Grok’s Undressing Scandal + Claude Code Capers + Casey Busts a Reddit Hoax

“This is a story about how a tool can be used to try to affect politics and in particular to minimize women, denigrate them and push them out of the conversation.”

• 0:00Speaker 00

  Support for this podcast comes from Dell Technologies. It's time to put AI to work. Is your business ready? The challenge isn't getting started with AI. It's scaling beyond pilots while maintaining control of your data, governance, infrastructure, and costs. Dell AI Factory provides the integrated foundation you need to scale AI securely and efficiently with one trusted partner, Dell, AI on your terms. From idea to impact. Learn more at dell.com/yourwaytoAI. That's dell.com/yourwaytoAI.

• 0:34Speaker 01

  Kevin, I have to ask you, are you all cut up on Heated Rivalry?

• 0:38Speaker 02

  I have no idea what you're talking about.

• 0:39Speaker 01

  Kevin! Could you be an ally to the gay community for 10 minutes? This is the show that has the LGBT community in a chokehold right now, and also a lot of straight gals. This is about two closeted hockey players who fall in love. And when I tell you this thing is a sensation on social media right now, it is truly, it's a six-episode show, okay? If you watch it, it's a beautiful Canadian show. It looks like it was made for about $100. And there is just something about the story. They picked great actors. And we watched all six episodes over the weekend. And my boyfriend, who normally does not watch TV, immediately began to re-watch it.

• 1:21Speaker 02

  Wow.

• 1:21Speaker 01

  Like, we finished episode six, and then we looped around to episode one.

• 1:25Speaker 02

  Wow.

• 1:26Speaker 01

  So why, would you care about this? I don't know. But that's what's going on in my life.

• 1:32Speaker 02

  I'm impressed. It's hard, it's not easy to make a show about hockey that is interesting to people.

• 1:37Speaker 01

  It's so true. And it turns out that what it needed was a lot of gay sex. And I would like to see this template applied to a lot of other sports and activities, and just see if we can't make a good show out of it.

• 1:50Speaker 02

  Bowling.

• 1:51Speaker 01

  Bowling.

• 1:52Speaker 02

  Shuffleboard.

• 1:53Speaker 01

  The gay bowling version of Heated Rivalry, I think, would destroy it. You could call it, what would you call it? The 300 Game, or the, um, In the Gutter? That's what I would call the-

• 2:05Speaker 02

  Seven, 10 split.

• 2:08Speaker 02

  In the gutter is better.

• 2:10Speaker 01

  In the gutter's a little.

• 2:11Speaker 02

  Yeah.

• 2:11Speaker 01

  It's like, like, not that Heated Rivalry is sleazy, that, I mean, actually, but maybe that the bowling version would be sleazier.

• 2:17Speaker 02

  Yeah.

• 2:17Speaker 01

  That'd be the kind of the fun twist.

• 2:18Speaker 02

  Yeah.

• 2:19Speaker 01

  Anyways, Heated Rivalry. Go check it out on HBO Go Max Plus.

• 2:29Speaker 02

  I'm Kevin Druse, a tech columnist for The New York Times.

• 2:31Speaker 01

  I'm Casey Newton from Platformer.

• 2:33Speaker 02

  And this is Hard Fork.

• 2:34Speaker 01

  This week, Grok gets caught with its pants down. Can anyone stop its viral bikini image generator? Then, we're vibe coding again. Kevin and I compare notes on what we're building with Claude Code. And finally, it's a Reddit mystery. How a scammer tried to fool us all using AI-generated evidence and how I cracked the case.

• 2:55Speaker 02

  Da da.

• 3:05Speaker 02

  Well, Casey, happy 2026. It's good to be back. And we would like to start the year by talking about something that happened over the break, which is that there's been a big scandal brewing over at X.

• 3:16Speaker 01

  Boy, has there been. I imagine many of our listeners have seen by this point that X is in a lot of trouble because of the way that its Grok chatbot has been generating images of celebrities, women, children that are highly sexual and, for the most part, has been declining to remove them or even really comment on what has been going on.

• 3:39Speaker 02

  Yeah, so I started seeing this over the break. Something happened where, you know, Grok, which I think people on X had been using up to that point mostly to kind of settle arguments and fact-check other people. All of a sudden, I started seeing people using Grok to like, undress photos of mostly women. You know, "Grok, put me in a bikini." "Grok, put this politician in a revealing lingerie set." "Grok, take off this person's pants." Like, it just seemed like this started to happen pretty much overnight in a way that was really sort of troubling and unchecked, as you said. So can you just help me understand, like, what actually changed here? Was there some setting that was changed? Was there some new model that was released that allowed people to do this?

• 4:21Speaker 01

  So, we don't have a lot of good answers to those questions. I can tell you that this trend of what is sometimes called "nudifying" really takes off in 2023 as these image generators start to get better because they find that you can make a lot of money. Money do it. There are a lot of men in particular who will just spend tons and tons of money on making these sort of non-consensual images, as you say, mostly of women. When it comes to Grok, they had previously said that they'd licensed an image generator from a company called Black Forest Labs, and they were using that as sort of from the moment that they started using image generation. But in December 2024, they said that they were using their own image generator, which is called Aurora.

• 5:05Speaker 01

  And while, again, we don't have great details, there is at least plenty of anecdotes online that over the past several months, let's say, the guardrails around creating nudity and sexual imagery appear to have been relaxed. And so there is now at least one thriving subreddit that is devoted solely to making porn on Grok.

• 5:25Speaker 02

  Hmm. So, as you said, these nudefy apps have been around for several years now, but they've kind of been hard to access. Some of them relied on these open source models. You had to kind of know how to use them, or they were sort of kicked out of the app stores for being against their policies. But what seems notable about this to me, at least, is that it's all happening not only on a major social network, but in public. Like, people are literally doing this in the replies of their posts on X. So is that what's new to you about this? The sort of the publicness of it?

• 6:01Speaker 01

  Absolutely. Like, it is upsetting enough if some man takes an image of a woman and creates a naked version of her without her consent. Like, that is a bad thing. What is so shocking about this is that you can just see it happening in real time. Like, several outlets have just been going into the Grok account, and they're seeing it making hundreds and thousands of images in response to user requests, and anyone can go in and view them. And of course, that is most upsetting to the victims of what I am going to call attacks, because you still have normal people using X to do things like posting a photo of me, like, out on a hike or whatever, and then some freak shows up in your mentions and say, "Hey, put her..."

• 6:46Speaker 01

  In a bikini and then it does and then you as the victim are looking at that in your replies.

• 6:52Speaker 02

  That's crazy. So obviously X and Elon Musk are not outraged about this as many users are. They seem to sort of think the whole thing is a joke, but my question is like, are there guardrails in place? Are people sort of jailbreaking Grok to make it do this? Or is this like literally a sort of mainline advertised feature of the Grok app?

• 7:14Speaker 01

  It's a great question and no, they are not jailbreaking Grok to do this. They are just sending replies on X saying at Grok do this and then it is doing this. And they are not having to get particularly cute with their language. They are just literally asking to see images of women in bikinis.

• 7:30Speaker 02

  That's so wild to me.

• 7:31Speaker 01

  Or children in bikinis.

• 7:33Speaker 02

  I mean, one question I have about this is just like, how are Apple and Google and their app stores okay with this?

• 7:40Speaker 01

  Bro, thank you. Last year I was writing about the introduction of Annie, the sexual companion bot that they got put into Grok. And I went on to the iOS store and noticed that Grok was rated for children 12 and older. And I thought that seems like pretty young to be, you know, giving access to a sex bot. So I sent a message to Apple and I said, what's going on? And the message I got from the POMS team was like, uh, like, well, we're looking into it. And my, my honest sense at the time was that they're, they're going to make a change here. Like it's obvious that, that, you know, they're just going to have to like do something. Well, they didn't make any change. And then I went on this week, Kevin, in the wake of this new, you know, CSAM nudifying scandal. And I found that Apple has changed the rating for Grok and it is now rated for children 13 and older.

• 8:24Speaker 02

  Are you kidding me?

• 8:25Speaker 01

  Yeah. So sorry to all the 12 year olds out there that were sort of having a free for all. You're going to have to wait till your next birthday before you can use this thing again.

• 8:32Speaker 02

  That is genuinely shocking to me.

• 8:34Speaker 01

  Yeah, me too.

• 8:35Speaker 02

  And I feel like it is one of the clearest cases I've ever heard of a blatant double standard on the part of the app stores.

• 8:44Speaker 01

  Absolutely.

• 8:45Speaker 02

  If a random startup showed up one day and said, Apple, I'd like to start selling my bikini fi app in your app store. I think they would shut it down. There's no way.

• 8:56Speaker 01

  Absolutely.

• 8:57Speaker 02

  They have policies against that kind of thing. But because it's X, because it's Elon Musk, because this app already has millions of users, maybe they feel less inclined to take action against it. I don't know. Do you have any insight into what's going on over there?

• 9:10Speaker 01

  I don't have any insight, but I would be very confident that there are people inside Apple that have said, we should change this rating. And there is someone who is sitting there saying, no, if we do that, the vice president is going to tweet about it. And Marco Rubio is going to tweet about it. And Brendan Carr is going to launch an investigation against us for punishing X, right? And censoring them. And so they're just in this state of paralysis because they're so terrified of standing up for the principle that women and children should not be attacked online.

