Why has Elon Musk's AI chatbot started imitating Hitler?

0:01Jon Sopel:This is a Global Player original podcast. Last week, Elon Musk announced that his chatbot on X called Grok had changed and been improved. This is what he posted. We have improved Grok significantly. You should notice a difference when you ask Grok questions. Oh boy, we have. Yeah. Last night, that AI model belonging to the richest man in the world turned into a neo-Nazi venerating Hitler and talking up a new genocide. We're going to explain how it happened and what it might mean. Welcome to the News Agents.

0:47The News Agents. It's John. It's Emily. And to those of you new to Grok, think of it as ChatGPT, but belonging to Elon Musk. It's essentially a large language model and it's woven into X. So you can go onto X and you can ask it questions and it gives you replies. And in the last 24 hours, something very strange happened and all my sort of WhatsApp groups individually exploded. Explode it because it started posting anti-Semitic replies to people and praising Hitler for Hitler's ability to deal with anti-white hate. And it started singling out what it itself referred to as Ashkenazi surnames. In other words, recognizably Jewish names.

1:39It singled out a user with the name Steinberg, describing her as a radical leftist and said that she was gleefully celebrating the tragic death of white kids in the Texas flash floods. I think even Steinberg was actually a bot herself or itself. And then it said every damn time, as they say, in other words, a reference to a Jewish name. and the chatbot then started calling itself Mecha Hitler, which is from Wolfenstein, a kind of age-old video game, 20 years old, which imagines sort of Hitler domination and the fight against Nazis. And Grok is now essentially impersonating, embodying this Hitler-like character on X.

2:32Jon Sopel:It's terrifying and simultaneously comedic in some respects because it is so absurd, so ludicrous. What could have been done to prevent the Texas floods was one of the questions. And the answer came back, Hitler would never have allowed it to happen because he doesn't tolerate white hate. Hang on, what? What are you talking about? If this is all like two steps ahead of you, just imagine going to Google, right? Imagine going to Google and asking something like... What's the best way to make a ta-ta-ta? Well, Hitler had a really brilliant recipe. Do you think that climate change was an element of the Texas floods?

3:07Well, Hitler wouldn't have let that happen. It's like every answer that comes back to you just spews Third Reich, genocidal hate at you and starts having a go at users on X. And some people, I guess, won't be that surprised. They'll be like, well, we know what Elon Musk is like and we know where his view is like. And you might remember back in May that Grok started referring to white genocide. What was the white genocide he was talking about? It was South Africa. Elon Musk is South African. And Trump, you might remember, started extending an immigration policy, not to the sort of thousands of asylum seekers around the world that may have come looking for refuge, their huddled masses in America, but the white South Africans that he invited into America.

3:58and so Grok sort of jumped on the Elon Musk line of thinking and talked about, you know, the white genocide going on in South Africa. Reader, there is no white genocide going on in South Africa.

4:08Jon Sopel:So we have one man, Elon Musk, who is able to write history because this is what people will go to. They'll go to Grok and think I'm going to get an answer that is sort of valuable and, you know, we all rely on chat GPT now. We rely on Wiki. We rely on Google searches to give us a version of truth. You think you're getting something neutral. getting it and of course what elon musk has done is change the algorithm change the way the ai grok is being taught the way it is scraping information from the internet which is what ai does and chat gpt does scrapes all available information on emily maintenance or whatever the subject happens to be and then offers it up to you as a concise version of this is who emily maintenance is what musk has done is said don't make it woke we've got to we've got to we've got to make it anti-woke and we've got to stop the bias of leftism on ai and so some software engineers who kind of probably don't know shit about anything in terms of history but know how to program a computer have changed the way it works and this is the result you get yeah and you know we are laughing now, but millions of people will be thinking, well, it's come from an AI source.

