You can't ignore Google Zero anymore | The Vergecast

Hello and welcome to the Vergecast, the flagship podcast of AI mode. I'm your friend David Pierce. Nilay Patel is here. Hello, sir.

What's up? How are you? How's life? I'm good.

You had an electrician come into your house today and you explained this to me as I have an AI vignette. What is it? Tell me what's going on.

So, we got solar panels at the end of last year. You know, we chased up all the Biden tax credits we could get. Sure. So, we leased an EV and we got solar panels. And now the government pays me to charge my car.

Sorry. I know you don't want to, Donald, but you're doing it. You are the deep state now.

You're paying me to charge the car. The way the net metering works out and the blah blah blah in New York State, they pay me to charge the car. This is all great. The problem is ConEd is a utility and it is bad in the way it's bad. And sometimes they don't get it right and you have to call them and argue with them about metering. Great. And so to do that, you need the data from your solar panels to be accurate. I discovered the data for my solar panel was not accurate. I was losing arguments. I hate losing arguments. So, I just had Claude pull the 15-minute interval data of how many kilowatts my system thought it was doing, how much ConEd thought it was getting, reconcile it all, and then generate a troubleshooting email for the solar people and the electricians. And I will say it was so rude. This email was so rude that I had to tone it way down, but it was right.

So you just basically had Claude read your meter for you.

I had Claude read both meters, the ConEd meter and the Enphase meter. We have an Enphase solar system and it just pulled the interval data from both systems over a pretty long period of time. This is something I would never have the time to do. Sure. Generate the report and then diagnose the problem from that report, which is that the consumption CTs were backwards and basically ConEd was getting half of what it should be getting.

That was going to be my first guess. So I was like, 'Oh, I bet it's the consumption CTs.'

But the email it wrote was like, 'Rain or shine, production or not, air conditioners on, like it's wrong and you know it's wrong.' And I was like, 'Calm down.' Like bring it down a couple. I just rewrote the email. Like it was very demanding.

It's not my problem, man. It's your problem.

But it was one of those things where it's like, 'Okay, I've got like two databases that don't agree. What is the tool?' And the tool is right in front of you. There's my AI vignette. It's pretty good.

Did you present your findings to this electrician as Claude told you?

No. No. No. No. No. You don't do that.

That's good. Good. That's so rude.

I was like, 'Look, I made a table. I'm not going to presume a computer was involved.'

Is that computer in love with Kevin Roose? I declined to say it this time.

No one knows.

Yeah. Okay. That is good. And that actually feels like a good and correct use of AI. You're like, 'Look, I have a bunch of numbers. Can you make sense of these numbers for me?' And I'm sending so much power back to the grid, it offset my energy consumption.

How's that?

And you poured water on the ground as you were doing it just to prove a point. That's good. All right. We're not here to talk about electricians. We have a bunch of stuff to talk about today. There's a lot of AI news. The second two-thirds of the show is going to be very lightning roundy. We just got a bunch of stuff to talk about, a bunch of stuff to get through. But first we have to start with Google Zero. Big week for Google Zero. Google reported earnings. Some really interesting stuff particularly around search. So we have some big Reddit news. You are in the New York Times taking credit for Google Zero as you were wanting from time to time.

Finally. Shout out to K Conger at the New York Times for being a real one and being like, 'I'm doing a Google Zero story,' and calling your boy. But I want to start with a story from the Wall Street Journal, which I think is actually kind of indicative of this whole thing about how Reddit, which we've talked about on the show, made a big AI data sharing deal with Google a while back worth $60 million a year to basically turn Reddit into a database full of stuff for Google's AI overviews and things to pull from. This turned into putting glue on pizza. This has turned into some messy stuff, but has also made Reddit a crucially important part of Google's entire AI enterprise. And there was reporting this week that Reddit is actually thinking about cutting that deal out entirely. And there's a quote that says, 'Reddit executives are assessing what the upside is of continuing to feed its content to Google.' Like just straightforwardly that they're like at $60 million a year, this is no longer worth it to us to feed our data to Google. And I think if you start there, you can get at the whole problem with everything Google is doing with AI right now. What did you make of this change?

This has been coming for everyone for so long. If Google doesn't send you traffic, giving things to Google for free or cheap, and $60 million is cheap. Yeah. For a company the size of Reddit, giving things to Google for free or cheap is economically irrational and everyone will stop doing it. And we can just go down the list of big publishers who are saying we're going to stop letting Google index us at all because we're not getting traffic or money. USA Today, Politico, The Economist, People Inc., Reuters. I'm going to bet you the New York Times will end up on this list sooner rather than later. They have a very testy relationship with Google. A company that we don't yet work for but which is merging with Vox Media in a joint venture that will be called PMX when that deal closes. Penske Media Corporation, they are openly suing Google over AI overviews not delivering any traffic. It's the whole publishing ecosystem is saying, 'Yeah, this isn't worth it.' And then importantly, they paid Reddit $60 million. That is a social media publisher, right? That's a platform. That's not a news publisher, a media organization. Reddit saying it's $60 million. Giving our users' content to Google is not worth it. We might turn this off. And everyone has known this is coming because I have been telling them it's coming for years and years and years. When Google's traffic, when the amount of traffic you see coming to you from Google goes to zero, what is left of your business? Do you have one? Do you have an audience? Do you have anything at all? And if Google turns off the economic incentives for all of these people to make new information, what will Google summarize in the AI overviews? What will be left? Right? And I don't think either side of that equation has had a good answer.

No. And in fact, the thing that has really shifted, you named all those publishers, I think we've been talking about this tension for forever, right? That even if you didn't like what Google was doing with AI overviews, there was still this sense that the deal with Google overall is important enough that you have to keep doing it, right? That we will continue to give our content to Google. We will let Google index and crawl our website and in return it will send us traffic. That has been the deal for forever.

What are you going to do? Turn off Google? I've heard so many publishing CEOs say this to me. What are we going to do? Not have Google traffic? And this has been Google's trump card because you can fight ChatGPT, you can fight Claude, but you can't fight Gemini without fighting Google search. And this is, I would say, a matter of concern for people who are worried about monopolies in this country. It's also just a foundational problem for all of these people. Google has tied these things together that if you want to not be crawled by Google's AI systems, you can't be crawled by its search systems. It's not exactly that simple, but it is fundamentally that simple.