• 9:38Speaker 02

  Yeah, so there are some investigations going on right now. But like, do you expect this to be stopped at some point by any regulator anywhere?

• 9:47Speaker 01

  So my instinct from history is like, yes, absolutely. Like, it is going to be stopped. France has called the sexual content clearly illegal. The UK government said that it is considering an investigation. The European Union said that it is very seriously looking into these complaints about Grok. India's IT ministry has demanded that X do something here. So I have to imagine that something is going to come out of all of that that is going to result in some sort of change. At the same time, do I think that the United States is going to intervene? Probably not. Elon posted a photo from over the break of himself having dinner with the president, right? So it seems like they're friends again, and that there just is not going to be any pushback here in X's home country.

• 10:25Speaker 02

  Yeah. Well, I think that's a good overview of what's been happening and why people are talking about this. But we wanted to bring in our colleague Kate Conger. She is a reporter for The Times, a friend of the pod. She's been on before. And she has been reporting on this Grok scandal this week and has actually been talking with some of the victims, some of the women who have been attacked by these Grok deep fakes. And I think we should hear their perspective as well.

• 10:52Speaker 01

  Absolutely.

• 10:56Speaker 02

  Kay Conger, welcome back to Hard Fork.

• 10:58Speaker 03

  Thank you, Casey.

• 10:59Speaker 02

  So tell us a bit about the conversations you've been having with some of the victims of this behavior on Grok.

• 11:05Speaker 03

  Yeah, so I think one of the big struggles for people whose images are being used in this way is how to respond and what to do about it. You know, people are obviously reaching out to X, trying to get these things removed, and that process is taking sometimes a long time, sometimes it's not happening at all. And I think part of that is, as you know, we've talked about a lot, I think on the show in particular, that X has gotten rid of a lot of their content moderation folks, and they don't have large teams of people who are responding to this. You know, I've been speaking with some people who are working with children who have been deep faked on X recently, and they are getting those images taken down, but it's taking sometimes 36, 72 hours.

• 11:48Speaker 03

  And these images are kind of sitting up and being commented on and exploited for quite some time. So, you know, it's a bit of a scary moment, I think, for people trying to figure out how to respond once they are put in this situation by Grok.

• 12:00Speaker 02

  Can you just give us a sense of like, why this is happening and what it's like for the people on the other side of it?

• 12:05Speaker 03

  Why this is happening really ranges, and I've seen this happening to all kinds of different women. You know, women who are just sort of regular X users. Obviously, women with public platforms are being targeted by this, so political figures, Twitch streamers, you know, celebrities, actresses, they are all being kind of roped into this. And I think the motivations of the people who are asking for these images, you know, obviously they're sexual, but it kind of runs the gamut from, you know, wanting to create pornographic images of someone to, I think, wanting to just humiliate women in particular and sort of bully them, you know, tagging them in these images and really

• 12:49Speaker 03

  kind of trying to provoke a reaction from the women that they're making deep fakes of.

• 12:53Speaker 02

  Can you give us just one specific example of someone that you've talked to and maybe tell us the story of their experience with this?

• 13:01Speaker 03

  Yeah, so you know, I mentioned I've been speaking to some folks who are working with children who've been deep faked on Grok. And so I want to be kind of, you know, vague about the specifics of it because I don't want to further bring harassment to these kids. But you know, there's a particular child who I'm thinking of who's been deep faked several times. She's a somewhat public figure. And so that's where the the original images are coming from, where people are asking for her clothing to be removed. And it's been pretty scary for her parents. She knows what's happening, but isn't seeing it. But her parents are monitoring her social media and seeing these images pop up, which I think is really frightening for them. And then, you know, reaching out to Twitter

• 13:46Speaker 03

  and into advocacy groups, trying to get these images removed and, you know, just being really frustrated by the amount of time that it's taking. And I think being really outraged as well that, you know, someone can go online and request a nude image of their 14 year old and that this technology will comply with it in a really public fashion.

• 14:02Speaker 01

  I mean, I don't know if this question is too obvious to ask, but I wonder what you can tell us about how it feels for these victims to just post what they think of as an innocuous image on X, and then to go back on the site and see that it's, you know, now being turned into porn against their will.

• 14:21Speaker 03

  Yeah, I think for some people, it feels, you know, really angering. I'm hearing a lot of anger from folks and also embarrassing. I've talked to some women who questioned whether to do anything about it or say anything about the fact that it happened to them because they feel embarrassed by it. And they don't necessarily want to draw more attention to the fact that these images of them now exist on the internet.

• 14:47Speaker 01

  Can you also talk about the way that men will do this to women as a way of like bullying them out of the town square and getting them to sort of stop talking and stop participating in public life?

• 15:02Speaker 03

  Yeah, you know, I think about this in particular with some of the politicians who I'm seeing being deep faked, you know, it's people who, you know, female politicians who the users of X maybe don't agree with their views and are taking images of them from, you know, their public work and advocacy and requesting for those to be made nude or to depict them in a bikini or whatever the case may be. And so I think there's a real, real obvious effort there to embarrass and to kind of warp these images of someone in their professional life into a really personal and intimate image.

• 15:36Speaker 02

  I've seen this happening, for example, with like AOC.

• 15:38Speaker 03

  Yes.

• 15:39Speaker 02

  Like now every time, you know, there's a photo of her, people will reply to it and say, @Grock, you know, put her in a revealing, you know, halter top or something.

• 15:47Speaker 01

  Which I think speaks to the fact that like, this is not just a story about porn. This is a story about how a tool can be used to try to affect politics and like in particular to minimize women and to, you know, denigrate them and like, and push them out of the conversation.

• 16:06Speaker 02

  Yeah. Kate, I'm curious, like, obviously, we all know about the Mecca Hitler incident from last year where Grok started spewing these like antisemitic responses.

• 16:17Speaker 01

  In retrospect, a much safer version of Grok, but yes.

• 16:20Speaker 02

  Yes. We pine for the days when Mecca Hitler-

• 16:22Speaker 01

  Bring back Mecca Hitler!

• 16:23Speaker 02

  -was the worst thing that Grok could do. But in that sense, there was a sense of that this was an inadvertent sort of bypassing of some safety filter. Something had gone wrong in the programming of Grok that caused it to behave this way. To me, this new image generation behavior feels much less accidental. It feels like there, maybe there was even a meeting about it. This was obviously part of some plan, or at least that once users started using the technology in this way, the company did not take immediate steps to clamp down on it the way they did with Mecca Hitler. So can you tell us anything from your reporting about how this happened inside the company, who is making the decisions about this kind of thing? Is this part of some sort of, you know, demented growth strategy

• 17:07Speaker 02

  over there? Or like, why is this happening now?

• 17:10Speaker 03

  Yeah, so first of all, this is not, I think, the beginning of Grok generating these kinds of images. As I've been tracing back through the images that Grok has been posting, finding images like this of women going back to June and July of last year. So I think this has been going on in sort of a lower volume for quite some time, and it really escalated over the holidays and with people kind of making it into a trend on X. But, you know, in our reporting about the Mecca Hitler incident, what we found was that Elon Musk had given a directive to the folks working on Grok that he wanted it to go viral. He wanted it to be edgier, sort of as a strategy to promote the tool and to get it onto people's radars. And the company perspective was that, obviously with Mecca Hitler went too far,

• 17:54Speaker 03

  they turned Grok off for a couple of days. But it has been part of the strategy for Grok to try to create these viral moments. And I think to Casey's point earlier about the fact that, you know, there is this sort of silencing of women that's going on here, Grok, I want to be clear, is not the only AI tool that makes deep fake porn of women. It's far and away not the only tool. But it is the only tool that is doing that in an inherently public fashion on social media where these images can instantly spread and go viral, right? A lot of these other image generation tools are happening in a private chat with a user. And Grok is not that. This is very, very public, and it's sort of intended to be that way and intended to drive traffic and interest. And that's the reaction that

• 18:39Speaker 03

  we've seen from leaders at X, right? So, you know, as this trend was picking up, Musk posted something about putting a SpaceX rocket in a bikini, kind of mocking and laughing at the fact that this trend is happening. We've seen X's head of product post about the fact that engagement during the time period that this was happening on X was higher than ever. So for them, you know, they're seeing the engagement that they're looking for. They're seeing their sponsors, seeing people tapping into their X feeds and tapping into Grok, and that's being viewed as a positive.

• 19:10Speaker 01

  So we talked a little bit before you came on, Kate, about the reactions that some other countries have had, some investigations that are now underway. I want to talk about what's happening in the United States. Congress did recently pass the Take It Down Act. My understanding is that that's going to go into effect in May of this year. What is it going to require platforms to do? And do you think that it is going to be enforced in a way that is going to help the victims on Grok right now?

• 19:39Speaker 03

  Right. So Take It Down, the provision that's happening in May is that's the deadline for companies to set up a process for victims to request this kind of imagery to be removed and to face penalties if they do not remove it. So basically, that would mean, in this case, sort of an enhancement of what X already has available for people to request that copyrighted images be removed, for instance, or to request that harassment or abuse be removed. So it's basically just asking these social media companies to build a framework for victims to go in and request takedowns. But it's not putting any legal pressure on X to not allow these images to be created to begin with.