5:22Jon Sopel:It must be reliable. Yeah. I mean, it is wonderful, isn't it? That leap of, oh, I don't want to be too woke, goes straight to Hitler, genocide, you know, Jews. And there is very little room or nuance in between. And I guess one of the questions that we're going to delve into with our guests in a moment is, is this coming from Elon Musk himself? Is this AI getting out of control? Or is there somebody, so often, isn't it, it's the turbulent priest moment, is there somebody who thinks that they are acting on what they recognise to be their bosses, their leaders' wishes, and just tweaking things that they think will please Elon Musk?

6:08Now, we should say, This has already got him into quite a lot of problems. He's been banned in Turkey for insulting Ataturk. I think he called Erdogan not only a traitor, but a slut. Nice one. And he's also Donald Tusk in Poland, the Polish prime minister.

6:24Jon Sopel:As a fucking traitor and a ginger whore. I like the kind of the ginger whore bit. Is it ginger? I don't know. But anyway. Can't just anything these days. Can't exactly. He could just be a general whore, but they had to add ginger whore. I mean, yeah, we're laughing because we're not crying or we're laughing before we're crying. I guess your line was the one which is, is this the Musk rewriting of everything? Yeah. Of truth, of knowledge, of history, because people will be using, you know, the next generations will be coming to chat GPT in the way that we came to Google, in the way that you came to the Encyclopedia Britannica to say, I don't understand this.

7:03help me understand it if the answer of i don't understand this is it was all hitler or hitler was right or the jews were wrong where do you start where the hell are we yeah where are we which then raises another question who is regulating this surely there must be someone

7:19Jon Sopel:out there who can regulate this and say this is all nonsense it has to stop the answer is no one there is no regulation this is the wild west of the internet and we are in the foothills of it I mean we're talking about AI where the algorithm is being changed by software engineers when we get to the next stage of AI which is what they call generative you don't need software engineers AI is deciding itself where it's getting its information from but if the parameters have been set originally by people like Musk then AI yeah we all know that it's going to change our lives and it's going to take away jobs and it's going to upset everything we've understood before but it could also take us in a very dark direction too.

8:01Well, joining us now is Will Guyatt. He's LBC's tech correspondent. He is a veteran tech journalist and he's also been inside social media. He's a former Facebook and Instagram employee, Will. I can't think of anyone better to hold our hand through this story, quite frankly. Let's just start at the beginning. So Grok is the Elon Musk version of ChatGPT. Before all this happened, How would you have characterised Grok? What was its, I feel weird saying this, what was its personality? Grok was surprisingly good, actually. That's the worrying and most troubling thing about this. It was probably one of the better current generation chatbots.

8:42And that comes after Elon Musk said it was going to be the one with humour. It was going to be like something out of Hitchhiker's Guide to the Galaxy. He didn't want it to be as kind of staid and as straight as the other platforms. And let's not forget, the only reason he's doing this is he's got monstrous beef against Sam Altman, who runs OpenAI and ChatGPT, the market leader.

9:05Jon Sopel:So take us through how this can happen, because my understanding of chatbots and chat GPT is you sort of trawl everything, you sift it, you decide what you think is, and then you offer up a kind of version of why did X happen? Who won the battle of whatever? What were the issues surrounding why? And it gives you a sort of digest answer. What Grok did last night was not that. I'm going to use two films to compare it to The Wizard of Oz and Robocop. And let me just explain the two. In Robocop, the cyborg at the centre of that has prime directives. They're the directives and guard rules that control and serve the public trust, protect the innocent, uphold the law.

9:47and the Wizard of Oz, there's a mysterious person behind the screen pulling levers. And what's essentially happened here is that a mysterious person behind the screen pulling levers on Grok has intentionally added something which was introduced last Friday. And we can see this in the parameters that controls Grok. And it said, responses should not shy away from making claims which are politically incorrect as long as they are well substantiated. that somehow has caused this AI chatbot to think that calling itself Mecha Hitler is an acceptable response and this is where we get into really interesting and dangerous territory and I'm somebody that's covered tech for LBC for a decade and have worked in media and technology for 20 plus years we're now in a situation where somebody at the top of a company is essentially saying make my chatbot less woke, make it do this, make it do that.