It is that simple. I know Neil Vogel, the CEO of People, has said that that will result in more antitrust litigation. I'm confident that the European Union, which just fined Google a billion dollars for self-preferencing its products in search, they will look at the tie between search indexing and AI training. It is coming for everyone all the time. And you know, just to make the other argument, I think the publishers have been equally stupid here because they were all hopelessly addicted to Google and they never thought that Google would do anything to harm the web, right? Most of them did not make other plans, right? Just straightforwardly, most of these companies did not make other plans. They did not see the asteroid coming at them. Even as I'm going to say it one more time, I would run around this industry being like, do you see the asteroid? Do you see the increase in the percentage of zero-click searches where Google keeps the traffic for itself or redirects it to its own platform? And do you see what AI overviews will do? I've asked this question to Sundar Pichai who will happily tell you that the web per Google's own estimation is bigger than ever and Google per Google's own estimation is sending out more traffic than ever and the publishers and kind of left unspoken all that is and these idiots didn't do anything about it, right?

And I find it hard to argue with that inside the house, right? I get why they're all making noise out in the world like, you killed our businesses, you should make a lot of noise about that, we're going to turn off Google search, we're going to turn off indexing because we owe you nothing, right? This was never really an economic relationship. We're just turning you off. Goodbye. But inside the house, it's like, guys, you did nothing. Like, you saw the problem and thought you could do nothing to solve it. Sure. And here we are at the end of that road. Because I think that belief that there was something fundamentally mutually beneficial that would keep being mutually beneficial just persisted in lots of industries for way too long, right? The SEO business has been relying on this for a very long time and it makes a certain amount of sense. Google invented search advertising, the greatest and most lucrative business in the history of the internet. And it worked as long as it kept sending a lot of traffic to websites. This was the whole idea. People will go, they will click the links, everything will get better. And Google had incentive to grow that and to make more people click the links. People had incentive to make more content that ranked highly. This actually was in a certain way mutually beneficial. It got very messy as Google ate the advertising business and lots of things became complicated. But that basic idea that we give data to Google and Google gives us traffic in return seemed like it was relatively immune to problems because it worked for everybody because on the other side of that is an incentive to make new information for Google to index and surface, right? And if you don't have that in Google search, what do you have? And I think everyone keeps conveniently forgetting that Google also owns and operates YouTube. And so Google's source of new information is YouTube. There was just a Google core update, that's what they call it when they update the Google search algorithm. Every recent Google search core update has basically boiled down to and now we're going to rely on YouTube even more. Yes. Right. The idea that they will summarize YouTube videos or drop you directly into the part of the YouTube video that most clearly answers your query. There's just a bunch of, oh, they had a hedge against us. There was a hedge against publishers and social media and all this stuff and it was always called YouTube and here it is. Yep. And the economic or traffic incentives to make new information on the open web are plummeting. Now, there's a flip side to this which I think supports Google's argument the web is getting healthier than ever. We just had Amy Lansy, who's the CEO of Digit House North America, a huge advertising agency, and she told me on a Decoder a few weeks ago, their number of inbound requests to make landing pages and build new websites and make things that particularly for commerce, like they're an advertising agency obviously for commerce, that are the web so that people can come and transact more directly on the web because they've had search experiences or AI experiences. She's like that I would you would never think that that would be up and that is up. So off in e-commerce world, the web might actually be healthier than ever because the agents are running around clicking on websites for you, right? And that's great if you're an e-commerce business. It is horrible if you're a consumer and you want information that isn't branded content from e-commerce providers. And it is horrible if you are a publisher or a social media site that would like to traffic in non-commercial information.

Yeah. And I think the idea that actually all of the internet is just becoming a series of databases to be mined by these other things just continues to be true, right? Like a lot of people recently have been citing this big internet report that Cloudflare did that said that more than 50% of internet traffic is now nonhuman which suggests that it is in some way AI bots. There was also I just pulled this out of that report. It says today for every hour spent online searching for information only 15 minutes is spent on the open web. Traditional search behavior is collapsing as users shift to AI-driven discovery and consumption. Instead of visiting multiple sites to source and compare information, users simply type a prompt and receive a nearly instantaneous consolidated answer. That's step one. Step two is agents. They just go buy the thing for you. And right, as we talked about on the show on Wednesday, everybody's idea about agentic AI ends in transactions. And so that is the next thing. That's how everybody makes money. Everybody understands how to make money off payment processing. We can kill the web as long as payment processing exists is fundamentally what a lot of these companies are betting on and fine with.

Every day a new startup making payment rails for agentic AI gets its $30 billion wings. Like truly that is TechCrunch every day in some way, right? Yeah. And fine. And by the way, no shade of TechCrunch. That's just the business they're in companies talking about their funding announcements. But you see that news all the money flowing into rails for agents across different things and right next to it is are we having good experiences? Are we getting better information? Is the deal that all of us on every side have made with Google that it would organize the world's information in a compelling way and useful way. Is that being upheld? I don't know the answers to this. And even in their earnings, they didn't quite hit on search, right? There's like a little bit of a miss on search.

