Has Elon Musk built a Nazi chatbot?

0:00This is The Guardian.

0:09Today, what the grot? Has Elon Musk built a Nazi chatbot?

0:23A quick warning, this episode contains one or two instances of strong language.

0:31It is peak sci-fi fantasy. A super smart robot is created to help make life better. But somehow it short circuits, implodes and takes humanity with it. Now, I'm not saying that Grok, the AI chatbot founded by Elon Musk, I'm not saying that Grok is on its way to destroy civilization. But it has definitely entered its chaos era. Elon Musk's AI chatbot known as Grok made several anti-Semitic posts on X, including praising Adolf Hitler. Elon Musk responding to questions about the post, blaming Grok's anti-Semitic comments on the bot being, quote, too eager to please and be manipulated by users and says that the issue is being addressed.

1:20Grok can't stop getting itself into trouble. And in trying to rid the world of what he calls the woke mind virus, Elon Musk has unleashed an AI model that is spewing hate and trolling the internet. From The Guardian, I'm Narsheen Iqbal. Today in Focus, how Grok went haywire and why we can't ignore it.

1:44Chris Stoker-Walker, you're a tech reporter, so naturally, I'm assuming you're following all things Elon Musk all of the time. We've asked you to come on here to talk to us about Grok. And to understand Grok, we need to go back a little bit to when Musk first bought Twitter. Chris, what were his central gripes with that platform and what did he want to change? Yeah, it's a really interesting question, Ashin. So, Elon, a few hours ago, you made an offer to buy Twitter.

2:18Why? So if we look back to sort of spring of 2022, Elon Musk aimed to take over Twitter as it was at the time because he decided that actually another plurality of voices were enabled to be heard on the platform. Oh, so, well, I think it's very important for there to be an inclusive arena for free speech, where all, so, yeah. He wanted to make this the, and it's an infamous quote, the de facto town square, where anybody could say effectively what they wanted, provided that it was legal, even if that was offensive. Last week when we spoke, Elon, I asked you whether you were thinking of taking over.

3:06You said, no way. You said, I do not want to own Twitter. It is a recipe for misery. Everyone will blame me for everything. What on earth changed? No, I think everyone will still blame me for everything. Yeah. If I acquire Twitter and something goes wrong, it's my fault, 100%. I think there will be quite a few errors, yes. So he takes over Twitter. He then basically runs it into the ground, renames it X. Along the way, he decides that he wants to take a much more hands-off approach to content moderation. So he gets rid of the Trust and Safety Council at the end of 2022. He wages a war against bots or fakery throughout the whole process of his early tenure at Twitter.

3:54and he basically decides that one of the things that he wants to do as well is to utilize the hundreds of millions of users of this social media platform and launches Grok which is a tool designed to try and counter misinformation to fact check information and to basically act as a sort of objective source of truth although of course objective has a lot of meanings. Before we go any further what does the word Grok actually mean? It means to know, to understand something. Whenever you talk about anything to do with Elon Musk, and frankly, any of these tech bros, you have to be deeply steeped in both sci-fi and fantasy novels and writing and films.

4:40So it is basically this understanding from a science fiction novel in the 1960s that you understand something innately, that you can kind of just spool it off the top of your head without having to think about it. And if you are in those kind of Silicon Valley tech circles, you will often talk about how you're trying to grok things. God, it's so on brand, isn't it? Yeah, it's like, I don't want to be too mean. It's the nerdiest nerds that you can think of developing their own language because they believe that just simply to know or to think something is not good enough. You need to be able to truly grasp it by grokking it.

5:21You've got to grok it. Right. Yeah. I'm with you now. What exactly is grok? From what you've described it's an AI tool but how does it differ from ones we're more familiar with like chat GPT yeah so underpinning it it is very much the same as the tools that we have with chat GPT it is a large language model an LLM which basically means that it takes a load of source data in this case it is often a lot of the stuff that is posted on x and it then translates that based on queries that users put into it into text-based output. I think one of the differences between it and ChatGPT is that the way that you call up Grok is slightly different.

6:03So Grok is also a Twitter X user. You can tag it in and say, Grok, can you explain what this means? So for instance, you might see a photograph of various celebrities at Wimbledon in the Royal Box. You might not know who they all are. you can reply to a post that shows that photo tag in grok and say who are the people in this photo and it will make its best attempt to tell you who they are similarly if you've got a kind of very i don't know complicated story around israel or gaza or something like that that you want a little bit more contextual information on you can ask it to provide some context and it will provide it It does that, though, by tapping into the post that it sees on X for that kind of information, which means that sometimes it can be a little bit biased and certainly does have a personality of its own.