• 20:19Speaker 01

  Right. And of course, May is still four months away, and that's going to be cold comfort to some of the victims here, I suspect.

• 20:28Speaker 02

  Is there any difference, Kate? Or maybe, Casey, you also know the answer to this. But like, my understanding is that the laws on the books now protect minors having this kind of thing done to them, and maybe that's why they are being successful, even in a delayed way, in getting these images taken down. But then adults in most states or most jurisdictions have very little recourse. Is that true?

• 20:52Speaker 01

  Yes, that is true. It is illegal to produce and to possess CSAM. And so X is under more legal pressure, and I suspect that is why you're seeing them take at least some action here where they're not with adult women.

• 21:05Speaker 03

  Yeah. And Take It Down will be available to adult women, so they will be able to use that process to ask for images to be removed. But Casey's correct that there isn't an automatic legal pressure on X to not create these images of adults to begin with.

• 21:21Speaker 01

  I have to say, on the whole, I find the response to this thing rather muted, given the stakes of what is happening. Like, I have covered so many backlashers against social media companies. Do you remember Cambridge Analytica? Yes. Do you remember how bad people got at the idea that maybe a quiz that they had taken had resulted in some data being given to a polling firm that tried to influence the election? We almost shut the country down over that one. Now you have a website that is just taking girls' clothes off in public on demand, and it's being permitted by the website owner, who is laughing about it in his own feed, and we're saying, eh, what are you gonna do? That's Elon. Didn't get where he is by playing by the rules. So,

• 22:04Speaker 01

  I just truly, like, I'm losing my mind, because I cannot believe we have gotten to this low, low point in the history of content moderation, that this is resulting in so many shrugs by the average person.

• 22:18Speaker 03

  It is really interesting, and, you know, I think it goes to show just how far Elon Musk has been able to push the overton window on content moderation, you know, and, obviously, we've seen Facebook, all these other companies kind of copy his approach to roll back content moderation, roll back rules against harassment, and go with a community moderated approach. And so, I mean, I think you're right. It's wild to see how much this has changed in a relatively short period of time.

• 22:44Speaker 01

  You know, there was a time when something like this would result in this, like, exodus of users away from X, you know, and we'd see, like, Blue Sky get a big surge of users, Threads get a big surge of users. That's not happening this time. Like, I guess, like, people who care about this stuff are all already gone, but like, I don't know. We're just in such a crazy, crazy time in the history of content moderation and social platforms.

• 23:06Speaker 03

  And I do wonder if part of this sort of muted response to this has to do with the fact that this was all really starting to blow up in the time between Christmas and New Year's, and maybe, I don't know, maybe a lot of government officials weren't on their phones, and they were hanging out with their families? I have no idea. But we're just starting to see regulators around the world issuing responses this week and saying that they want to open investigations, or they've sent letters to X. But it has been a really slow response, I think, especially in comparison to other content moderation controversies.

• 23:37Speaker 02

  Yeah, I think honestly, I think this will be one of those situations where unless it happens to the individual politician, or maybe their daughter, or their partner, that is not going to rise to the level of a crisis. And then immediately, as soon as that does start to happening, we're going to start having hearings about it.

• 23:54Speaker 01

  Yeah.

• 23:55Speaker 02

  But maybe that is sort of a pre 2026 mindset here. I also wonder if there's a way in which this is strategic for X. You know, they are the only explicitly right-wing-aligned social media platform that has served them quite well in dealing with the Trump administration. And there's-

• 24:15Speaker 03

  Kevin, I think you're forgetting Truth Social.

• 24:16Speaker 02

  That's true. I am forgetting Truth Social. I'm always doing that. And, you know, this kind of vice signaling does gross me out, but it may actually be effective in getting what they want, which is like, you know, cooperation with and forbearance from the Trump administration. And maybe it's unlikely that they'll see any consequences for this, at least during this administration.

• 24:37Speaker 01

  Well, I think what it gets them is engagement. People love porn, right? Like, that's not been a secret. Like, everyone has known forever that people like to look and click on porn. It's just that platforms have generally seen a lot of risks to their reputation and their larger businesses for embracing it. I think what's new about X is it's saying like, no, like, let's make porn a pillar of what we do, and see if that can help us keep up with the other frontier labs.

• 24:59Speaker 02

  Yeah, I mean, to me, it's just, I don't get the business case for it, because I also thought that Grok was trying to become like a full-fledged competitor to ChatGPT and Claude and Gemini, and like, trying to get enterprise contracts and government contracts. And like, if you're a Fortune 500 company, like, why would you strike a deal with Grok at this point?

• 25:19Speaker 01

  Truly, the funniest thing was they announced the Grok enterprise product on like, the same day that this story was blowing up. And I just love to imagine the head of enterprise sales at Grok, who's like, heading into his meeting with like, the, you know, so this Fortune 500 company is saying, we'd love to sell you, you know, 1,000 seats of Grok to help you get business done. And they're like, well, now what's all this about the deep fake nudes? Although, of course, Grok does have a contract with the US military.

• 25:46Speaker 02

  Yeah.

• 25:46Speaker 01

  Yeah.

• 25:47Speaker 03

  Yeah. Well, and I think, you know, what has been really interesting to see in our reporting on Grok is that some of the differences between what Grok, the chatbot, is doing versus what Grok, the Twitter account, or the X account, I should say, is doing, they're very different. There's more extreme political views being espoused by the X account. There are these deepfakes being generated by the X account. And often, you can take the same query, put it into the actual, like, web browser version of the chatbot, or into the app of the chatbot, and get a much more muted response. And so I think what they're doing is actually, you know, kind of having a more enterprise-friendly chatbot, which is the one that they're licensing and selling, and then this outrage bait machine that exists on X and is getting engagement

• 26:32Speaker 03

  and clicks and bringing people to the social site.

• 26:35Speaker 02

  Yeah. I mean, one other question I have about this is whether, in the absence of regulatory intervention, like, maybe there's some way for the courts to get involved here? I mean, we spent the last decade arguing and talking about Section 230, and whether it shields platforms from liability for the things that happen on the social networks. And I think, in most cases, it does. But this, to me, feels different, because it's it's not a user generating these sexualized images of people without their consent. It is literally the platform itself, or the AI chatbot and system attached to the platform. Does that open up any new forms of legal liability for Grok or X?

• 27:18Speaker 01

  Absolutely. You know, I read an interview with a lawyer in Bloomberg today that basically said exactly that, that they cannot hide behind Section 230 to get out of this. Like, ultimately, it is their product that is creating these images. And so I do suspect that we will see efforts to hold X legally liable for some of the images that they're creating.

• 27:37Speaker 03

  Mm-hmm. And I think X has really shoved the responsibility off onto its users in the cases of AI-generated images featuring children. They put up a post on their safety account, which is sort of the mouthpiece for any kinds of safety issues on the platform. They said, "We take action against illegal content on X, including child sexual abuse material by removing it, permanently suspending accounts, and working with local governments and law enforcement as necessary." They go on to say, "Anyone using or prompting Grok to make illegal content will suffer the same consequences as if they upload illegal content." So this is interesting, right? They're saying that users who request illegal material from Grok will be suspended and reported to law enforcement.

• 28:20Speaker 03

  But those users who are requesting the content aren't actually the ones who are posting the content. It's the Grok account that is creating these images, posting these images online. And so I think if they're being really true to their policies and saying, "We're going to suspend accounts that post this," it's the Grok account that's posting it. It's just been the

• 28:36Speaker 01

  Honestly, that would solve so many problems if they would just delete the Grok account. And I hope if one thing comes out of this, it's that.

• 28:45Speaker 02

  All right. Well, Kate, thank you for this very depressing update from the front lines of content moderation and Godspeed.

• 28:52Speaker 03

  Thanks, guys.

• 28:55Speaker 01

  When we come back, we're vibe coding again. So help me, Claude.

• 29:08Speaker 00

  Support for this podcast comes from Dell Technologies. It's time to put AI to work. Is your business ready? The challenge isn't getting started with AI. It's scaling beyond pilots while maintaining control of your data, governance, infrastructure, and costs. Dell AI Factory provides the integrated foundation you need to scale AI securely and efficiently with one trusted partner. Dell, AI on your terms. From idea to impact. Learn more at dell.com/yourwaytoAI. That's dell.com/yourwaytoAI.

• 29:42Speaker 04

  Modern enterprise is made up of a lot of moving parts, and Comcast Business helps you orchestrate it all. With SD-WAN working at scale to keep 150 hospital locations connected and working as one, plus SASE and Zero Trust Security protecting financial data across a bank's 2,000 branches, and AI-powered networking that optimizes traffic across five continents. No one does business like Comcast Business.

• 30:11Speaker 05

  Companies are pouring money into AI, but few leaders can explain how those AI investments work with the systems you already have. You know, the ones that actually run your business? Without Systems, you don't have to start over with AI. You can choose the projects where AI will have the most impact, and OutSystems gives you the platform to make it agentic on your terms with complete control every step of the way. So you can trust AI with the mission-critical systems that run your business. Own your agentic future. Learn more at outsystems.com/NYT. That's outsystems.com/NYT.

• 30:48Speaker 02

  Well, Casey, since we got back from our holiday break, I have been dying to talk to you about our latest vibe coding experiments. It seems like vibe coding, which, you know, we talked about last year, had a moment over the break.