Yeah. Yeah. Like barely. It was a very funny thing. They missed by like a hundredth of a billion dollars which is it was a very tiny miss in the scheme of things. But like again Google search has been growing like crazy for forever. This business has been immune to everything. And then you have Google coming out and saying we're spending way more than we thought on all of this AI stuff. Our compute costs have gone crazy. We're kind of becoming a chip company as much as anything. Like Google has changed as a business so much and it's betting on so many things in so many directions and there's a perception that Gemini is behind all of that is okay as long as search continues to be search and any teeny tiny like hole in that case starts to look like a big scary problem for Google and so when they report these earnings they're like oh look we're going to spend $200 billion this year on AI related things and also our search business is not doing as well as you thought even by this much we're not doing as well as you And everybody panics because it's clear and Google even alluded to this in their earnings call that they see some pain in the transition to basically how do we make money from AI because surprise they don't know because people don't click the links. This is what I mean right like when I search best lawnmower Google relies on the fact that a lot of people will click the link to the sponsored site for the lawnmower. And if Google has built a thing that has trained people to not click any of the links, it has a problem. Like Google is creating a world in which its own best business model doesn't work anymore. And Google will tell you that people click the links. And I can tell you from my own experience and from all of these publishers and from Reddit, people aren't clicking the links. They're not. And they also hate the AI spam that is infecting all these sites because the spammers are trying to get themselves in AI summaries on Google. We have a big story coming out from me about AI spammers attacking Reddit every single day. I mean, I see it and Reddit mods having to fight back. I'm really excited for that story. I think that's going to come out next week. But you see the problem is the incentives to put information on the web are getting away from Google in real ways. And so the web is getting polluted. Google's best resource for ages has been their index of the web. Yeah. Right. That's the thing that they had that no one else can compete with. There's the Bing index of the web which isn't as good. No one else has an index of the web, right? That can compete. So Google search was the dominant winner and because they had the best index of the web, the Google search interface was the dominant winner. Whether or not you thought it was good or getting worse over time, even though it was getting worse over time because SEO was destroying the Google search interface, ChatGPT shows up in 2023 and says, 'What if we just told you the answer?' And I think this will go down as something Google really has to reckon with. They changed Google search to look almost exactly like its disruptive competitor. Yep. Right. So you enter an AI overview. Now you're just chatting. That is exactly the same interface as ChatGPT. It's exactly the same interface as Claude. It's exactly the same interface as Grok. Everyone is just giving you answers in pretty much the same format, stealing all the same information in exactly the same ways. Now, Google isn't differentiated because it doesn't have the best interface on top of the best index. It has the same style of output as everyone else, right? What are you going to do? And that index isn't getting better. It's actually getting more and more polluted with AI slop, right? And it's potentially going to start hemorrhaging some of the actually useful sources of information who no longer feel like it's a good trade. Like you can just see how this starts to turn on Google in kind of a death spiral kind of way that the worse Google search work gets the worse it will get. Do you know what I mean? Like that I believe you know the answer and here's my optimistic case. The answer for all these companies is to start paying for information in a way that keeps that information away from their rivals. Sure. So if all the publishers quit letting their sites get indexed by Google, the first company that will show up on their doorstop is OpenAI. You don't want to be indexed by Google. We will pay you a bunch of money to be indexed by our new web search agent, which will do whatever it's going to do. Great. You know who will pay more money is Google. Sure, they have more money. They're a more stable company. They can probably make the economics that work in different ways. But that is, I think, good for the ecosystem. But then you're putting yourself in the same position that Reddit is in now, which is to say, 'We've given you all of this data, and what you're giving us back is essentially neutering the traffic to our site.' No, what they're saying is what you're giving us back is not enough money. Maybe, maybe. But then what you're arguing is at the end of this road, everyone is just in the business of providing stuff to AI indexes. And I think that is bleak. And I hope that's not the answer, but there is probably money on the way down in that answer. And this is I mean this is the thing that's every publisher is going to have to stare hard at right now, right? Is do we want to take the money out on the way down like in the way that a lot of people took a lot of money from Facebook to make video that no one ever watched and it all eventually fell apart. Do we want to do that again knowing that at the end of this we have essentially obviated ourselves out of business by training people to use a different product to access our information or do we need to completely pivot the way that we have thought about what our thing is and how people find it and access it which for 20 years has not been the media's strong suit.

Yeah, David, you've met these guys. You know which one it is. I'm like Google Zero is coming and they're like no it's not. That is five years of my life. I started a news startup in 2019 and one of our very first conversations. A very important person who I will not name said, 'What if we don't even have a homepage? Who needs homepages anymore?' And I don't think we've learned a single lesson as an industry since that meeting in 2019. My one piece of tiny push back on this is they all did just get burned, right? These organizations, they all gotten burned. The thing came true. The thing they did not want to believe would happen. Absolutely. The traffic numbers are pretty staggering. There's a chart in that Wall Street Journal story showing how much people's traffic are down and it's like it is the sort of thing that has become absolutely unignorable for everyone in this industry. You might say that Google traffic has gone to zero. Somebody should really there's a term traffic zero maybe we'll call it. What do you think? But it happened so now they got to do something. And if the outcome here is it is economically viable to employ people to make the raw information that gets fed into AI engines that people use to summarize it all. At least there's an economic incentive to make new information. Right now there isn't really one. Like the information itself isn't valuable unless you're selling something. And the whole world being lists of great products to use on a hiking trip sponsored by REI is just the bleakest possible. Yes. So if we do get to a place where in some roundabout way of payments reporters exist, that might be fine. But I don't know which way it's going. I think we just might disagree on the degree to which it's optimistic to suggest that the best possible case for a reporter going forward is to feed data into ChatGPT. Even if it's new data. Do you know what I mean? I don't know that I find that as blissfully rose-colored as it sounds like you do. Well, if you know any media CEOs who are interested in product innovation, you give me a call. Looking forward to that. All right, we should take a break and then we've actually got a bunch more AI adjacent stuff to talk about. So, we'll get into that. We'll be right back.

Support for the show comes from MongoDB. AI assisted and agentic coding can help you build faster than ever. But if your data layer is a bottleneck, what's the point? MongoDB actually gets out of your way. MongoDB is a unified AI ready data platform that empowers you to build scalable generative AI applications. It eliminates the need for separate specialized vector databases by combining a flexible document model natively with semantic vector search, full text search, and real-time operational data. Instead of wrestling with rigid schemas or translating data formats, MongoDB's native data model mirrors the language LLMs already speak. Plus, MongoDB gives you the flexibility to ship at the speed of AI. The ACID compliance lets you sleep soundly at night and scales to handle massive Fortune 500 workloads. Developers have a word for that kind of reliability. Actually, five words. It's a great freaking database. Start building at mongodb.com/ai.