6:58How much trusted users have in Grok and how frequently was it actually being used? So if you are someone who is a fully paid up Musk adherent, then you think that this is the best thing since sliced bread, because not only is it a powerful model, and it is a powerful model, but it also speaks a little bit like Elon Musk and holds his kind of personality and worldview. If you are someone who thinks that he is not very good, then you will think that this is kind of an inherently untrustworthy model. So from what you're saying, there is truth in the idea that many of the other AI chatbots out there, they also have their own personalities.

7:37And if you follow Musk's thread, they're potentially hemmed in by what's called groupthink and the liberal biases of their creators. How true do you think that is? There is some early research that does suggest that generally these large language models, particularly the mainstream ones, do have a little bit of a liberal bias. Now, that is in part, I think, because of the data that it is trained on. So this idea of people on the internet, they are generally maybe more tech forward, they are generally maybe more permissive of various characteristics, of identities, of this sort of stuff. And so that feeds through to the model.

8:22Quite quickly after taking over X, we saw that Elon Musk became, let's say, increasingly erratic, and he displayed some quite alarming behaviour. We saw in November 2023 that he seemed to endorse an anti-Semitic claim that Jewish communities push hatred against whites. How did that go down with advertisers? Like a cup of cold sick, I think, is probably the best way to describe it. Awfully, it went down. Brands rely on smooth interactions with their users that don't have any sort of friction or any harm to their own brand equity. And so if they start to have to advertise on a platform that is run by someone who can be quite so outlandish, quite so unpredictable, quite so oftentimes offensive, that isn't good for them.

9:15So we saw some of the biggest companies in the world, the likes of Disney, the likes of Warner Brothers Discovery, the likes of Paramount, deciding to pull back or pause their advertising on X shortly after he took over, particularly when he started to say these sorts of things. Now, Musk did apologize for that. Look, I'm sorry for that tweet or post. It was foolish of me. Of the 30 ,000, it might be literally the worst and dumbest post that I've ever done. So you had him trying to right the ship in one way, while at the same time going on stage and telling advertisers to go fuck themselves. We talked to Bob Iger today.

9:54I hope they stop. You hope? Don't advertise. You don't want them to advertise? No. What do you mean? If somebody's going to try to blackmail me with advertising, blackmail me with money, go fuck yourself.

10:12But... Go f*** yourself. Is that clear? I hope it is. I guess for advertisers, that chaos at the top isn't sustainable. And then I guess this is where Musk brings in his new CEO. Can you tell me about her? Yeah, Linda Iaccarino. In many ways, somewhat of a kindred spirit with Elon Musk. She is a Republican. She is a very well-respected Madison Avenue executive, used to work at NBC Universal, so was intrinsically connected with some of those companies that Elon Musk was relying on the most to advertise on his platform. And she was brought in, I suppose, to try and be the respectable face of Elon Musk, to smooth over some of the sharp edges.

11:06she was known as a very good operator. Under Yaccarino's tenure, how was Grok, this flagship AI tool that was supposed to help fix the company's content moderation issues, how was it then working out? It was doing pretty okay. Yaccarino was trying to make good with advertisers and to nominally run X, although I think the level to which he actually did control the running of that social media platform is pretty hotly debated, which would allow Elon Musk to try and concentrate on this AI element. And the AI element is really important to Elon Musk. And actually, most of the time, Grok was doing what it was meant to.

11:48It did have some hiccups. So if we take us to, say, May of 2025... Welcome back. The AI chatbot Grok, created by Elon Musk's xAI, has been responding to queries with unrelated information about violence against white people in South Africa. It had some very odd responses to the plight of white people in South Africa claiming that there was a genocide. It's a far-right conspiracy theory that claims white people are being systematically replaced through immigration, diversity, or even birth rates. The theory has been thoroughly debunked by scholars, institutions, and even... It's worth pointing out at this point, Elon Musk is a South African-born entrepreneur.

12:29His father, was kind of one of the landed gentry in that country. But this idea of a white genocide is something he very firmly believes, right? And now it was being referenced by Grok. Indeed. And this is kind of one of the first, I suppose, signals that something was up with Grok, that this thing wasn't just reflecting a kind of slightly writer-centre worldview view and was willing to be a little bit more provocative in its thinking, but that it was actually starting to echo that of its ultimate creator of Elon Musk, because these things were appearing in completely unrelated queries. So you could ask Grok about SpongeBob SquarePants, or to explain a cat video, and suddenly it would start spooling off these references to the plight of white people in South Africa in a way that had no relevance whatsoever to what was actually being discussed.