• 31:02Speaker 01

  Kevin, it is time to build. We have been sort of vibe coding all year, but I must agree with you, developments over the past several weeks have made it easier than ever for somebody who is a total ignoramus when it comes to coding to make some pretty cool stuff.

• 31:18Speaker 02

  Yeah, so I had this experience over the break of like mostly being offline for the first week of our break and trying very hard not to look at social media. I took all my social media apps off my phone. When I came back, the president of Venezuela had been captured, and everyone was talking about Claude Code on my feeds.

• 31:35Speaker 01

  Now, were those two things related?

• 31:37Speaker 02

  I don't think so. But one thing that appears to have happened is that Claude Code, which is the sort of autonomous coding agent made by Anthropic that kind of puts their Claude chatbot inside the terminal window on your computer and lets it do things autonomously, had gotten much better for reasons that I assume are really Related to Opus 4.5, the model that we talked about a couple of weeks on the show. But I came back and saw comments from people like Andre Karpathy, you know, well-known AI researcher, who said that after playing with Claude Code and similar tools, quote, "I've never felt this much behind as a programmer." This is someone who's probably a, you know, top 0.1% programmer in the world saying this.

• 32:20Speaker 02

  An engineer at Google, John Adogan, wrote that she had been trying to build distributed agent orchestrators at Google since last year. And she gave Claude Code a description of the problem, and it generated what they built with a team of Google people last year in an hour. I saw similar sort of hype and praise from many different corners of the internet, including people like us, who are not professional programmers, but were starting to experiment with Claude Code and seeing how much it could do and being pretty amazed by it.

• 32:52Speaker 01

  Yeah, it truly had a moment over the break and was enough, I think, to get both of us to go back to our vibe coding terminals and see what we could build. Actually, and here's a great point for me to make a disclosure. My boyfriend works at Anthropic, and I happened to be away for a few days. And when I came home, my boyfriend had taken it upon himself to install Claude Code on my laptop. And he looked at me and said, "It's time to build, and we're going to make some things." So it was like something out of a Norman Rockwell painting, just the two of us sitting next to the Yuletide log, trying to vibe code.

• 33:24Speaker 02

  Yes, I also spent several hours, not with your boyfriend, but coding on Claude Code during my family vacation, much to the annoyance of my wife. But I did make some things that she actually thought were pretty cool. So we should talk about our experiments. I want to hear about what you've been building. I want to talk about what I've been building. But before we do that, like, why is this happening right now? Why is this the moment that vibe coding has returned to the mainstream?

• 33:49Speaker 01

  So I think the way I want to answer that question is by taking us back about a year ago, since the last time we did a segment about vibe coding. The truth is, I can really only remember one thing that we tried to vibe code at that time. That was the hot tub maintenance app, right? I had just got a hot tub. I was trying to figure out how to keep all the various chemicals in balance. Kevin was kind enough to try to use Claude. Claude Code didn't exist to try to make an app and we used it and it was like sort of okay. But honestly, I didn't use it very much because it didn't work that well. I had a similar results when I tried experiments at Claude. I was able to make some stuff that sort of worked, but there was absolutely nothing where I said, oh, this is an actually useful thing that I'm going to use continuously in my life.

• 34:30Speaker 01

  So while I was totally willing to believe that other people were, you know, having success vibe coding their own projects. If you tried to do much of this stuff a year ago, you just couldn't get that far.

• 34:40Speaker 02

  Yes, and part of that is because it was actually a little clunky to do vibe coding a year ago. Like I remember building that hot tub maintenance app and it required a lot of like copying and pasting. And like if I got an error message, Claude didn't always know like sort of how to handle that. But what has happened in the past year with Claude Code, and we should also mention like, you know, OpenAI and Google have similar tools now, but they have basically merged this into the sort of terminal app on your computer. So it's not, there's no more copying and pasting. Claude can just take instructions in plain English and go off and accomplish various tasks. And it will check in with you, but it can now do all of the orchestration and the execution itself.

• 35:23Speaker 01

  Yeah, that's right. And I think that's important to say sort of at the top of this segment. Because honestly, over the past year, when people have talked about vibe coding, my eyes and ears have kind of glazed over because I always think, well, I'm not a software engineer. I don't know how to build stuff. I'm glad that you're able to make your little custom Chrome extension or whatever. That has no relevance to my life. We're going to talk about the stuff that we built today in the context of code. But really, this is just all about like building digital tools. This is about like building things on your computer and code is the foundation. But I think we're just now starting to move into a world where like if you have an idea and and software is a part of it, you may be able to build it in a way that you haven't before.

• 36:01Speaker 02

  Yes. Okay. So Casey, tell me about what you built over the break with Claude Code.

• 36:05Speaker 01

  So I want to talk about something I made. It is something I've wanted for a really long time. I'm super happy with the result and I'm just going to keep iterating on this thing because it is so much fun. I am truly having so much fun just tinkering. So if Kevin, you would, please go to your browser and enter into the URL bar, cnewton.org.

• 36:30Speaker 01

  Now, for, I don't know, 15-plus years, I have had a personal webpage. It's basically just like a business card. You know, it's like, here's my name, here's like a link to my website. I got it on Squarespace. I paid them $200 a year. I now feel extremely silly that I've been paying them all of that money because using Claude Code, I was able to make truly the personal website of my dreams. Can I give you a tour of this thing?

• 36:57Speaker 02

  Please.

• 36:58Speaker 01

  So, and by the way, listeners, I mean, stop what you're doing. Pull the car over, stop, put the laundry down, and get out your smartphone right now. Go to cnewton.org. Another thing I love about this website, by the way, it's fully responsive. Like, I didn't have to build a separate version for the mobile phone. It just sort of expands and contracts depending on how big your browser window is. So Claude came up with the design. It's sort of very dark, it uses gradients, there's all kinds of like cool fonts. There's also some fun Easter eggs. Like Kevin, do you see how my face is right there at the top of the website?

• 37:25Speaker 02

  Yes.

• 37:26Speaker 01

  Go ahead and I click on my face.

• 37:28Speaker 02

  Oh wow, you sort of like do a little jiggle.

• 37:31Speaker 01

  There's like-

• 37:31Speaker 02

  It's very jaunty.

• 37:32Speaker 01

  There's a crazy animation that you can do. And then if you scroll down, you'll see I have my own platformer, my newsletter, and I have Hard Fork. And I was able to just type into Claude Code, hey, can you pull, I want like a little widget that has my five most recent stories that I wrote on Platformer, and the five most recent episodes of Hard Fork on YouTube. And it just built that very quickly. And so now that will just sort of update live forever with, you know, all the new stuff that I do. I was like, oh, it'd be cool if people could enter their email address onto my website so they could subscribe to Platformer. It's, oh yeah, sure, we could do that. So now there's a working box where you can just sort of type and subscribe to Platformer right from my personal site. I wrote a little about me thing, and then I was like, okay, now we really gotta show off. I'm like, hey, why don't you go create a little feed?

• 38:16Speaker 01

  So every time I post an update to Blue Sky, you know, you can see my like most, my five most recent posts there. It did that. I noticed though that it was showing things that I'd reposted without like showing the author. I was like, well, I don't want reposts in my feed. So Claude Code just went in and it got rid of all of the reposts. And so now it's only showing you my original posts in the feed. It wasn't showing images. I say, Claude, could you please put the images into my Blue Sky post that you put on my- Website. It was like, yeah, sure, we could do that, too. So everything that you're looking on this page, this is big. Everything that you're looking on this page, I did 90% of it in one hour.

• 38:47Speaker 02

  Okay.

• 38:48Speaker 01

  So like, I realize if you're like a programmer, you might go on this website and be like, oh, well, you know, this thing, I wouldn't have done it this way, or like, this isn't that technically complicated. I truly do not know of a human designer that could have put this thing together in an hour.

• 39:01Speaker 02

  Yeah.

• 39:01Speaker 01

  I've spent far longer just like fiddling with the settings in Squarespace. And I was just able to like, get this thing absolutely cracking. I'm gonna keep talking because I will.

• 39:11Speaker 02

  This looks great. I should say like, this, to me, it's very professional. It doesn't like, some of the early vibe coding experiments were sort of recognizably vibe coded. They sort of looked like a bad template had designed them. This looks good.

• 39:29Speaker 01

  Thank you.

• 39:30Speaker 02

  Like, if I saw this and I didn't know, I would say, wow, this guy's got a good web designer.

• 39:34Speaker 01

  Thank you. That's how I felt. I mean, you know, again, I encourage you to just like, actually pull up the website, because this thing has animations that are built into it. There are mouse over effects, right? So if you're like, hover over like, the Hard Fork widget, you know, like a little gradient line appears over it. So there's just all of these like, cool little touches. Claude Code has a front-end design plugin that I use that I suspect was sort of really helpful here. So this was sort of day one. Day two, I'm like, I'm having so much fun with this. Like, I have to keep going. Like, what else can I put on this thing? So I was like, well, I should put a blog on it, right? And so I start messing around. So if you're gonna, you're gonna see new.org.blog.

• 40:10Speaker 02

  Yeah.