Support for the show comes from Rippling. Retaining your best employees should always be top of mind, especially if they are in high demand. And while it's impossible to predict the future, Rippling AI can help you identify who might be a flight risk. That's because Rippling AI is built on your live global workforce data, giving you full visibility into your business and the ability to stay ahead of the curve. Say you want to retain top talent. Just ask Rippling AI, 'Who are my top performers this year?' and instantly receive a workforce report highlighting your top employees with supporting data like recent performance reviews and engagement metrics. But it doesn't stop there. Rippling AI can turn these insights into a proposed retention strategy, including a recommended 10% spot bonus for top performers. All you have to do is tap confirm and the spot bonus is added to next month's pay run. So don't settle for AI that's all talk. Head to rippling.ai/verge and get AI that turns insight into action. That's R-I-P-P-L-I-N-G dot A-I slash V-E-R-G-E. Sign up for exclusive access today.

All right, we're back. So, we have a bunch of AI news to talk about. There's no sort of big broad theme here, but just a bunch of stuff going on. We should talk about it. Nilay, I genuinely need you to explain this first one to me because either there is something about this story I don't understand, or I'm going to melt down in an aggressive way on this podcast in a couple of minutes. Can you explain what happened between OpenAI and Hugging Face this week?

I do like the way we have to say Hugging Face a lot.

Hugging Face, a company I don't know anything about. What actually is Hugging Face?

Hugging Face is like a platform to store AI models in and it's like GitHub for AI. Okay. And that really implies that people know what GitHub is, but it's the first guest audience. So, do you know what I mean? It's a place to put things for AI. It's a place to put things and collaborate on AI stuff and lots and lots of people use it and is a central source of truth in the AI community. So there's Hugging Face. OpenAI is testing GPT-5.6 Soul on cyber benchmarks. There's one called xpoint gym and they basically said go figure it out find a solution for it and the model broke out of the sandbox it was in and it went and broke into Hugging Face and stole the answers. This is as reductive as it gets, but it stole the answers to the test and it's not supposed to do that. Hugging Face says its own AI agents detected and stopped the breach. They're now trying to spin this as a partnership or this is actually a partnership to solve it. Who knows? But they're investigating everything. They're going to bring it to some kind of conclusion. And you really have to either believe that the models are so capable they could evade OpenAI's own systems and that we're all in danger all the time or that maybe Hugging Face was too easily breached. There was no social engineering which is like a good and important thing to note if you're the sort of person who cares about the model like tricking the CEO of Hugging Face into inputting his password. Yeah. Okay. But here's the summary here. In one example, the model chained together multiple attack vectors including stolen credentials and zero-day vulnerabilities to find a remote code execution path on Hugging Face servers. It stole answers for the test. How much you want to overreact to this is like it is up to you. Can I tell you my personal choice? Yeah.

A lot. How is this not like a screaming five-alarm fire about what these models are capable of and what they can do? Hugging Face is saying, 'Oh, isn't this adorable? They came and hacked our system and stole the answers even though we were just trying to prevent you from doing so.' And OpenAI is like, 'Oh, look how adorable. We hacked into and stole their information.' And it's like we're just supposed to think this is like a normal okay occurrence because everybody is just trying to laugh it off and pretend it's fine. This to me is absolute disaster on all sides. This means Hugging Face has a problem. This means OpenAI has a problem. This means these things can do the things that they keep promising us they can stop them from doing. Like this is they kept putting guardrails on Mythos and calling it Fable. And everybody's like, well, we can just make it Mythos again. Like the idea that these models can either be put on rails by the company that makes them or forced into a box by the companies that employ them is clearly not correct. And for all of us to pretend that this is just a delightful silly thing that happened is not okay. Like it turns out I do understand this story and everyone needs to be angrier about it. Like do you want me to make it worse or do you want me to make it better? Let's go with worse and then better. Let's end in a happy place for David. All right. So worse is it is true that the models keep getting out right like OpenAI was not in control of its own model and they didn't even detect that it was doing this even though it was in a sandbox and by OpenAI's account which we have to trust the sandbox had no external connections like aside from being able to download packages to the sandbox the model should not have been able to have network access and they blew it just straightforwardly that that's bad that is bad. And so we gave it nothing and it did it anyway. Is not good, right? And then you can make it even worse. You say this is what you would call the paperclip problem where you like tell the machine to make a paperclip and it vaporizes the human race and turns it all into paperclips because it's just doing what it's told which is to get the solution to the problem. Which by the way it cheated. And I just I do want to point out that it cheated on the test. It didn't look at the problems and solve the problems. It was like, I can probably steal the answers and it went and did that anyway. So, this is all bad. Yeah. Okay. Here's how I can try to make it better. Okay. This model in the testing environment had no guardrails at all. So, this it would not be the model that is deployed to any normal people or any normal customer outside a very select set of security researchers or whatever we're going to come to in whatever regulatory framework we'll eventually come to. So it is true that OpenAI this all happened but this model is the guardrail free model. I don't know if I made this better because so we'll So what you're saying is this will all be fine once we have a completely rational and thorough regulatory system for artificial intelligence in like I said I don't know that I made it better. I'm just pointing out that this is the model without any guardrails. If I'm the CEO of Hugging Face, am I not sitting in my office going, 'Gosh, it's good that Soul didn't decide that the best thing to do was burn my office down in order to get like I just how is the either everybody is lying and these things are vastly less capable than we think and everybody is just trying to make marketing moments out of cyber security disasters, which is weird but sort of a thing, or everyone should be more alarmed by this than everyone seems to be.' And I really think both of these companies are trying to pretend that this was not a terrifying overreach by an AI model and a terrifying lack of protective ability by Hugging Face. And everybody has real incentive to try and save face here. So they're trying to just like lock arms and laugh it off. This feels bad to me. Maybe I'm missing something. If I am missing something and you are a more sophisticated AI person than I am, I would love to hear from you. Vergecast.com. Tell me why I shouldn't be at like complete run around the neighborhood screaming levels of panic about this because this is where I am. Enough with this one. I'm feeling bad enough as it is. Let's move on to another AI thing that frankly won't work nearly as well. Which is that Substack added an AI detector to theoretically figure out which of your many many newsletters have been written by AI. I should say it's powered by Pangram, which we had a whole Vergecast about last week with Jake Castanakis. Really good. Pangram is sort of the one tool in this space that most people seem to agree actually works at detecting AI. My immediate reaction to this from Substack is just kind of what a self-own to have to roll this thing out on your platform. But what do you think about this, Nilay? All the distributors of content, all the platforms, if you will, are stuck in a trap where they're getting flooded with AI slop and many of the platform owners are also purveyors of AI or invested in AI companies or have taken an investment from VCs that are heavy on AI and they're stuck in the middle. Right? If you're Google and you run YouTube and YouTube is being overrun with slop, you have the finest of lines to walk. You cannot say the output of Google's AI tools are bad, but everyone hates the

