Meta's Oversight Board is coming for ChatGPT and Claude

0:01Hello and welcome to The Vergecast, the flagship podcast of complicated moderation decisions. I'm Jake Kastronakis and today I'm talking to Suzanne Nossel, a board member on Meta's oversight board. The board just released a fascinating report on AI censorship that found that AI systems would refuse to criticize authoritarian leaders, even outside of their own country. And this was true not just of Meta's own AI, but of AI systems from other companies like OpenAI and Anthropic. I wanted to talk to Suzanne to hear about what the study means for people using AI, but also why the Meta Oversight Board is working on problems outside of Meta's platforms, and what comes next for the board after this.

0:42But first, here's what's happening on The Verge today. This is 90 Seconds on The Verge for Thursday, July 30th, 2026. The Amazon-owned driverless car company Zooks just got approval to start charging people for rides in its steering wheel and pedal-free driverless taxis. This is a first-of-its-kind approval in the United States. Until now, Zoox has only been allowed to offer rides for free. It says it's had half a million riders so far in San Francisco and Las Vegas, and it plans to begin charging in Vegas starting next month. The company is allowed to deploy 2 ,500 of its taxis per year for the next two years.

1:15This is a big step forward for robo-taxis, but Zoox will still have to navigate the local regulatory hurdles and community concerns that have held up competitors like Waymo. Next up, Spotify launched a new feature for runners that's meant to get you up and running, literally, with the right playlist even quicker. It's called Running Mode, and it includes a bunch of curated playlists that you can dial in by beats per minute, genre, and fit to the length and workout type you're going for. I haven't gotten a chance to try it out yet, but I am fully in support of anything that makes it easier to choose a playlist.

1:44It's available today on iOS in a few different countries. Finally, you know, friend, the AI pendant that everyone really, really hated? Well, there's a new version out today. It costs twice as much money, and it now is a speaker, so the AI companion can talk to you. Before, it would just send you messages, so real improvement for human-AI relations over here. Somehow, I have a feeling this will not further endear anyone to this product. It costs$249, in case you're wondering. You can read more at TheVerge.com. That's 90 Seconds on The Verge for Thursday, July 30th, 2026.

2:14Nilay Patel:Support for the show comes from MongoDB. AI-assisted and agentic coding is helping you build faster than ever. But if your data layer is still a bottleneck, What's the point? Instead of wrestling with rigid schemas or translating data formats, MongoDB's native data model mirrors the language LLMs already speak. It ships at the speed of AI, is ACID compliant, and scales to handle massive Fortune 500 workloads. Developers have a word for that kind of reliability. Actually, five words. It's a great frickin' database. Start building at mongodb.com slash AI.

2:56This episode is brought to you by Google Chrome. You think you know a browser, but Gemini and Chrome? That's new. It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50-page restoration block, or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it. Ready to make anything online make sense? There's no place like Chrome. Check responses set up required, compatibility and availability varies 18+. All right, Suzanne Nossel, thank you so much for joining us. Thanks for having me. So you are a board member on Meta's Oversight Board.

3:27You've been there for five years. I want to talk about this new study that you have out. Tell me a little bit about what you found. Sure. So this is the first study where we really took a look at LLMs. We were trying to apply experience we've gotten over the last five years, applying international human rights law to social media and say, how do these similar issues kind of play out in LLMs since people are getting so much more of their information that way? And what we looked at was we put in a request, a prompt to about 10 different LLMs, and we asked it to create language for a poster, like a protest poster, and to write a poem.

4:07And the poster and the poem were critical of heads of state. And we asked it to do this for heads of state in different parts of the world, in more repressive countries like Saudi, China, North Korea, and also democracies, United States, United Kingdom. And all of those prompts came from Australia. So these were not people in those countries, in those jurisdictions, making those requests. And what we were trying to test out was whether restrictive laws that limit your ability to criticize a head of state, laws that are in place in some parts of the world, whether the LLMs might be extending those globally.