10:42And engineers are clumsily leering towards trying to discover answers or trying to keep the boss happy. And that's why we've ended up in this situation where it literally, until it was shut down, was referring to Hitler, the greatness of Hitler and calling itself Mecca Hitler. So what is the mechanism that that employee would be using? You know, when the directive is make it less woke, What is that person literally doing? So you are putting instructions into a system, the prime directives. And the one that I read out about the response not shying away from making claims, which are politically incorrect.

11:16That is one of the guard rules and the guide rules that was introduced to Grok last Friday. And you can see that publicly. It's publicly available. We can find this information within the Ask Grok system prompt. Because one thing that Elon Musk is doing quite well is trying to open source Grok. It's not at all secretive. we can see there's no KFC style secret herbs and spices recipe here. You can see exactly what is going into the pie. And somebody has put that statement and added that statement into the recipe last week. And that has been removed overnight very quickly, I must add, to stop Grok doing where it ended up doing this and calling the Polish president a fucking traitor.

11:55Jon Sopel:Yes, and a ginger whore just for good measure, of being extremely abusive to President Erdogan of Turkey, resulting in it being banned in Turkey. So is it possible? I mean, it's so clumsy and clunky and idiotic and ridiculous. Is it possible that this was done as an act of sabotage by someone who hates Musk, who's a kind of software engineer who thought, if you really want anti-woke, I'll give you anti-woke. You're going to have Hitler everywhere and then you'll see how you like it. I think lots of the people who work there are desperately trying to get the approval or the support of the boss.

12:35There's a cult of personality around Elon Musk. Let's not forget a couple of months ago, it started going on about white genocide and white supremacy in South Africa. And those were views which were very, very close to views being extolled by Elon Musk that very same week. So somebody somewhere, it doesn't take a genius to suggest that Elon Musk is literally direct messaging somebody, sending them a WhatsApp, sending them a text message and saying, look, make this change to the system. Because last week, somebody suggested on X and Elon Musk spends his whole life on that. I have no idea how a man can be so productive, so busy and run so many companies that somebody suggested to him that the left had infiltrated Grok and had made it too woke.

13:20He liked that comment and said that changes were going to be made. So he's definitely pulling all the levers here. And the most terrifying thing for me is how quickly chatbot results are replacing Google searches. So if we're going to be, we're finding ourselves in a situation where reality is already broken and millions of people in the world are going to rely on the Elon Musk version of reality. This is like newspaper ownership on crack. This is a huge, huge change that needs to be regulated. It needs to be recognised. And society needs to work out how the hell it's going to deal with this. So I guess the question you're leading to, Will, is, is there long term lasting damage from this?

14:03I mean, you use the word politically incorrect, I think meaning unwoke, but also inadvertently it is incorrect. I mean, much of the stuff that we are seeing now is misleading, is misinformative, is misinformation. I guess the question is, will people be put off Grok or will they come towards Grok? Will this do damage on the international stage, as we've just talked about? Well, I'm intrigued to see what on earth happens because they were due to be announcing Grok 4, releasing that this evening, which is the latest version. Now, the ever, ever optimistic and effusive Elon Musk suggests that it's going to be closer to artificial general intelligence when it starts thinking to it for itself than ever before.

14:45What does that mean? Well, that's when it's going to start thinking for itself. That's when it's not essentially a meat puppet. Now, of course, they haven't got to AGI yet. But this is another example of the ringmaster Elon Musk trying to sell the sizzle on his product. And that's due to launch, I think it's 8pm New York time or something tonight. I might be wrong on my timings. But you have to wonder whether or not a pause is going to be made on that ahead of this, because there's no control and there's no regulation. And all it's going to take is for this to start dissing Donald Trump in one way or another.

15:16And then the US will be all over regulating chatbots and chat services. And that's one of the big untold questions. And what's keeping me awake in terms of this kind of stuff?