Slop. Meta is caught in this trap. Substack, the startup — I have a lot of feelings about Substack, but all their VCs are like heavy AI guys and undoubtedly publishing AI slop. This is the challenge for every platform right now: you have to somehow embrace the fact that AI is coming and people want to use AI in whatever way they want to use it, and people hate the slop.

Yes. And you have to detect it and you have to tell people to turn it off, and what are you gonna do?

And they keep trying to pretend that it's not about getting rid of the stuff that sucks that no one wants, but it's just — it's about transparency and helping people understand the difference. Like, here's a quote from Chris Best, the CEO of Substack. He says, "The core problem is not people using AI or the quality of its output." Strong disagree, but let's keep going. "The problem is when there is a mismatch between a reader's expectation and reality, especially when they unwillingly invest their attention in something with no human thought on the other end." That's Claude fishing. I want to say a Claude fishing is an excellent term and kudos to Chris Best.

I'll throw one to Chris. He and I have beefed, but I'll give him — Claude fishing is excellent and I will use it. This is not correct, right? Like the idea that actually what people want is sort of separation of two good things is not at all actual reality. What people want is a way to not have AI in their feeds. It is — people don't want an AI feed and a not AI feed. Sora did a really good test of if people want an AI feed and the answer was they don't. Meta did a really good test of whether people want an AI feed and the answer was they don't. People want a non-AI feed, and all of these things are designed to get AI out, but for all the reasons you just said, they can't say that and they're stuck.

Let me ask a question of you and, by extension, Chris Best, the CEO of Substack. He's here right now. I can get him. What is an automated system to delete content that people find offensive if not content moderation? Chris, would you call that content moderation? Uh-huh. Would you say that it's weird that you are more permissive of outright racism and Nazis on your platform than the output made by Claude? Oh boy. The core problem is not people being Nazis or the quality of their output. The problem is when there's a mismatch between a reader's expectation and — you can just put it all right there. What are we describing? Is it content moderation? Is it the ability to detect certain kinds of content and make sure it does not find an organic audience because the people don't like it?

This is a really interesting question, actually. So, Substack relatively famously has been very hands-off on this front and — famously to your face. Yeah. A clip that has been seen by every person on TikTok at this point. But look, I sincerely think reasonable people can disagree on how Substack should moderate the people who write newsletters on its platform. Feel that way however you want. The idea that it was actually willing to take a step this drastic, and in a time like this with technology like this — this is a drastic step to say we are going to loudly announce that something is AI — is a big step. What do you think it means that Substack actually pushed over the line to say we are going to do this in a way that they have not been willing to before?

One, I am not as high on Pangram as everyone else is. Whether or not they can actually detect AI text at scale over time using more AI is just a big question mark. I saw a thing on Threads the other day that was like, here are 12 prompts to make your AI writing sound less like AI. See what I'm saying? It's just like, oh, right, we're just — this is the cat-and-mouse game forever and ever and ever and ever. This is just one problem. Maybe more existentially and more depressingly, the market has spoken, and it is clear that the market will reject AI content more vigorously than it will reject outright racism and outright sexism. I don't know how to feel about that, right? Like, maybe you should just let it all ride. I have a lot of feelings about Substack just surfacing Nazis over and over again, which is a problem that just keeps repeating itself. And just to be totally blunt, Substack has built very, very good content moderation systems. They have them for a lot of reasons. They have them for copyright infringement purposes. They have them because there is some stuff they will not put in their recommendation system. They've built it all. It's all there. And now they're using it to filter and moderate AI content. They've just added another category to their existing content moderation infrastructure. And the market has not demanded that they use that infrastructure to moderate outright racism. And I, you know, maybe that's just the internet, maybe that's the world we live in. Maybe that is 2026 and dark woke is going to make its comeback whenever it makes its comeback because pendulums do swing back. But this part of the information ecosystem is much more tolerant of one thing than another. And I do think it's very notable that an ecosystem that will tolerate some horrible speech seems to have no tolerance for AI. And that is, you know, you talk about data center protests and the negative public perception towards AI — there it is again. We will have outright racism and sexism on these platforms that will occur with no penalty to them and the usage numbers will keep going up. They will steal all of your information. They will invade your privacy. All that stuff. Targeted advertising will come at you. They will have problems with children and all the stuff that they do. Nothing. None of that hurt them. But people hate AI content so much that we're going to build in the filters. Huh.

Yeah. Can I offer you the funniest possible counterexample? Sure. What if you did this in the exact opposite direction? That's what's going on at Prime Video at Amazon right now. So there was this great Reuters story that we quick-posted about earlier this week, basically saying that they're going through a big reorganization of the way that Prime Video, the actual platform, works, and rethinking the way the homepage works, rethinking the discovery systems, the personalization systems. And apparently Jeff Bezos himself has shown up and demanded — from what it seems — just more AI. He says do more AI to it to the Prime Video team appears to be what's happening here. Like make the interface — vibe code the interface. No, I think what's happening is they seem to be wanting to make it much more AI-programmed in a really interesting way. Like use more AI, presumably Amazon's AI, to do more personalization, to do more algorithmic recommendations. Presumably there will be a million Alexa-powered things on the — I don't know. He just — like there's a thing we posted that there's an image that Prime Video is using to promote Death Race 2000, which is very obviously AI-generated. So maybe that's the future too, where instead of — we're going to get personalized posters for every title generated with AI. Like this really has like Sergey Brin coming back to Google and demanding to AI everything energy to it, where Bezos is like, take Prime Video and AI it, and we just — what if we put more AI in front of people, which in a moment where AI slop is filling up all of these platforms everywhere and people are aggressively fighting against the use of AI, particularly in movies and shows, this is a very funny energy to be putting into your streaming service right now.