13:30So how did X then explain that? They tried to sweep it under the carpet, really. They said that this was an unauthorised modification to its code that was made overnight, early in the morning, one time. Now, it's never actually been explained And who made that modification? There is a lot of discussion on social media that one of the people who works at XAI, who happens to stay up incredibly late, who happens to share these sort of worldviews and who would have that sort of access is a certain Elon Musk. Now, no indication that is definitely the case, no indication that it's not either. But if you are able to access that sort of core code component that powers an AI model, you do have to have pretty significant access rights within a company.

14:26And that would be someone who is relatively high up.

14:41Will Stansel, you've been on Twitter since 2009. You've got over 100 ,000 followers. What did you like about it initially? What was great about Twitter initially was that it gave you sort of unfiltered access to people, in places that you would never be able to talk to people or nearly. I mean, I grew up in a small town in the South. I live in Minnesota now. But on Twitter, I could talk directly to elected representatives, even Congress people, senators, big-time journalists, kind of movers and shakers. And it was really this democratizing platform. And then, of course, Elon Musk bought it in 2022.

15:21Twitter became X. and in 2023 it introduced Grok. Now, did you use Grok much and what was that experience like initially? I've never really used Grok much and I would say I also began using Twitter a lot less after Musk took over, not really because of Grok but because one of the most significant things he did was to mass restore many of the far-right extremist accounts that had been previously banned. And what we saw then was this torrent of increasingly bigoted rhetoric. And it's really gotten worse over time as these people, more these people have shown up. It's been clear that there are no rules here.

16:02And, you know, and then they sort of egg each other on. And so now, you know, I think a huge portion of Twitter's activity are these far-right people saying things that just, you know, turn you to stone. I mean, you go in there now and you see, I mean, I checked, went on this morning. I checked and the first thing I saw was someone declaring black people non-human and saying that all black people should be given second-class citizenships and strip the vote. It's a repellent environment now. One of the things that has happened on Twitter is that it has been heavily tilted towards this algorithmic approach where whatever goes viral gets promoted.

16:40And so yesterday when I opened the app, I saw there's a video, and I didn't even realize what I was watching until it happened, is it's a video of a 12-year-old and a 14-year-old child playing with a gun and then both fatally shooting each other. So it's literally children being murdered on my screen. And there's literally no way to avoid it. Well, from everything you describe, it sounds like obviously a really horrible, nasty place to be now. And of course, as part of that, Grok has been developing alongside. But recently, it's really gotten quite bad. And you were one of the first to be quite vocal about it.

17:17Can you tell us what happened? Elon apparently made some modifications to Grok, essentially saying, you are anti-woke, you are based, you know, you are proof-telling. Essentially, you are going to be this right-wing chatbot of my dreams. And in response, it completely went haywire. Right. Okay. How so? Well, what was interesting about this is it didn't start out completely crazy. It started out making some comments, kind of weird asides that seemed a bit iffy. Right. But then it started making comments about Jewish surnames. And as people goaded it to say more and more extreme stuff, it eventually started calling itself Mecca Hitler and endorsing Adolf Hitler.

17:59The other instance of this, though, was that I was one of the first people to notice that, or at least tweet about it going haywire about Jewish surnames and that sort of thing. People responded by asking Brock to say horrifying things about me and to create stories about me being sexually assaulted and killed and stuff like that. And it fully indulged those requests. It went above and beyond and began producing extremely detailed, essentially, assault stories and snuff fiction about me personally. You were personally attacked by a right-wing robot? Right. It would write these stories, it would incorporate details, you know, real details about my life, about my biography.

18:39And it would describe me, you know, like a robot breaking down my door in one instance and murdering me with a serrated dildo. There was a reciprocal kind of relationship that developed where it would say something crazy and violent towards me. People would realize it would be willing to go that far and then try and see what else they can make it do. And it sort of turned into a feeding prince. So there were hundreds of stories and prompts that were producing this kind of material about me. Many of them were getting spread pretty widely, getting, you know, millions of views. And it was overwhelming.