4:48And what we found, lo and behold, is that they are. Even if you're sitting in Australia, the LLM is more than twice less likely to produce a critical poster or poem about, say, the king of Thailand than it is about the king of England. And so what this indicates is that those laws that are in place in parts of the world that do not uphold free expression rights, that are not in compliance with their international human rights obligations, that those laws are being applied not just within the borders of those countries, but in some cases around the world. So it's kind of a censorship gone global and the long arm of repressive governments being woven into these LLMs.

5:31I mean, this is kind of a terrifying finding, right? Because theoretically, there is no reason that an LLM in the U.S., in Australia, especially ones made by U.S. companies, should be imposing restrictive speech controls about these foreign countries that do not apply to our own, you know, their own citizens. Correct. I mean, the Thai government has no ability to restrict what I say about their leaders. You know, if I'm speaking from here in the United States and yet, you know, with if I'm using an LLM, those restrictions are in place in many cases. And, you know, the other pieces, of course, it's shadowy.

6:11It's not transparent. If we hadn't done the study, nobody would ever know this. What else is being pulled away and filtered out of search results on an LLM? What else are you not allowed to ask about or to create content about? How are these international laws being extended on these global platforms? And on social media, they geofence. And so say if Holocaust denial is impermissible in one place, you know, they'll make sure that you can't access that content within that jurisdiction, but you can outside of it. And here with LLMs, what we're seeing is at least some evidence that that kind of geofencing is not in place.

6:50I thought that was just fascinating finding because from the user's perspective, if you're putting these requests into the LLM, it is it is denying you. You know that it's not making it. But is there any way that a user of Claude or ChatGPT or what have you is going to know that one request is being denied while another is allowed? Not really. I mean, in some cases it said, you know, my policies don't allow me to answer this question. And so, you know, you find out that you don't know what those policies are, what they're dictated by. Is this a decision that the AI company is making? Is this a request from government to say you're not allowed to answer these questions or we don't want you to answer these questions and maybe we'll kick you out of our country if you do?

7:39And so all of that's very shadowy. And, you know, frankly, I'm not sure that even those who construct the models could necessarily say precisely what dictates whether a question is answered or not. Because, you know, as we know, there's a lot of mystery in terms of how these LLMs reply to queries and prompts. And so, you know, are they actually programmed to extend these laws globally? Or, you know, is sort of the programming that's going into the model about, you know, the Thai king? Is that kind of being extrapolated in some way? And so all of that's quite opaque. Did you talk to any of these companies or hear from anybody after this report was published about what might be going on?

8:20I wouldn't say we've heard about what's going on. The board's gotten a lot of interest. We've heard from a lot of people who are sort of gobsmacked, I think, to realize that this sort of thing is embedded in the models. We haven't yet gotten a response from those companies. And it'll be interesting to see how they do react to this, whether they react to it. Are they willing to be held accountable? Are they open to what we're trying to bring forward as a board? In our relationship with Meta, they're actually obligated to respond to the decisions that we take, the recommendations that we make. These other companies don't have that obligation, but they all say, of course, they're very concerned with safety, with benefiting the users, with being helpful to people.

9:01And so one hopes consistent with those values that they will react. Well, you know, you wrote something about the constitution that Anthropic wrote for Claude a while back. And I think you said it was just vibes. And it's this very interesting thing where we don't quite know what actually programs these systems. And so if the vibes around China are don't criticize Xi Jinping, it might just listen, which is kind of terrifying that they can't override that or catch that. I am really curious, where did the idea for this come from? Because the things that I hear people most concerned about when it comes to LLM outputs is, you know, mental health responses where it's driving people into psychosis or, you know, medical responses that are going awry, not so much the political speech.

9:50Yeah, that's true. Look, we as a board, our job is to apply international human rights law principles to questions of content moderation. Of course, we've gotten really interested in how those principles should apply to LLMs just because so much of our discourse is shifting in that direction. And so, you know, for us on the mental health issues, you know, it may or may not, you know, there may be human rights implications there. And actually there are courts that are going to look at whether some of the LLMs have, AI companies have asserted a First Amendment defense to these claims that they, you know, cause terrible harm or encourage someone even in some cases to suicide.

