like a large language model, we aim to improve from version to version.

As we add new parameters—

Yeah.

And computing power.

Yes.

Yeah.

Before we go, a reminder that we are still soliciting stories from students about how AI is playing out on the ground in schools, colleges, universities around the country. We want to hear from you. Send us a voice memo telling us what effect AI is having in your school, and we may use it in our upcoming Back to School AI episode. You can send that to hardfork@nytimes.com.