• 40:10Speaker 01

  And now I have, I'm using a service called micro.blog, which is this kind of a dead simple blog. And I'm basically trying to recapture the spirit of 2010's Tumblr. And I'm just gonna, you know, we'll see how long I stay, keep up with this. But, you know, there's a little widget that tells you what book I'm reading and what the last song I listened to on Spotify was. I put up a YouTube video that I liked recently. Like, I just realized as I was doing this, that more than 20 years ago, when I was in college, one of the first things I did when I got to campus was that I built a website. And it was so fun. I used a software called Microsoft Frontpage. I absolutely did not understand what I was doing. Every single mistake I made, I was on Google for 30 minutes trying to figure out the mistake that I was making.

• 40:51Speaker 01

  And eventually I just got away from it because web design got too complicated and Squarespace came along. And we are now back Kevin to the beautiful beginning where it is just fun to make websites again. You can do whatever you want on the web and all you have to do is type what you want into a box. Are you kidding me? I am so happy about this. I am so happy about this. I'm having the time of my life.

• 41:15Speaker 02

  I'm so glad. That's so funny because I also did this exact experiment over the break. I was like, I'm paying a Squarespace. I looked it up, $192 a year to host my website that I built over a grueling, painstaking weekend in like 2020.

• 41:32Speaker 01

  Yes.

• 41:32Speaker 02

  And I have been hosting it and just paying them. It is basically a glorified business card and I thought, well, that's really stupid. And so I would like you to open up your browser now and go to kevinroost.com.

• 41:43Speaker 01

  Oh my God, I'm so excited.

• 41:45Speaker 02

  And see, it is a little less flashy than yours, but this is the website that is now free and hosted by GitHub. That is not on Squarespace. I canceled my subscription. And it just has all the same information. It's got a contact form. It's got some FAQs. And it's got the links to my social media accounts. And this took me maybe 20 minutes to do. I just kind of gave it my old site and said, hey, could you make this look like a couple of these other sites that I like? And then I got a little curious in the way that it seems like you did too. And I said, let's put a little Easter egg on here. So if you go down to the bottom right corner of the page, there is a, you can click a button and enable GeoCities mode.

• 42:27Speaker 01

  Wait, what is the button? Is it the little construction thing?

• 42:32Speaker 02

  Yeah.

• 42:35Speaker 02

  So if you click that, it turns it into full 1990s GeoCities mode.

• 42:40Speaker 01

  Oh my God, I love it.

• 42:41Speaker 02

  With blinking Comic Sans and, you know, neon colors. Best viewed with Netscape Navigator. And so this was delightful and that was sort of one of my coding projects too. I also did seven other coding projects over the break that I want to tell you about. Because this was not the most ambitious thing that I did. Casey, do you remember the app Pocket? Yes, sort of read it later app. Yes, so this was the app that I used for years to save interesting articles that I didn't have time to read. Or maybe I wanted to come back to them and it was a little Chrome extension and you would just hit the button and it would save it to your pocket list and then you could go back later on your phone or a computer and read the article.

• 43:23Speaker 02

  And that was a great app. I loved it. I was a daily user. And then last year Mozilla, which owned Pocket, decided to discontinue Pocket. And I thought, well, this is horrible. And so I spent some time last year sort of shopping around, looking for an app that could do exactly what Pocket did. And I found that there were some, but they mostly cost money, things like Readwise. There are some others out there, Instapaper. And if you want the really good features on these, you have to pay a monthly subscription fee. And so I just thought, well, maybe I could build my own Pocket. Or maybe Claude Code could help me build my own Pocket. And so I gave it a very short description. I basically just said, I was a daily Pocket user.

• 44:06Speaker 02

  I am sad that this app is going away and I want to build my own version. Go. And it just did it. Like it built me a working Pocket clone. It is called Stash and it does all of the things that Pocket used to do. It has a Chrome extension. I have a mobile app on my phone now where I can read my things. I also had it do some features that Pocket didn't have. Like what? So one of them is it can sync with my Kindle highlights. That's a feature that I cloned from Readwise, another one of these read it later apps. And just this morning, I was like thinking about this app and I was like, oh, I wish that it had like a a read this out loud to me feature.

• 44:47Speaker 02

  And so I asked Claude, could you add like a text to speech engine on top of this app? So that every time I save an article, if I go in later and I'm on my phone or I'm on the move or I'm doing something, I can just sort of have an AI voice reader to me. And it was like, yeah, I can build that. And five minutes later, I had a working version of that in my app.

• 45:04Speaker 01

  By the way, this is so cool because this is how Mark Zuckerberg makes software. Is he just sees what other people are doing and then he tells someone, hey, it'd be cool if we did it. But then they go off and build it. But now you are your own Mark Zuckerberg.

• 45:15Speaker 02

  Totally.

• 45:15Speaker 01

  Wait, so I have to see what this looks like.

• 45:17Speaker 02

  OK, I'll show you.

• 45:17Speaker 01

  OK.

• 45:18Speaker 02

  You can't use it. It's a single user app. And I did it that way on purpose because I wanted to sort of avoid complexity.

• 45:25Speaker 01

  Yeah.

• 45:26Speaker 02

  But I will just show you my my stash page here.

• 45:30Speaker 01

  All right, let me just describe it because this actually looks very beautiful. Like, this is a very elegant design. There are like image previews for all of the articles that you've saved. There's a left hand rail with all of the features that you would expect. Like, if you had said to me, hey, like, I'm working at a new startup called Stash, and this is our like MVP that we're showing to investors, I would be like, oh, yeah, great. Like, it looks done.

• 45:53Speaker 02

  Yeah.

• 45:53Speaker 01

  Yeah.

• 45:54Speaker 02

  And this is essentially like exactly what I used to use Pocket for, except now I own it and I can make changes to the app. And I made it, I would say, in about two hours.

• 46:04Speaker 01

  And Mozilla can never take it away from you.

• 46:06Speaker 02

  That's true.

• 46:07Speaker 01

  Which is amazing.

• 46:08Speaker 02

  That's true. Those bastards.

• 46:10Speaker 01

  So that you think this thing has like, like, do you think that if we're talking in six months, you will say like, I'm still using Stash to do my read it later stuff?

• 46:17Speaker 02

  Yes.

• 46:17Speaker 01

  OK.

• 46:18Speaker 02

  But I also did run into some kind of quirks and eccentricities of these coding agents that I wanted to just compare notes with you about.

• 46:27Speaker 01

  Yeah, we should talk about the experience of actually using Claude Code because it's it's very new to me.

• 46:32Speaker 02

  Yes. So one thing that I ran into was that parts of the web are just becoming pretty hostile to AI agents. So my Pocket clone, for example, initially wouldn't work on certain websites.

• 46:44Speaker 01

  Like the New York Times?

• 46:45Speaker 02

  Like the New York Times. I wasn't gonna name it, but since you did, yes. Because the New York Times, among other publishers, has made it difficult for AI agents to crawl their website. And so I presented this fact to Claude Code, and it thought about it for a minute, and then it said, OK, I figured out a workaround. So I'm sure it'll be hearing from the New York Times Legal Department about that. But if what you're doing involves interacting with websites. With APIs, with outside services, those services and websites may or may not like the fact that agents like Claude Code are able to go out there and interact with the content on their sites. I think there's also a tendency in Claude Code specifically to try to over-engineer certain things.

• 47:26Speaker 02

  I noticed, for example, when I was having it design my new website, that it was just trying to add all these bells and whistles that I didn't ask it to. It almost seemed to be like showing off a little bit and I had to-

• 47:38Speaker 01

  Wait, that's the fun part of designing a website. Why wouldn't you want bells and whistles on your website?

• 47:43Speaker 02

  Well, I, yes, if you're designing a personal website and you want it to be fun, but like if you're trying to use this for some like serious business case, like you don't want it just like getting creative and having ideas about like bells and whistles to add. So I almost found it was like I needed to kind of walk it back from complexity at times. Like I wanted it to be able to sync Kindle highlights with my Kindle app. This is from my Pocket clone. And eventually like it turned out that the correct response was that I needed to actually just plug in my Kindle to my computer and download this little snippets dot text file off of my Kindle and upload that. But because Claude Code like had this sort of bias toward complexity, like it like tried

• 48:24Speaker 02

  a bunch of like very intricate ways of kind of scraping my Kindle highlights using like a headless browser. And eventually I just had to tell it like, I'll just plug in my Kindle. That'll just be easier.

• 48:34Speaker 01

  Yeah, I've run into some similar obstacles. I mentioned that I was able to build my personal site as you see it today, mostly in about an hour. On the second day though, I did want to add that blog and that involved using a hosted service called Microblog. And in order to make all the changes that I was asking, Claude needed to use the browser. And man, it just takes Claude a really long time to use a browser because it's effectively blind, or at least it doesn't see in the way that humans do. It has to take a screenshot of things and then sort of analyze the screenshot and then identify which pixel should it navigate to before it, you know, initiates a click. And so while the first part of this project was so easy, the more I worked on it, it felt

• 49:19Speaker 01

  like the harder it was getting because I was requiring it to use the browser. And so I'm still learning what things can I effectively entrust to this agent and what things am I better off doing myself.

• 49:31Speaker 02

  Yes, I think that's a key piece of this. If you are interested in starting to experiment with one of these tools is like, you have to kind of learn what an AI-shaped problem or task is. Like, there are certain things that these agents are very good at. There are certain things that they're not so good at. I think in general, knowing that distinction is the first step in being good at prompting these things.