There's a Lauren Sanchez joke here and I can't find it. And maybe I don't want to. Maybe the future of this is every one of the posters is AI-generated and Lauren Sanchez is in every one of them. She's flying a helicopter over, right? I'm going to — I'll set that aside. You can send me the jokes if you want. It does have that AI has a marketing problem energy to it. People would like us better if we would just show them more AI, which I hear from every corner of this industry. And it also — there's probably some amount of Prime Video's interface is already just like a carnival. Like it's already just trying to sell you so much stuff, it might as well just try to sell you some more stuff. Amazon has always just been the least concerned about taste or class in its video streaming interface. Like it is just the most capitalistic hellhole that has ever existed in the history of television. And so of course he's like, I don't know, use AI to do more recommendations of the last thing you saw on Amazon.com and make that shoppable. I could not tell you what this actually means, but it is very good that it's just going to be loud and stupid all at once. And maybe that's the joke. I do think there is a real energy behind a lot of this stuff that suggests all of these people think that all we need to see is more examples of how AI is already making our lives better, and what if you were just confronted with how great AI is more often, you would learn to love AI. And I just don't think that's true. Like, I mean, the confidence with which this is repeatedly said to me is off the charts. If only people could see it — if you had three great use cases and then you're an AI convert for life. And it's like, have you looked at the polling data? I know. Well, and I think about this next to things like Letterboxd, which there have been all these rumors that somebody's going to buy Letterboxd recently. There is this big push happening towards recommendations for movies and TV shows and people wanting them to be human, right? Like staff picks are coming back. People want the experience of having another person tell them what is good. Letterboxd is so important because it has lots of data from actual people telling you what is good. And there is just this rebellion happening against what I don't want is a personalized list of movies that are perfectly tuned to me and then actually movies that you made because of that data you have about how to perfectly tune them to me. People understand the pipeline that is feeding this slop to them and they want out of it. And so what we're getting is just louder AI recommendations from Amazon Prime Video. Sure, why not? Knock yourself out.

I do like that he looked at all of Amazon. He's like, I know where to put the AI. What's the ugliest homepage we have? How do I put some more stuff on it? Do you want to talk about this Meta AI detection system before we go? I know you love an AI detection system. Do you want to talk about this one before we switch gears? I mean, it's the theme of our little segment here is can you find it and can you kill it? Can you — can you predator the AI? Could Content Seal, Meta's tool, have stopped 5.6 Soul? I'm going to say probably not. No, definitely not. So shout out to Jess Weatherbread who is on our team. I long ago told Jess that her beat was creators and creativity and creative tools, which in the age of AI has meant she knows more about the systems to detect AI content than almost anybody alive. Yep. Apart from the people who are making the systems. But I think that she knows more about these systems than the people who made Content Seal at Meta. You can go read her story. She has spent more time reporting on C2PA and Content Authenticity Initiative and Google's SynthID. She is in the weeds of these things. Her reporting is great, and she saw that Meta was announcing this, and I think we have the only story that's like, what? Because she understood that the only rational response is what? So the entire industry is basically coalescing on Google's SynthID, which is an invisible watermark over images and videos that is pretty hard to strip away, and a lot of people want to use SynthID. So Apple's going to use it, they announced it at WWDC. OpenAI is going to use it. There's another standard called C2PA that promotes content credentials. Meta is on the board of C2PA. So's Google, right? So's Google. Yeah. I mean, these are like evolving. There's a lot of complication about how they work together. Doesn't matter. There's just a lot of this going on in the world already. Standards that exist. And a lot of the arguments for these standards are they needed to exist. So they exist. We're all going to use them. They all get better, right? And importantly, they're only really valuable if the distributors of the content, the platforms, use them, right? Which is an important reason to only have a couple of standards, right? So everybody understands that all the images are being labeled the same way and then the content provenance is being labeled the same way. Like it doesn't matter if you open up X and you don't see the label. Then it doesn't matter if the standard exists or not, and maybe X will never show the labels, and like here we are. This is the whole problem with these standards. Anyway, Meta introduced a new one called Content Seal and it does not work with any of the existing standards. It doesn't work like SynthID. Meta has not built any of the tooling for people to detect it. Meta's own detector is rate-limited. So even if you get images that are labeled in Content Seal and you try to detect them, they will stop you, and that — you partially — this is for security so people can't reverse-engineer it. They're not using it on their own platforms. It is just — it's malicious compliance. This is like Stellantis making hybrid Jeeps to sell gas Jeeps in Europe. Like it is that thing exactly. It's the Ferrari Lucho of content standards. Brutal. Do you know what I mean? Yeah. Like look, we made an EV. We're so proud of it. It's to get around EU regulations. 100% what Content Seal is. Super cynical. Meta answered some of the questions on the record. Didn't answer some of her other questions on the record. Why are we doing this instead of using the already existing standard? It's because Meta does not want its AI images to get detected and filtered by other people's systems, which is what will happen if you give people the tools to turn the AI off. Yep. I think Meta is also not super eager to filter it in its own systems. Like Adam Mosseri has been on the record saying we don't want to filter out AI content, that they are also, like Substack, interested in letting you know what's what but not letting you filter it out. And if you actually do a good job of telling people what's what, they will filter it out and they will demand a way to filter it out. And so Meta is actually in a weird way sort of incentivized to not have a system that works. And this is a particular problem for Instagram because every photographer of any note uses tools that now have a bunch of AI in them. Not even necessarily generative AI. If you use denoise in Lightroom, you're right up against the line of what AI is. And you should probably label it because the AI is inventing some pixels, but it is just denoising an image. Yeah. Should you be punished for denoising an image using the most state-of-the-art denoising tool that exists? I don't know. I truly don't. I do it without any regard for anything anymore. Like maybe — should every piece of advertising content ever to be made in the modern era be labeled as AI-generated because it has all gone through Photoshop and someone somewhere has undoubtedly used generative remove? Well, and increasingly all of those companies are just going to happily make AI slop advertisements about everything all the time. And in fact, that is fundamentally the new business model for Meta and Google and everybody else is just make creative with AI and target it with AI and we'll solve all of your problems with AI. I'm going to tell you the CEO of Adobe, Shantanu Narayen, who's retiring, and he should come back on Decoder and tell me who's going to be next. That'd be great. He was on Decoder and he said the generative fill and remove tools in Photoshop are being used at the same rate as layers, which means literally everyone is using them. Everyone is using them all the time. I mean, that's their stat. You can believe it or not, but that is what Adobe is claiming. And this is the problem for all these labeling tools because if you over-apply them, everyone gets mad. If you under-apply them, everyone gets mad. And Meta is like, what if we made a fake standard no one uses and then claim it exists? Pretty cynical. Yeah, it's rough out there. Where is the AI goes up right with where is the computer in foundational Vergecast questions going forward. All right, that's enough AI for now. We've got a bunch of gadgets and stuff to talk about. Let's take one more break. We'll be back.