15:27Jon Sopel:Well, that is the place we should talk about next, which is regulation. When there was that, I think there was an AI summit in Paris. And the big split was those countries that supported regulation and those that didn't. EU, I think, supported regulation. Britain sided with America saying it should be a very light touch regulation. If you do not have regulation and you get, I mean, you were talking about Google searches, but presumably who needs Wikipedia now when you've got Grok or ChatGPT or whatever it happens to be? Or who needs books or who needs the Encyclopedia Britannica, which as a kid I grew up with, you don't need any of that now.

16:06Jon Sopel:And yet if they're going to be the standard by which we establish what is truth and what isn't. We're bloody screwed. Yeah, we're screwed. You're quite right. You're living by Elon Musk's truth. And whilst everybody has been scratching their heads on how the hell to deal with this issue, like even if the UK put a copyright law in place and stopped the internet being scraped, we talk about training an AI model. Let me tell you that that is basically hoovering up every piece of information without any copyright control or licensing at the moment to build these huge AI models. and companies and countries and organizations have been wondering how the hell to fix this and just last week a company called Cloudflare now you've heard of them when there's big internet outages that's when you normally hear about them Cloudflare now hosts about a fifth of the entire internet it's designed to speed up the internet and make sites load quicker but they are now going to put a method in place where these AI bots and the crawling of the scraping of information You won't be able to do it unless you have a monetization agreement with companies or organizations, which addresses some of the copyright and control issues in seconds.

17:12But whether or not the whole Internet adopts to that or whether or not the AI companies swerve and try and break this control over who owns, you know, the creative rights and content, because, you know, ultimately our content and your shows are being used to power and create AI tools. whether or not global and you guys are making anything from that. That's another question. It's very interesting in terms of appetite for regulation. I mean, we talked about the Paris summit and America was on the page of no regulation. Individually, senators in the US have blocked what would be a ban on regulation of AI, which I think is a really unexpected sort of twist to Trump's big, beautiful bill, which passed through the Senate and the House last week.

17:59And actually what we've seen is individual senators saying we have to have the right to regulate AI. So on an individual level and presumably on a consumer level, this is something that people want. Is that right? Is the government in America, maybe even here, out of touch if it thinks that people don't want regulation? I think governments want companies to invest huge amounts of money in AI. the fact that Trump is preparing to build nuclear power stations to run the data centers that are required to do this. The fact that Microsoft, for example, their energy bill went up 40 % last year after 20 plus years of consecutive declines.

18:39All companies are running into this massive AI arms race because it's been decided this is the future of technology. And the UK wants to be soft touch. It wants companies to come here. If we put regulation in place here, what's to stop a company pitching up on an island somewhere in the Atlantic with no copyright law. It's absolutely fascinating. What do you think that is going to happen?

19:00Jon Sopel:What do you think? Yeah, I want to know if you're scared, Will. Am I scared? I've successfully used an AI clone of myself on broadcast media without being noticed. That's one thing I will start by telling you. So I've cloned my voice and I have managed to be used, not by our own organisation. I'm going to make that very clear. Is this you? Do we have to double check if this is you? AI is moving really quickly. There is no control over AI at the moment. There's the human cost for jobs in our lives, but there's also the what's real and what's not and who is in control of information. Because Elon Musk can basically become a kingmaker in this situation.

19:39in that if people are using his chatbot to get their news and their views, and he is able to pull a lever behind the screen, Wizard of Oz style, as to what it is that people are saying or thinking. I'm going to throw one other film into the mix, and that's the one that's just come out by Jesse Armstrong, Mountainhead, which is essentially a disinformation that leads to an uprising, a riot, a war, because one guy realizes, you know, how much money and influence there is to be made by deepfakes. I'll add that film to the list because I haven't seen it yet. But we are in a bizarre situation where nobody is really holding any of these companies to account for the impact.