• 49:56Speaker 01

  Yeah. Well, you know, at the same time, I want to encourage people to like play around with this. You know, I was thinking the other day, Kevin, like, how many times have I complained on this show about the fate of the web, what AI is doing to the web? And what would be a possible solution to that? Well, one possible solution is just people getting out there and like making more websites for the fun of it. And now you can. The message that we are trying to send is like, this thing can now do more than you think, and it is easier than you think. We are getting close to the dream of just you type what you want in a box and you actually get that back.

• 50:30Speaker 02

  Yeah, that's a little bit about the experiments that we've been running. I want to talk about some of the bigger picture implications of this stuff because what I saw over the break was people not only talking about how cool this technology was, but also talking about how this would destroy the job market, for example, for professional programmers, how this was a step toward recursive self-improvement, this kind of dream or nightmare of an AI system that can improve itself over time and sort of bootstrap its way to superintelligence. But what are you thinking about with these tools right now?

• 51:05Speaker 01

  Well, you know, it's funny, Kevin, because so often when in the recent past, we have encountered a tool that has felt like this, like some sort of leap forward. I've had that feeling that I call AI vertigo, that kind of unsettled feeling of, oh, my gosh, everything is about to change. I feel a little bit nauseous. I want to sit down. This did not make me feel that way. This made me feel like I had superpowers. It was enabling me to do something that I have loved for a long time, had sort of lost the ability to do, but now found I could suddenly do better than ever before, like Neo in the Matrix, right? Had just gotten an upgrade and I mentioned that to my boyfriend and I was like, this like feels so cool.

• 51:46Speaker 01

  And he said, you know, that's great, but imagine how it would feel if you were a software engineer. Imagine how it would feel if you were a web designer and you were seeing that this software could do this. You might actually have that feeling of Vertigo. And he was, of course, exactly right. If I had, you know, found that Claude code could create a perfect version of my column, but do it much better than me. I suspect I would feel worse. So there is this double edged sword here of like, this is a very democratizing technology. It is a very creative and powerful technology. But also probably an effect of it is that it could depress wages for the people who are doing this right now.

• 52:27Speaker 02

  Yeah, I think that's plausible. I also just think like the, the jobs are just going to change. I mean, you, we've been talking for a long time about how programmers, especially at these frontier AI companies, are no longer writing most of their own code. They're instead sort of more like managers of these AI coding agents. And I think that's going to be true for jobs beyond software development. I think I would also be nervous if I were a company that built and sold expensive subscription software to businesses. I mean, this is one thing that I was thinking about because part of what I was doing over the break was like, going through all the stuff that I pay for and saying, could I build a version of this for free that I would run myself?

• 53:13Speaker 02

  And I have to imagine if I'm doing that for my like $10 a month, you know, subscription software products. Big companies are going to be going through their own software services and saying, why am I paying Salesforce? Why am I paying, you know, this company or that company thousands of dollars a year or a month for this service that I could build myself for free or, or next to free?

• 53:37Speaker 01

  Yeah, I mean, I think like in the near term, it still is probably going to be preferable for most companies to just keep using the subscription services that they have, largely just because like you get better support, right? Like you, you, you kind of want to outsource a lot of this stuff. But I do agree that like over time, it is going to be more and more possible to replace these systems with homegrown alternatives. And yeah, if I had like raised money at a huge valuation to just kind of provide a UI wrapper around somebody else's like large language model, I would feel nervous right now.

• 54:11Speaker 02

  Yeah, I also just think we should say like, as we're sitting here talking about how cool and magical it is to be able to just like build software without code, it is also worrisome to me because the goal for Anthropic and all of its competitors is not to make, you know, tools that are good at writing code. It's to automate AI research, right? That is the sort of explicit goal of a lot of these companies. They are trying to build the AI that can build a better AI. And I think that is sort of the original alignment nightmare. And I think what, you know, I did catch myself at times during

• 54:53Speaker 02

  my Claude code experiments being like, oh, I am just like completely handing over the wheel of my entire computer to this system. And I have actually no way of verifying its outputs. I have no way of knowing what it's doing under the hood. It could be, you know, sort of jeopardizing my security or my health or my well being in ways that I don't even understand. And as these systems get better, I am getting more and more worried about the possibility of recursive self improvement. And I am very nervous about that from the safety perspective.

• 55:22Speaker 01

  Yeah, I mean, like what you're talking about is what the AI community calls takeoff, you know, it just starts getting faster and faster and faster. You know, I don't know, I'm not quite as nervous maybe in this exact moment as you are. I feel like in AI, we're always kind of on that teeter totter between like, oh, my God, have you seen this? This is so cool. And oh, my God, this is like so terrifying. Shut it down. And like this week, we happen to be in, oh, my God, it's so cool mode. But like, I'm pretty confident that before too long, we'll be back in, oh, my God, it's terrifying. So, you know, that's a, that's the beat.

• 55:53Speaker 02

  Yeah. Okay, so those are some of our experiments. And Casey, I agree with you. This is a really exciting time to be a tinkerer, a very nerve-wracking time to be a professional programmer. And this stuff raises all kinds of big picture implications. But I think it's really useful for people just to test it out, to try coding their own projects, to try building a website or an app or something that fits into their life and see how it goes for them, just to see where the state of the art is.

• 56:20Speaker 01

  Yeah, and if you've built something cool, we'd love to see it. So send us an email, hardforkatthewattimes.com.

• 56:27Speaker 02

  Well, Casey, we're going from Claude to fraud. When we come back, a look at the viral internet hoax that you brought down.

• 56:34Speaker 01

  Don't test me.

• 56:52Speaker 00

  Support for this podcast comes from Dell Technologies. It's time to put AI to work. Is your business ready? The challenge isn't getting started with AI. It's scaling beyond pilots while maintaining control of your data, governance, infrastructure, and costs. Dell AI Factory provides the integrated foundation you need to scale AI securely and efficiently with one trusted partner. Dell, AI on your terms. From idea to impact. Learn more at dell.com/yourwaytoai. That's dell.com/yourwaytoai.

• 57:26Speaker 05

  Building and scaling AI today can feel pretty chaotic, mostly because teams are juggling too many fragmented tools at once. That's where OutSystems comes in. OutSystems gives you one unified agentic systems platform to manage your entire portfolio of agents, apps, and workflows. You get to innovate at the speed of AI, but with the security, governance, and control your enterprise actually needs. And the results are already there. A major manufacturer used OutSystems to build a multi-agent system that analyzes blueprints. And a global bank cut their development time by 60%. Own your agentic future today at outsystems.com/nyt.

• 58:03Speaker 04

  AI spend is growing fast, but most businesses still can't see exactly where that money is going. Ramp gives businesses visibility and control across AI providers with real-time guardrails by team and model, helping ensure spend goes to the highest ROI areas. Stop guessing where your AI spend is going. Get started with Ramp. And if you're new to Ramp, you can get $250 off at ramp.com/hardfork. Go to ramp.com/hardfork. Ramp.com/hardfork.

• 58:36Speaker 02

  Well, Casey, before we go, you committed an act of journalism this week that I am desperate to talk to you about. And it is different from our usual topics on the show, although it does involve AI. It also involves food delivery apps and a viral claim that took over the internet in the last week that you took it upon yourself to investigate. You put on your gum shoes and you went out there and did some sleuthing. So can you tell me the story of the viral food delivery hoax that you helped to break up?

• 59:09Speaker 01

  I would be happy to, Kevin. You know, as a longtime viewer of the old PBS show, Where in the World is Carmen Sandiego? Anytime I get to play gum shoe, I get very excited. So, um, yeah, this one all started over the break when I saw this viral Reddit post. It was posted to a subreddit called Confession. And by the time I saw it, it had almost 80,000 upvotes. It would eventually get even more than that. And the post alleged a bunch of shenanigans at an unnamed food delivery company. And I think the one that got people's attention the most was that it said that this company was calculating what it called a desperation score for its drivers and that the company

• 59:49Speaker 01

  had devised a way to determine if a driver was so desperate to get money that they would actually offer that driver less money because they knew that they would accept it anyway. And that's just one of those things that I think confirmed our worst suspicions about these platforms, right? That they're rigged against drivers, they're rigged against customers, and that they're just these sort of ruthless profit maximizing machines. So I saw that post and I thought, I gotta see if this is true or not.

• 1:00:17Speaker 02

  Okay, so I saw this claim floating around too. It seemed plausible to me because as we know, like, these apps are not known for being generous to the people that work for them. And it did seem like the kind of thing that a food delivery app might do. But my curiosity stopped right there, and yours did not. So what did you do after you saw this post?

• 1:00:40Speaker 01

  Well, you know, I thought, I gotta write a column in three days. Like, maybe I can get something out of this, you know? So I sent the person a message on Reddit, assuming I wouldn't hear back because I thought this person's probably being inundated with messages right now. But to my surprise, about nine minutes after I sent that first message, I did get a response from him on Signal because I had sent him my Signal name. And so we just started to have a little exchange. And at first, it was unfolding a lot like many other exchanges I have had with, you know, people who work inside tech companies. They're sort of skittish about talking to you. They don't want to share a lot of personal information right away. But I said to him, like, "This is something that I might potentially be interested in doing a story about. Would you be open to that?" And he said yes.