Support for the show comes from Bitwarden. Quick gut check: how many of your passwords are actually strong and unique? If you had to think about that for more than a second, today's sponsor is for you. Bitwarden is the trusted open-source solution for passkey, password, and secrets management, built to make secure password management accessible to everyone. With its free-forever plan for individuals, you can create, securely store, and autofill strong, unique passwords across all your accounts and devices, so you never have to remember dozens of complicated logins or rely on the same password everywhere. It also makes signing in faster by securely autofilling passkeys, passwords, payment cards, and identities on virtually any device. Behind it all is end-to-end encryption that helps protect your information from phishing attacks and data breaches. Get started with a basic free plan for individuals or protect your business today with a free trial. Visit bitwarden.com/vergecast to get started today.

Support for the show comes from Upwork. I have bad news for you: there's just no way to grow your business if you're doing everything yourself. But I also have really good news for you. Upwork is a one-stop platform to find, hire, and pay expert freelancers. It gives you fast access to specialized talent across more than 125 categories. That covers everything from web and software development to data and analytics to marketing, business operations, and more. With Business Plus, you can access the top 1% of talent on Upwork. And with AI-powered short listings, you'll get matched to the right freelancer in under 6 hours. That means you can hire with confidence and get started quickly. It's free to sign up and posting a job is easy. Visit upwork.com right now and post your job for free. That is upwork.com to connect with top talent ready to help your business grow. That's u-p-w-o-r-k.com. Upwork.com.

All right, we're back. It's time now for the hype desk, where our friends Ross and Ashley come and tell us about what's cool in the world. This one is unsponsored, which I'm telling you because at some point in the near future that will not be true, but for now it is unsponsored. And today it's just Ross. Ross, welcome back.

Welcome to Flavor Town. How you doing? I'm Guy Fieri. Diners, Drive-Ins and Dives, Devices. We got it all. Anyway, now we're getting sued.

Diners, drive-ins, dives, and devices. There's something there, you guys. There's something there.

This is your hot one, so you're welcome.

I'm dying.

We're going to take all that sponsor money and we're going to pay it directly to Guy Fieri.

What are we doing here?

There it is. If you'd like to sponsor us, that'd be great.

I don't know if the Food Network is buying podcast advertising at this time. We're open to it.

Welcome to Flavor Town is a thing I am willing to say every week for enough money. Ross, what do you have for us this week?

Nope.

That was not the pitch I received when this began.

I mean, can we just, can we, can we just pause on this for just...

Yes. So, one could call it Infinity Vision. One could also call it lies. Is that, is that a fair...

Well, to me, IMAX right now is, is like an ever escalating series of lies, right? There's like proper IMAXs and there's like fake IMAX and there's like levels of fake IMAX.

Then there's Zack Snyder IMAX, which is just 4:3 grayscale on HBO Max.

I don't know, man. It's... Go ahead. Tell me what Infinity Vision is.

This is what I mean by it's like true IMAX is what, 40 theaters in the whole world, something like that.

Didn't you say you were going to make a case?

It, it does, because the trailer in many ways feels like homework, right? It's like, remember this, remember this...

Chris, it's Ready Player One, the Avengers Experience. Do you remember this guy? What about this guy? Do you remember this guy?

Listen to this music. Needle drop this guy. Like the whole...

And at the end, another guy appears. Oh my god.

Let's not forget it's Robert Downey Jr. practicing this kind of faux Eastern European in-universe. It's called Latverian. It's a Latverian accent.

It was wonderful. It was a great experience for everyone.

It, it is. I just want to point out they already sent off the Fox X-Men in Deadpool. That literally, that movie was like goodbye to this.

Well, so wait, so this is actually my question.

There was a montage sequence at the end that was literally goodbye to this, Ross.

I think they ran Time of Your Life.

They, they did actually. Yes, they did.

You can't come back from time. You can't ungraduate from high school.

Well, okay. Professor Xavier has died on camera for a film like eight times at this point.

Wait, I'm sorry. The argument that I heard here was they did a fake IMAX and a lot of people watch this trailer. Wait. So, okay, Ross, let me, let me ask you a question that I think is, is at the center of what I'm trying to figure out here, which is they tried to do this thing where there was one era of Marvel that, that ended and then there was going to be a new era of Marvel with new stars and new characters and new movies and that, that none of that worked. Like, it didn't, it, it just, it all sort of died with, with the original characters. And this movie seems like a, a good idea in the sense that it is an admission that that first era everybody liked and kind of nobody came along with them into the second era they were trying to build. So they're just going back to the first thing. They paid everybody a lot of money to go make the old kind of Marvel movie with the old characters that everybody liked. It's probably going to work, right? Like this, it, it feels sort of cynical and sad that this is what it's come to, that like it turns out everybody liked Robert Downey Jr., so we're going to give you more Robert Downey Jr. Is it as simple as that? Like we're, we're just going to pretend that the sort of second era post all of these characters didn't exist and just go back to them again at much higher salaries. Jesus, $500 million to do two movies. Unreal. Congrats, Robert. Congrats.