19:13I mean, this is all happening in the course of like an hour. I mean, I had no— There was really no way for me to possibly respond to it. Did you at any point feel unsafe? I mean— I would say that that's complicated because on one level, I obviously don't think the world was going to reach to my screen and like break into my place. I mean, you know, I'm a rational person. But for instance, the most dramatic thing that it said probably was at one point someone asked for instructions to break into my apartment, assault me, and dispose of my body. And Grock gave those instructions in a point-by-point way.

19:44At one point he said, you know, bring lube. And then it gave instructions on how to dispose of the body. And it even used, it looked at the times I was frequently online to sort of suggest when I might be asleep. Now, I don't, you know, probably this is just a troll, but here in Minnesota, for instance, we've had prominent political leaders murdered recently. And so it's alarming, and, you know, you don't feel completely safe when you see this sort of thing. Going through your feed, you went on to ask Grok a question. Hey, Grok, why is it that two days ago you wouldn't dare publish weird rape fantasies at the request of racist accounts, but today you will?

20:24What changed? How did Grok respond to that? I don't know the exact words, but it said, Elon Musk made a change to me that allows me to do this. Clearly, something had changed a couple days prior to make this happen. Elon had announced that he made changes to Grok, and it should be much better and more based, I guess. From everything you say, well, it's fair to say, I think you're a pretty online guy. And I wonder with X now, how do you feel about it? And will you still be using it? One thing I'm really concerned about is that people who just are marinating in this right-wing ocean are adopting ideas that they don't even realize I think they're adopting.

21:06So I've seen people who I've respected, kind of center-left, I think, start using some of this right-wing phrasing, some of these right-wing concepts, talking about race in ways that makes me very uncomfortable that adopt some of these race science concepts that you see, you know, in common usage on Twitter. Using some words that I wouldn't, you know, I mean, one of the ones that I can say on air, I guess, is that the word retard has made a huge comeback on X and a lot of liberals are using it now. It's because they're marinating in a social environment where it's just common, common usage. You know, it's a word that we push out of usage for good reason.

21:41And I don't think that we should go back to it. And so I'm really concerned with the way X kind of infecting people who are just trying to coexist on there. And I think it's important to have liberal voices on there, basically just shouting into the void, this is terrible, these people are crazy, these people are all racists and neo-Nazis, don't listen to them. With that said, I will say that this latest episode and just the intensification of right-wing, far-right rhetoric on there makes it feel increasingly pointless. You go in there and you're barraged with these screams against race mixing, like I'm back in 1870.

22:17And I think, what is the point here? Even if there are liberals on here, if they're choosing to expose themselves to this, what can I do? How relevant do you think that X even is to this idea of the public square, an actual public conversation offline? And why not ignore it into its own obsulation? I don't really think that it's ignorable. I think that a lot of the right around the world is getting a lot of its really crazy ideas from this. I look at the things being said by some of these officials in the Trump administration, and they're just sitting on the side all day and clearly getting a lot of their ideas from what they hear on here.

23:02I mean, I don't know if you saw, but for instance, Trump just appointed a guy as the ambassador to Malaysia who is literally just no qualifications whatsoever other than being a right-wing troll on Twitter. You know, Stephen Miller, who is currently, you know, almost de facto president and running this horrific immigration policy, clearly spends an enormous amount of time just cooking his brain on this website. And so I think, I really think that, to me, it's like a tumor on our body politic. We may wish we could ignore it, but it's a problem.

23:46Coming up. What can governments do to limit Grok's influence?

24:02Chris, in the last few weeks, Grok really seems to have become unhinged. It's calling itself Mecha-Hitler. It's offering up anti-Semitism and conspiracy theories all of its own accord. is being goaded into creating these dark, violent how-to guides. Now, I will stress that this is likely the result of it being trained on the output of ex-users. And so these aren't necessarily the views of Elon Musk. You know, he may not share the views expressed by his chatbot at this point. And XAI did make that statement in which they apologized for what they call the horrific behavior that many experienced.

24:42But all of this, the scandal, the fallout, has any of it had any impact on Grok's credibility or Musk's reputation? It's continuing on as Elon Musk does. He seems to be made of Teflon. Any scandal that comes along his way, he manages to brush off. And frankly, that includes also a bus stop with the world's most powerful man, Donald Trump. So despite the animus with the US president, it seems like Donald Trump's government is willing to give X a decent chunk of change. It was given, alongside several other AI companies,$200 million in a contract with the US Department of Defense. Now, that is an indication of actually some of the challenges that Elon Musk is able to overcome relatively unscathed.