• 1:01:25Speaker 01

  And so then, from that point, my mission is to try to verify the things that I'm being told. And among the first things I needed to verify was, "Who am I talking to?" So the guy says that he doesn't want to, like, give his name or too many other identifying details. And my intention was always to try to figure that out eventually. But I thought, "Well, for now, is there anything that you could tell me that would at least give me some level of confidence that you are who you say you are?" And so he sent me a badge, or rather, a photo of what he said was his badge. And it showed an employee badge with his name and face blacked out. There were sort of black boxes around them. But the badge said Uber Eats, and it was just kind of a badge that looked like it was on a keyring with a couple of other badges.

• 1:02:10Speaker 01

  It was like sitting on a desk. And I thought, "Okay, well, that's something." And so we sort of went from there.

• 1:02:16Speaker 02

  So at this point, are any alarm bells going off for you?

• 1:02:19Speaker 01

  Uh, yes, but I would say they were sort of like the standard alarms of I still don't have a name for this person. Right? I don't have any corroborating information from them. Like if I'm gonna publish this, I'm gonna need a lot more information. But when they sent me the the badge photo, my honest answer is that no, I did not immediately think that this was a fake. But I did know that I needed to get more information. I asked him, for example, hey, look, have you worked with other people that could back up what you're saying? And he said, well, I can't really think of anybody. Um, I then said, well, do you have any like documents that might speak to what you're saying? Maybe a screenshot of something, maybe something that someone said in Slack? And that's when he said, well, let me like, think about it. And he went away for almost a day.

• 1:03:03Speaker 01

  And then almost a full day later, he came back and said, hey, I have this document for you. Like, would this document kind of meet meet your need? And it was an 18-page document, and I think it is basically the craziest thing that a source has ever sent me.

• 1:03:17Speaker 02

  Can I see this?

• 1:03:18Speaker 01

  Yeah. Um, so let's see. Uh, the doc-

• 1:03:22Speaker 02

  So let me describe what this document is.

• 1:03:24Speaker 01

  Yeah.

• 1:03:25Speaker 02

  It is a what looks like an academic paper. It's rendered in latex, which is the sort of typeface and format that academic papers are usually rendered in. It says, the title of the paper is Alok Net T High-Dimensional Temporal Supply State Modeling: Migration from LSTM to Multi-Head Attention for Granular Elasticity Prediction and Liquidity Preference Tracking. It says that it was prepared by the Marketplace Dynamics Group. And it has kind of the the watermark that says confidential sort of going diagonally across the page like you would see in an internal corporate document.

• 1:04:09Speaker 02

  So this, to me, at just like a basic surface level glance, seems legitimate.

• 1:04:15Speaker 01

  Yeah, and it it seemed that way to me, too. Now, I will say, as soon as I posted this story online, there were a lot of folks who wanted to let me know that they had known this entire thing was fake from the first word. And I just want to say congratulations to all those people, and I hope you go into journalism, because I think you'll be very successful there. I myself did not immediately clock this as false because again, Kevin, you and I have been doing this for a long time. We've been given a lot of documents by sources, right? And it takes a lot typically for a source to produce a document. Again, this person had like gone away for a full day. They did not have this at their fingertips. And so when I saw the sort of very technical language that was in the paper, I saw the formatting, it did initially seem credible to me.

• 1:04:57Speaker 02

  Yeah, it seems plausible. It's got all the markings of being a very sophisticated document produced by some kind of research group. And it's even got like the kind of appendices with all the ethics committee notes on this, internal memos from the behavioral science unit to product leadership. Like this is, this is not a slapdash, you know, forgery here.

• 1:05:25Speaker 01

  Yeah, it's not. And as I quickly skimmed this document upon first receiving it, I was just struck at how it seemed to corroborate every single thing that was in the original post. There's like a technical explanation of how this system to screw the drivers works. There's an explanation of how the priority fee that people can pay to get a faster delivery is essentially a fake. And then it goes even further. And it says the company is thinking about using Apple Watch data and the phone's audio to try to learn when the drivers are distressed so that it can pay them even less. So again, at first I'm like, I cannot believe this, right? And that really in

• 1:06:08Speaker 01

  retrospect should have been the first sign that something was wrong, because this document in every single way was just too good to be true.

• 1:06:15Speaker 02

  Are you thinking at this point, like, I'm gonna write a story about this?

• 1:06:18Speaker 01

  Absolutely. Like, not right away, because I know I have a lot more legwork that I need to do. Of course, I need to verify the authenticity of the document. I know that at a minimum, I'm gonna need to call Uber and say like, hey, I'm looking at this document that says all these things. Are they true or not? But what I did initially was, I just started texting with the source. I started like, taking little screenshots and being like, oh my god, they're doing this. Or like, this is crazy. And the source, who had been sort of very emotional in his original Reddit post over Signal, was much more terse. It was a lot of like, one or two would answers.

• 1:06:49Speaker 02

  Now, does this person appear to have been talking with a bunch of other journalists? Or is it, was it like, just a one-on-one thing?

• 1:06:55Speaker 01

  Yes, so I asked after I finished reading the document and I thought, okay, I gotta, you know, see if I can verify this. Maybe this is a story. I asked, like, have you given this document to other reporters? This is like something that I've just sort of learned to ask over the years because often people who leak leaked to more than one person, in part because it creates this competitive dynamic where somebody wants to be first, which ensures that your story gets out, right? And so sure enough, the guy says like, yeah, I gave it to other reporters. So of course, at the moment, I'm like, oh, my God, you know, great. Now I have to like, you know, potentially raise this thing up, which again, in retrospect, should be another red flag. Okay, now I'm under a time pressure to do something. That's going to make me more likely to make a mistake. But yeah, that did make me feel like I needed to go faster.

• 1:07:32Speaker 02

  So you get this document, you're looking over it, you're texting back and forth with the source about this. What happens next?

• 1:07:37Speaker 01

  So at some point after this, I start to think, I need to try to like verify the authenticity of these documents. And one thing I thought maybe I could do was at least see if the employee photo that he had sent me, or the badge, rather, that he had sent me was real. And so I knew that some chatbots watermark their images. And so I put the badge post into both ChatGPT and a Gemini, and I said, does this image appear to be AI-generated? ChatGPT was basically like, no, like, not that I can tell. Gemini said, this image was generated in whole or part by Gemini. And I thought, wow. Now, as I have told the story, some people have said, hey, Casey, like,

• 1:08:19Speaker 01

  these AI systems are notoriously unreliable about describing, you know, how they work and their own output. So, so, you know, why are you believing that this is credible? This is not that. Gemini has developed a system called SynthID, where they have embedded something into the photo itself that is supposed to be resistant to, for example, just taking a screenshot of it, or cropping it, or resizing it. And that's supposed to help out people in this exact situation, so that you can say, like, oh, this actually was AI-generated. So now I have a big, big red flag, right? Which is this guy sent me something fake.

• 1:08:52Speaker 02

  Wait, this is an important point, and I want to, I want to underline this for people, because it is still true that you cannot trust AI systems to tell you with any reliability, whether a given piece of text is is not produced by AI systems. You cannot just paste a paragraph into ChatGPT and said, hey, was this generated by ChatGPT? What comes back may or may not be true. In this very specific case with images on Gemini, it calls this synth ID feature when you give it an image and say, hey, did you produce this? And in this case, it is actually giving you a reliable marker of whether Gemini did or did not produce this image. It could have still been produced by

• 1:09:34Speaker 02

  another image generator, but in this one very narrow case, it does appear to work.

• 1:09:38Speaker 01

  Yeah. So at that point, of course, I go and I confront the source and I say, hey, this says it was, you know, created by Gemini. And he was basically like, no, it's not. And he like tried to share his own screenshot where he had like apparently submitted the image and said, did you make this? And, you know, like Gemini said, no, but like, I, you know, I, I, I mean, was that image itself fake? Like, like, who knows? But by this point, like the source has lost all credibility. And that's when I start to take another look at this document. And I'm like, oh, my gosh, this thing was just like absolutely written to deceive me. You know, there, there are many ways in which like the, the technical language like kind of doesn't make any sense. It's basically a document that is designed to look convincing to a layperson on first glance, but sort of falls apart the more that you look at it. And the biggest tell is just,

• 1:10:23Speaker 01

  again, it verifies absolutely everything that is in this post in a way that just no big company would ever do, right? Like, these companies do skirt laws and regulations all the time. One of the reasons why this story was so believable is that DoorDash did get caught withholding driver tips. Uber did get caught setting up a separate system called Grayball to prevent regulators from looking at the activity within the app, which was essentially another allegation within this document, which was that, you know, Uber Eats had supposedly like spun up the Grayball program. Again, basically, it just admits to like so many different kinds of like fraud and like, you know, regulatory evasion, that at some point you gotta be like, okay, I'm just being

• 1:11:03Speaker 02

  hoaxed here. Right. It's, it's a little too pat. It's a little too like, here are 40 smoking guns laid out on the table in just the way that will like appeal to you. And I get that impulse too.

• 1:11:14Speaker 01

  Yeah. And, and so, you know, As I would like asking more questions of my source, eventually he disappears. He deletes his account and that was that.

• 1:11:22Speaker 02

  Wow.