How many times must we kill the X-Men? We're like, we're sending back in time now. They're young X-Men. We're going to kill them, too. Old X-Men and young X-Men die in the same like...

I will say just in defense of Robert Downey Jr., yes, he talks real fast and so words per minute, it's a deal.

Yeah, I appreciate the, the Doomsday spec sheet here. As a fellow fast talker, that's kind of the...

To me the whole thing here is like just to go all the way back to the beginning, the whole reason Infinity Vision exists is because Marvel lost all of the IMAX screens to Dune Part 3, which is going to be a vastly better and more interesting movie that I think is going to absolutely just eat Doomsday's lunch.

Oh, that is a take. No, no, no way. We're way over. We got to be done here.

But no way.

Yep. It's going to be. Are you excited about Spider-Man, Ross? Brand new day.

I, I, I love the optimism. And just so everyone knows, if you pre-order tickets for Avengers Doomsday now and then the movie is terrible, Guy Fieri will personally refund you your ticket. He's going to sponsor this show by buying everyone tickets to Avengers Doomsday. It's going to be amazing.

I did always wonder what would bring The Verge to its knees. And the Guy Fieri lawsuit was a dark horse.

Here it comes. I'm ready for it. This is the way I always wanted to go. All right, Ross. Good to see you. Thanks for coming. Cheers, y'all. All right, now it is time for America's favorite podcast within a podcast. Brendan Carr is a dummy.

Oh, yeah. All right. That just gave me very intense — and I mean this is a compliment — day rave at a Vegas pool energy in a way that I really enjoyed. Oh, I had Lowrider Cruise. I think it's — I think that's roughly the same energy. It's in there. You're like — you're bouncing up and down in a Camaro. I'm like kicking it in a cabana. Yeah. Your knowledge of lowrider culture is completely just like offensive. I don't even — what — I'm very sorry. That theme remix was from Chris Swick who said — and this is a very good part of this email — he says no AI was used. Not a fan of AI. Everything was done in GarageBand. Big fan of GarageBand. We love to see it. That's very good. Neil, what did Brendan do this week? Brendan, man, we actually have a clip this week. So, if you'll recall, last Thursday, President Trump gave a sort of rambly, incoherent, late-night speech about election interference. Sure. You know, and then he referenced a bunch of documents. Everyone looked at the documents and even the New York Times is like — president lied about documents, which is like a big step for them. Like none of it made any sense, but the thrust of it is that the elections were interfered with and he should have won in 2020 even though that all happened and he was in charge of the government in 2020, which is just a real fact no one can hold on to. Who was the president in 2020 is a real question that the Republicans of our country cannot deal with. Which has led to a full week of John Oliver TikToks. Have you seen these? Just him grilling people about whether Joe Biden won the election in 2020 and everyone obfuscates and he just asks them the question. It's very good. Anyway, so that was the speech. Many networks covered it, many did not. Like even like Fox News the next day basically was like, well, that happened. That was weird. And they moved on to the next thing. Anyhow, as is their right,

Right? They have First Amendment rights programs.

They're also just like on, I think, in Republican and also just like in the Republican party, this is not a winner. No.

Like relitigating 2020 is not a thing they all want to do. It doesn't win them elections. In fact, sometimes it loses them elections. Anyhow, some networks covered the speech in prime time, some networks didn't. NBC and ABC did not. Okay. NBC and ABC are broadcasters. Guess who loves to infringe the speech of broadcasters? Our boy Brendan Carr, whose FCC nominally controls the rights of broadcasters using the public airwaves. So, he was asked on Fox News this week if it was acceptable for NBC and ABC to not air Trump's remarks in prime time. Here's what he said.

Air the president's remark, and I said, 'Of course they should.' I'm Brendan Carr. I'm the chairman of the Federal Communications Commission, and they should do so because they have an obligation at the end of the day to operate in the public interest, to serve the needs of their communities. And I think making sure that people have access to significant presidential addresses is part and parcel of what all broadcasters should want to do. There's been a lot of interest in Congress. Obviously, we got a letter from Congress, you know, asking us to investigate the decisions to do this or not do this, and you know, we're reviewing the letter itself at least at this point, and there's been a lot of interest in this. So, we'll see where it goes.

Okay. I have several thoughts.

By all means. Thought number one, Brendan Carr recorded this video on like an iPhone 3G from the quality of the thing. Two, do you think he got an actual letter? Like, do you think someone nailed a letter?

There's some Republican backbencher who's like, 'You should arrest Mickey Mouse for not running Donald Trump.' And maybe like, sure, it doesn't matter. This is all just like pretext to can I punish you, right, for speech I don't like. This was thought number three, is that this is very much reminiscent of Brendan Carr saying we can do this the easy way or he can do this the hard way. It's very clear he has no moves here. Like he's not shy about saying what he's going to do when he thinks he's going to do something. But in this case he just kind of gestures at it. He's just like, 'Huh, we're going to look at this,' with sort of a meaningful look on his face, which is very Brendan. I mean, he can definitely incur costs and launch sham investigations. Sure. And bring the wrath of the federal government down on various speakers for not uncritically rebroadcasting the president's remarks on election interference. Do you want to live in a country where definitionally the government says it is in your interest to listen to the president?

No. That's not the country I was raised in. Being able to ignore the president is one of the great joys of being an American. Being able to just openly be like, 'No, I'm choosing to never listen to you.'

Yeah.

Is maybe the only quality of being an American that we all share.

It's a very funny time to remember that like we mostly didn't used to think about the president.

Yeah. Screw that guy. Who is he? Get out of my face. Shut up. I don't have to listen to you. Is the most important thing. And right next to that, if you listen to what he says, being able to provide access to the important presidential addresses is certainly in the public interest. There's so much access to everything Donald Trump says.

Yeah, there's so much access that Truth Social announced a new business line where they will give Wall Street traders early access to Truth Social posts because you can trade ahead of the markets based on whatever Trump says, which is just a remarkable kind of open corruption. But like that's how much access there is.