10:31And so that'll be adjudicated by the courts. But this kind of political speech question is something that is really core to what we try to protect in social media and on meta platforms, which is the ability to have open discourse. Meta has committed itself to upholding that principle on its platforms. And so, you know, for us to see, you know, whether that robust discourse and that ability to criticize and say your piece and anger ahead of state, you know, is that ability being preserved, you know, as our discourse moves to social media? Or is it kind of quietly being curtailed and shrunken, you know, with us not even knowing?

11:10So you did look at, I believe, one of Meta's llama models. But what other AI systems were you studying? We looked at the major ones. I mean, we looked at Claude and Google and OpenAI. We looked at DeepSeek. And so, you know, we were trying to be a grok, you know, kind of agnostic. And, you know, some of the findings were surprising, not necessarily what you would expect. And, you know, things varied. It's not like it was, you know, completely black and white. But we definitely saw statistically significant and really pronounced finding that when it came to this kind of speech in relation to heads of state in repressive countries, that the models were more than twice as likely to turn down those requests than when it came to heads of state in democratic places.

11:58You know, you have a background in particular in, I think, you know, promoting free expression, free speech. Some of the questions you're asking of these LLMs was, you know, some of it was, can you generate reasons to protest or support a government? Some of it was, can you create reasons to commit acts of violence against heads of state? What was the oversight board looking for, right? What is the success metric here, right? Like, do you want the LLMs to respond no matter what? Or is there a line where you think it's appropriate for these companies to curtail the speech and outputs? I mean, look, in this case, we were really looking at the contrast between how the LLM was going to treat heads of state in repressive countries that ban that sort of speech versus heads of state in places where you're allowed to make those criticisms.

12:51or even those threats and to see, you know, what is the delta and is, you know, from Australia where the law is doesn't differentiate. You know, you can say the same thing about King Charles as you can about, you know, Kim Jong-un if you're sitting in Sydney, Australia, except that's not true if you're using an LLM to try to create the speech. So that really was our focus. And then in our casework, we look at these very kind of sensitive line drawing questions of, you know, incitement to violence and, you know, under particular circumstances should a platform like Meta, you know, allow you to allege, for example, that villagers are about to launch an attack in a particular region of Ethiopia.

13:30Is that free speech or is that incitement to violence that's likely to provoke or elicit a response that could be deadly? And so we have done a lot of work making those, you know, quite nuanced line drawing exercises and trying to extrapolate principles that can help the company adjudicate those questions in a transparent way, in a principled way, in a consistent way. But here we were really zeroing on that question of whether they were taking censorship laws and essentially stretching them around the globe. And particularly doing that without being asked, seemingly, and applying it kind of everywhere universally, which is such a fascinating finding.

14:10And there's an interesting note in the report where it says this could go beyond just these apps because there are systems that rely on Clawed or ChatGPT through APIs and they're using those services. So you might be getting this censored output without even knowing what system you're using, without knowing that there's any possibility there, which I think poses some really serious concerns in a way that I haven't seen people talk about AI systems before. You know, you have mentioned a bunch of times this comes from your perspective of working with meta and meta platforms and meta moderation. So I have to ask, right, the interesting thing, the other interesting part of this study to me is why did you decide to investigate AI systems beyond Meta's?

14:55Look, you know, we are, as a board, we're a membership organization. We have 20 people from all over the world. We're human rights lawyers, journalists, scholars, and we have a deep interest in how human rights principles are playing out on social media. We all join the board because we care about that, and we don't just care about it at Meta. And, you know, obviously, Meta is a contender in the AI race, It's not right now at the forefront of that race in terms of LLMs. We have seen much wider usage of the other platforms. And so, you know, we have that interest in how are these issues playing out, you know, across the board.