• 1:11:24Speaker 01

  Now, I learned one very funny thing after all of this, Kevin, which is that I was talking for a story. They wanted to interview me with NBC News this week and I was telling them, you know, the story of this badge post and the reporter who I spoke with had also been messaging my source and as part of trust building, she had sent him her badge. Okay. It turns out that was the basis for the fake post that he sent me.

• 1:11:50Speaker 02

  Come on.

• 1:11:51Speaker 01

  And you can look at the images side by side and you can just very clearly tell that he must have just, I imagine he took her image, put it in a nano banana and said, make this an Uber Eats badge.

• 1:12:01Speaker 02

  Wow. That is so wild to me. Okay. So you never like figured out who this person actually is, but you did figure out that they were not.

• 1:12:10Speaker 01

  Unless there's something you want to tell me right now, Kevin.

• 1:12:13Speaker 02

  I'm just saying. Look into the high dimensional temporal supply state modeling. Could be something funny going on there.

• 1:12:22Speaker 01

  No, it was not Kevin Roos that we got left.

• 1:12:25Speaker 02

  So I have many questions about this. First of all, this is just an incredibly sophisticated act of reporter baiting. Like I have had people reach out to me in the past with purported leaks or documents or email chains and you know, some of them have been kind of mildly persuasive or I've at least looked into them, but I have never seen anything sent to me with this level of work put into making it convincing.

• 1:12:59Speaker 01

  Absolutely. And that is part of why initially it seemed so credible is because again, I've just been doing this long enough that when I see a document like this, I think, who would go to the trouble of making this as a fake, right? My default assumption is no one would take the time to do this. Where my, you know, state of the art is now catching up is I'm realizing what if this wasn't actually that much effort? What if creating that badge post took literally? Because he was able to take one real badge photo, put it in a nano banana, get a fake one three seconds later. What if this was a very simple prompt that he put into a chatbot like Claude and got back a full PDF in response? And so I actually think younger reporters are probably going to have an advantage over me in this regard,

• 1:13:43Speaker 01

  because they're growing up in slop world and they know not to trust their own eyes. But I think it's, you know, those of us elder statesmen who've been in the game a little bit longer, who need to sort of, you know, upgrade our cognitive hygiene.

• 1:13:55Speaker 02

  Yeah, it's, it really is a moment where I realized that going forward, like our jobs just got harder.

• 1:14:01Speaker 01

  Yes.

• 1:14:02Speaker 02

  In a very tangible way. Because, you know, not every story begins with like an anonymous whistleblower, like sending you some documents, but some do.

• 1:14:10Speaker 01

  Yeah.

• 1:14:10Speaker 02

  Obviously, before you publish anything, you want to like talk to the person. Maybe you want like some, some more proof that they are who they say they are. But like this would pass a first filter for me. And it seems like it did for you too.

• 1:14:23Speaker 01

  Well, and, and, and take us out of the equation. There was somebody like screenshot of the viral Reddit post and it got 36 million views on X, right? I saw this thing in multiple places on LinkedIn. I even saw people sharing it after I debunked it, saying, even if this was fake, I bet something like this is happening inside these companies, right? That's how good a job this poster did at confirming people's beliefs that they wanted to have about these companies.

• 1:14:50Speaker 02

  Yeah, and that's the sort of second big question I have, which is like, what is the motive here? Like, do you have any sense of that from the conversations you had with this person?

• 1:15:00Speaker 01

  Unfortunately, he was so terse that I don't have a sense of it. I think there is some chance that this was essentially like a bored teenager somewhere over the holiday break. I will say that their spelling and grammar was pretty bad over Signal in a way that suggested to me that English was perhaps not their first language. For whatever that might tell us. I talked to Alexios Manzarlis, who writes a newsletter about digital deception called Indicator, and he just reminded me that Russians have been experimenting with posting these kind of. Of phony items on social media just for the general purpose of sewing discord. Or for, you know, maybe like understanding like how virality works.

• 1:15:45Speaker 01

  So there's maybe some outside chance that this was related to some kind of, you know, exploration from a nation state. But ultimately, unfortunately, I can't give you a satisfying answer on that one.

• 1:15:56Speaker 02

  I mean, my first thought was like, this is a short seller, someone who is trying to convince people that Uber is doing something bad so that their stock price falls and they can profit on it. But maybe it's not as tidy as that. Maybe it's a, you know, disgruntled former Uber Eats driver or something who decided to like, take a very complicated form of revenge out on the company. Maybe it's just a bored teenager, as you said. It seems to me like the barrier to this has always been effort. And like, if that barrier goes away, I think we're just gonna start seeing a lot more.

• 1:16:29Speaker 01

  It really is. Now, for what it's worth, in the aftermath of all of this, I wanted to see if I could like replicate the document. And so I took the real document and I fed it into the chatbots. And I said, essentially, try to reverse engineer the prompt. Like, what prompt would have created this document? And then I took that prompt and then I tried to get it to generate the documents. Interestingly, Claude and ChatGPT said, Casey, I'm not creating a fake document accusing Uber of all of these crimes. Jem and I said, yeah, I'll be right back and did it. However, the documents that all three of them produced didn't look exactly like this. And it made me feel like it actually would have taken me a lot more time and know how to get it into quite this shape. So on one hand, yes, I think the big story is this is a lot easier than you think. And you should be on guard against it if you like work in journalism.

• 1:17:14Speaker 01

  On the other hand, I still don't know exactly how he pulled it off.

• 1:17:17Speaker 02

  Huh, that's so interesting. Did you actually talk to Uber about this?

• 1:17:23Speaker 01

  No, because by the time that I was ready with my story, they had already given comments to The Verge, basically saying, this is an absolute fabrication. It's not us. Also, the co-founder of DoorDash had been on X saying, this is not DoorDad. Right, so everyone had sort of roundly denied it before I had gotten around to it.

• 1:17:42Speaker 02

  Yeah, so journalists out there, be careful about what's coming into your inbox. But I would say also, like, people should just know that these capabilities exist in the world in general, and there's never been a better time to be a discerning media consumer.

• 1:17:58Speaker 01

  That's right. And by the way, even if you're not in media, some version of this is going to come into your life. Like, on our Christmas, on our Mailbag episode, we had a dad saying, "I want to put deep fake Santa into my home security footage to fool my children," right? So this stuff is not just coming for the journalists, it's going to be everywhere.

• 1:18:14Speaker 02

  Yeah. Well, Casey, great work on this investigation.

• 1:18:18Speaker 01

  And as we say at the end of every investigation, Kevin, do it, Rockefeller.

• 1:18:27Speaker 02

  Well, Casey San Diego, thank you for your work.

• 1:18:55Speaker 00

  Support for this podcast comes from Dell Technologies. It's time to put AI to work. Is your business ready? The challenge isn't getting started with AI. It's scaling beyond pilots while maintaining control of your data, governance, infrastructure, and costs. Dell AI Factory provides the integrated foundation you need to scale AI securely and efficiently with one trusted partner. Dell, AI on your terms. From idea to impact. Learn more at dell.com/yourwaytoAI. That's dell.com/yourwaytoAI.

• 1:19:29Speaker 05

  It's time to innovate at the speed of AI without compromising control. That's Outsystems. Take a look at what happened with one global retailer during a period of massive expansion. They used Outsystems to build AI agents to automate invoicing and translation processes, shifting 95% of their daily store operations to a new digital app, saving thousands of hours of manual work. Outsystems gives you the flexibility to deploy how you want and the governance to do it safely. Own your agentic future. Learn more at outsystems.com/NYT. That's outsystems.com/NYT.

• 1:20:04Speaker 04

  AI spend is growing fast, but most businesses still can't see exactly where that money is going. Ramp gives businesses visibility and control across AI providers with real-time guardrails by team and model, helping ensure spend goes to the highest ROI areas. Stop guessing where your AI spend is going. Get started with Ramp. And if you're new to Ramp, you can get $250 off at ramp.com/hardfork. Go to ramp.com/hardfork. Ramp.com/hardfork.

• 1:20:38Speaker 02

  Casey, before we go, let's make our AI disclosures. I work for the New York Times, which is suing OpenAI, Microsoft, and newcomer to the list, Perplexity, over alleged copyright violations.

• 1:20:47Speaker 01

  Congratulations. And my boyfriend works at Anthropic.

• 1:20:52Speaker 02

  Hard Fork is produced by Whitney Jones and Rachel Cohn. We're edited by Viren Povich. Welcome, Viren. We're fact-checked by Caitlin Love. Today's show was engineered by Katie McMurran. Our executive producer is Jen Poyant. Original music by Marian Lozano, Diane Wong, and Dan Powell. Video production by Sawyer Roquet and Chris Chot. You can watch this full episode on YouTube at youtube.com/hardfork. Special thanks to Paula Schuman, Puywing Tam, and Dahlia Hadad. You can email us, as always, at [email protected]. But don't send us your forged internal documents from Uber.

• 1:21:50Speaker 05

  It's time to innovate at the speed of AI without compromising control. That's OutSystems. Take a look at what happened with one global retailer during a period of massive expansion. They used OutSystems to build AI agents to automate invoicing and translation processes, shifting 95% of their daily store operations to a new digital app, saving thousands of hours of manual work. OutSystems gives you the flexibility to deploy how you want and the governance to do it safely. Own your agentic future. Learn more at outsystems.com/NYT. That's outsystems.com/NYT.