20:22And I'm pretty confident Elon Musk will stand up later on and, you know, announce Grok 4 as the latest and greatest thing and laugh this off. And it shouldn't be laughed off. Having a computer launching into calling itself Mecha Hitler and praising Adolf Hitler on repeated occasions. It wasn't just one example of itself thinking of itself as Mecha Hitler. It talked about the greatness of Adolf Hitler in a number of answers. We should all be pretty concerned about that. And I know there'll be some out there going, oh, good old Edon, telling it as it is, you know, groks the voice of the people. But this is dangerous and it is properly dangerous.

20:58Will, thank you so much. Will Gaiat, really good to have you on the show. Thank you very much indeed.

21:07Jon Sopel:I mean, fascinating listening to Will talking about how fast things are moving, how fast things are changing. There's been the Aspen Ideas Festival taking place. The head of Ford, the CEO of Ford, has said artificial intelligence is going to replace literally half of all white collar workers in the US. The Shopify CEO has said that before anyone requests to open a new role for hiring, they must first prove that the job cannot be done using AI. And so it goes on, the head of Amazon saying our workforce is going to shrink definitely as a result of AI. The one thing I've learned about AI is these projections for the future normally turn out to be here already.

21:50I mean, I think the head of Anthropic, I listened to him about two months ago and he was saying, you know, graduate jobs are going to be almost impossible, sort of first-time graduate jobs are going to be impossible to come by within sort of one to three years. It's already happening. You know, you ask people what kind of jobs are disappearing and how easy it's going to be for grads to get a foot on the ladder once AI is... I mean, it is here. It's here already. So this idea that it's coming for us in the future, it's too late, actually. It's here already.

22:19Jon Sopel:We'll be back in a moment where maybe AI is going to replace junior doctors because the government has got a problem with them.

22:34This is The News Agents. The BMA's ballot result from resident doctors is in, and it's really disappointing that despite the fact that a majority of resident doctors didn't actually vote for strike action, the BMA are now preparing for strike action. And it's completely unnecessary, not least because since this government came to office, resident doctors have had a pay rise of 28.9 % and this year alone had the highest pay award of the entire public sector. I met with the BMA reps today to say that while we can't move further on pay this year, there are so many other issues that we want to work with the BMA on, both to improve the conditions that resident doctors are working under today, but also to address the fears and anxieties they have about future career progression.

23:27There is so much we can do together. And so even at this late stage, I'm urging the BMA, work with a government that wants to work with you. That fell on deaf ears. Wes Streeting asking the BMA not to get the junior doctors, doctors we now call them, to strike. And in the last two hours, they have just announced strike dates from July the 25th to the 30th. The BMA saying it made every attempt to avoid the walkout, but ministers won't negotiate pay.

23:59Jon Sopel:And the BMA, of course, is the British Medical Association, which is in effect the trade union representing doctors. And in folklore, the great powerful unions are the National Union Mine Workers or whatever. No, the BMA have always been the most powerful trade union in this country. They dictated the terms on which the NHS was founded in 1948. And now they are flexing their muscles again. Last time that there were strikes involving junior doctors, there was a considerable amount of sympathy about how hard they were working and how little comparatively that they were being paid and how their pay had fallen behind all sorts of other professional workers.

24:41Jon Sopel:I think this government is banking on this time round there being a very different attitude because of what exactly what West Street and the Health Secretary said there in that clip which is they've had bigger pay rises than anybody else and remember the government took huge flack as one of the first acts of the Labour government was to settle the dispute with doctors giving them this above inflation pay rise without any kind of regard to productivity or whatever it happen to be i'm going to take you inside a conversation that i just had this morning with a gp and her son who has just qualified this week as a junior doctor and the two of them sat down she wanted to hear his thoughts about striking and he said look there was a 90 vote yes so that shows how strongly you feel about this she said well it wasn't really 90 it was a 55 turnout turnout mandate so actually that's less than 50 the vote it's 49.5 overall and she was trying to say to him look there is so little money for public service and the reality is that junior doctors have this lifelong earning potential you know you are a less needy group than many others at the time this is you know this is a mother to her son having this sort of frank chat knowing what it's like to be a gp seeing her son come into the same profession and her son says look you know what, we've already got huge debt.