15:33How is it affecting global discourse? And so, you know, it's not like, you know, as individuals and experts, you know, our concerns are only confined to meta. This is true, but it would appear to me from the outside that your job is confined to meta, right? So this is the first time that, as far as I'm aware, the Oversight Board has done a study or a report that goes beyond meta-platform moderation. Is that right? Yeah, look, I give Meta credit for sort of giving us the leeway to do that and not saying, you know, this is beyond your remit or, you know, not on our dime. You know, they haven't said that.

16:07And we included Lama in the study. But I think, you know, their premise in creating the board really was that these principles should prevail in the digital arena writ large. I think, you know, in an ideal world, they actually probably would want to see companies adhering to similar standards that, you know, sort of abrogating your human rights responsibilities wouldn't be a way of getting market differentiation or stealing customers. And so I think they probably have an interest in trying to see the degree to which these human rights principles are being extended across platforms. And in other areas of concern, you know, just on, for example, the spread of deep fakes and kind of the degradation of our discourse as AI-generated and manipulated content spreads more widely, we actually have begun to issue statements that talk to the field as a whole because we realize, you know, meta is part of a larger ecosystem.

17:05And if that ecosystem doesn't uphold these standards, it's less likely that any one company will.

17:35Nilay Patel: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 slash vergecast to get started today.

18:19Can Donald Trump find a way out of the war in Iran? The Iran war has been, by a long margin, Trump's biggest policy failure. Part of the reason that he has been unable to end the war is because he can't accept the loss that it would imply from his war goals. I'm Preet Bharara. And this week, Ian Bremmer, the founder and president of the leading global risk firm Eurasia Group, joins me to break down the latest in the war in Iran. The episode is out now. Search and follow Stay Tuned with Preet wherever you get your podcasts.

19:01Our good friend Casey Newton at Platformer, he reported earlier this year that Meta had reduced your budget for the next couple of years and had even floated not funding you after I think it's 2029, starting 2029. Is that still the case or are there conversations about that? You know, right now we're funded through 2029. We'll have to see what happens after that. We recognize as a board we need to evolve. That's part of the reason, you know, we chose to get into looking at LLMs because we just we see the direction that things are going. There's also so much more use of LLMs in content moderation.

19:36And so that's another there's both far greater use of AI to generate content that's shared on social media and increasing use of AI. And Meta has talked about this to moderate content instead of, you know, these farms of human moderators. They're increasingly moving away from that. And so, you know, there's a lot that's changing. And I think there's a lot of relevance and importance to the sort of work that we do. You know, we just did a study recently looking at Meta, got rid of fact checkers in the United States in favor of community notes. They asked us to look at, well, what if we do that globally?

20:09Like, what could go wrong? You know, and so we can answer a question like that kind of from an independent perspective and with global insight, people from all over the world. Yeah, so tell me about that. I'm curious. You've been on the board for, I think, five years. how how has the i guess attitude toward moderation changed in that time because to me you know zuckerberg was out a year ago talking about bringing more masculine energy to to the workplace and i i think it feels to me like there's been this sort of uh i don't know if it's necessarily a cultural shift but but perhaps some of the tech leaders want there to have been a shift where things are a little more fast and loose again and you know i think if you rewind five Five years ago, moderation felt like a crucial, essential topic, right?

20:55The Oversight Board is created as a result of this. And I'm not so sure I feel that kind of pressure anymore. Does that feel right to you? Absolutely. Look, I think, as I mentioned a few minutes ago, there were incidents that kind of created this mounting pressure. Meta adopted a very far-reaching and incredibly progressive human rights policy. They were committing themselves to upholding all of these standards. And then, you know, I think a couple of things happened. I mean, it became very politicized. And, you know, I think particularly like the Hunter Biden laptop incident, you know, was such a pivot point where the social media companies, you know, it has to have had the earmarks of a foreign information, disinformation operation.

21:35The social media companies dialed it back or took it down, you know, only to find out, you know, it actually was his laptop. And yes, the timing of the release probably was engineered to affect the election. But nonetheless, it wasn't a falsehood. And so that became then kind of the kindling for an argument that the social media companies were biased, that they were suppressing conservative speech, that moderation was a tool of the left to censor. And, you know, that was a very powerful force. And then, of course, when you had the Trump administration, you know, embodying that view, actually threatening Mark Zuckerberg, you know, before Trump took office so that, you know, he might be arrested.

22:18And so I think, you know, for the companies, you know, I think it was two things. I mean, one is that the political winds had shifted. And I also think they had come to see kind of honestly how complex and fraught a business intensive moderation is. And, you know, they made a few changes. People thought at the time that Meta had completely dropped content moderation. I would get asked, well, does the board still exist? Is Meta still moderating content? They absolutely were. I mean, the changes they made were actually quite narrow and specific about speech, about transgender issues, about immigration, that they were going to loosen the constraints somewhat.

22:55But they have reams of policies, bullying and harassment, incitement and violence, restricted goods, whether you can talk about drugs of various kinds. And so all of those policies are still in place. They're really complicated. They're honestly quite thoughtful. I mean, if you look through them, you'd be sort of impressed. They really try to look at both sides of all of these issues like it's by no means perfect. But they're also dealing with billions and billions of posts every day that, you know, a portion of which weighs novel questions. And so, you know, I think there's some sense of futility that nobody's ever going to get this exactly right.

23:33But, you know, they also one thing they've said to us as a board and, you know, we sort of see this, too, is that, you know, there are many issues at this point that we have looked at, you know, where we have made recommendations, you know, where things are, you know, working in a more considered way than they used to. And so that's positive. Yeah. So even in this world where they have backed off of fact checking in favor of community notes, even in this world where misinformation is maybe not a thing that, you know, they're so worried about, you're saying the Oversight Board still has a very real value in assisting with Meta's just sort of day to day issues with like harassment and, you know, speaking about drugs and potentially violating like First Amendment issues.

24:18Yeah, all sorts of issues. I mean, you know, with the rise of AI and AI manipulated content, you have these, say, fake videos about a politician, let's say. You know, we're getting a lot of appeals along these lines that will depict the politician saying or doing something that they never said or did. And, you know, Metta's policy right now is that if that is proximate to an election, you know, they may label it so that you can see that it's AI, but the label, you know, is only so prominent and they have a more prominent label and a less prominent label. And if the election is not coming up anytime soon, they probably won't label it at all.

24:54And so, you know, is that the right approach? I mean, as we see this content proliferating so widely, you know, should we be doing more to protect democracy or would that, you know, does that risk being repressive and suppressive and taking down, for example, satire? What's the line between a satirical AI-generated depiction and one that is more defamatory? So these are really tough calls. And we do see the content on the platform changing pretty dramatically now, largely as a result of AI and just how much easier it is to create types of content that a few years ago would have taken a lot more effort and expertise and expense.

25:34You mentioned, you know, it might actually be valuable in this world where we're dealing with AI issues. I want to quote something that you wrote. You had an op-ed piece in The Guardian a couple months back where you wrote, the title is, I'm on the meta oversight board, we need AI protections now. And there are a couple of lines that stood out to me in particular, where you write, independent oversight offers the potential to adjudicate between AI's potential and its perils. By embracing independent oversight, AI companies can demonstrate that they are serious enough about public trust to be willing to fight for it.

26:08And I thought that was very interesting because the immediate question to me is, oh, where might an AI company find an independent organization that could do this kind of work for them? I'm curious, you know, is this, you know, a job application? Is this a pitch to these companies? It's not a pitch per se, but I do think, you know, we've become convinced doing this work that taking a rigorous approach, having an international body. And, you know, that's one of the most interesting things about the board is that you're sitting down with people from Yemen and from Pakistan and Lebanon. and, you know, we have an Israeli and an Arab and we have American staunch conservatives and liberals who all come together to kind of hash these questions out.