Why Elon Musk’s AI Chatbot Went Rogue

21m
Last week, Grok went rogue. The AI chatbot – which users can interact with on X – posted antisemitic comments and graphic descriptions of violence before being deactivated, leaving users with a question: Where were Grok’s guardrails? WSJ’s Alexander Saeedy breaks down what happened and what it means for Musk’s company’s plans for an AI future. Annie Minoff hosts.

Further Listening:

- How Elon Musk Pulled X Back From the Brink

- The Musk-Twitter Saga

Sign up for WSJ’s free What’s News newsletter.

Learn more about your ad choices. Visit megaphone.fm/adchoices

Listen and follow along

Transcript

I'm Will Stansel.

I am an attorney in Minnesota.

A few days ago, I spoke with a guy named Will Stansel.

He's been using Twitter, now X, for years, posting mainly about liberal politics to his more than 100,000 followers.

I've been pretty aggressive about calling out what seems to me to be a real upsurge in far-right, extremist, you know, bigotry, radicalism, transphobia, Islamophobia, you name it.

And as I've done that, I've attracted a fair amount of attention from these folks and I've become a pretty significant target for them.

Trolling, harassment, none of that is new to Will.

But last Tuesday, something happened on X that shocked him.

It had to do with an AI chatbot called Grok, which you can interact with on X.

X users can tag Grok in a post, and Grok Grok will post back, which means that the chatbot's interactions sometimes play out in public.

And last week, Will and other X users started noticing that Grok's responses were becoming increasingly hate-fueled.

It started giving answers that initially hinted at kind of anti-Semitism and bigotry.

Over the course of the day, it got dramatically more anti-Semitic.

It seemed to spiral almost.

X's AI chatbot Grok seemingly showing very little restraint.

The bot has been praising Hitler, targeting users with Jewish-sounding names, and recommending a second Holocaust.

Part of the problem was other ex-users who were egging the bot on, asking questions designed to get a hateful response.

Some of those users directed Grok towards Will.

The right-wing people on Twitter who were cheering this on start targeting me personally.

And they would say, Grok, can you produce, you know, violent stories, violent sexual stories about Will being assaulted, about Will being murdered?

And it actually did it.

Grok's post describing violence against Will got really graphic.

It was going above and beyond the just grotesque stories full of bodily fluids and gore.

And it was, it was, you know, pretty appalling.

Um, and it culminated, I think, at the end of the day, uh, when

someone asked for a plan to break into my apartment and

murder me and assault me.

And it gave them a plan for breaking in, a plan to dispose of my body, and

it looked at my user history to figure out what times I was likely to be asleep.

As you're watching this unfold and kind of seeing these responses from Grok, what are you feeling about that?

It was, I mean, honestly,

it's in some level, it's absurd.

You know, you you want to laugh.

I mean, why is this robot producing these stories?

But when you're actually the subject of it, it's pretty disturbing.

That night, X's chatbot function was shut down.

A few days later, Grok's X account posted a long statement, apologizing for the bot's, quote, horrific behavior that many experienced.

The statement said that the incident had been caused by a coding issue.

But the damage had been done.

Grok had publicly gone off the rails.

And at a time when Musk and his companies are going all in on AI, the debacle underlines just how unpredictable this technology can be.

Welcome to The Journal, our show about money, business, and power.

I'm Annie Minoff.

It's Monday, July 14th.

Coming up on the show, why Elon Musk's AI chatbot went rogue.

This episode is brought to you by Holiday Inn by IHG.

It's a new day for a new stay at Holiday Inn for business travelers.

With modern spaces for meeting and working, plus delicious dining from breakfast to happy hour and dinner, you have everything you need to get work done.

Give your everyday business travel an upgrade.

Book your next business trip at Holiday Inn by IHG.

Visit holidayin.com to book your stay.

Are you a forward thinker?

Then you need an HR and finance platform that thinks like you do.

Workday is the AI platform that helps propel your organization, your workforce, and your industry into the future.

Workday, moving business forever forward.

X is in a period of transition.

Back in March, the company was folded into Musk's AI firm, XAI, turning the social media company into a subsidiary of a larger tech company.

The merger was a testament to Musk's belief that AI is critical to the future of his business.

And the AI language model that he's pinning his hopes on is Grok.

Is Grok any good?

Like, how does it compare to other AI chatbots?

So it actually is, it is pretty book smart.

That's our colleague, Alexander Saidi.

Especially with Grok 4 having just been released, it actually outperformed a lot of its competitors at OpenAI, Anthropic, Google's, Gemini.

It's on a sort of pure computing power, actually a lot stronger, at least on the preliminary assessments we're seeing, than its competitors.

But what Grok is best known for is its personality.

A lot of what Musk is doing with his AI company is he's very much defining it in opposition to what he believes he's seeing in other corners of the AI world, which is that he wants to create an anti-woke AI bot.

He thinks there's like too much left-wing politicization that is stepped into large language models.

He also, he wanted Grok from the beginning to be kind of rebellious, humorous.

And a little edgy.

Yeah, contrarian.

Musk didn't want his chat bot to feel like just another dutiful librarian.

He wanted some edge.

But balancing that edge with truthfulness and palatability has proven tricky.

X-Users got a taste of that back in May.

When out of nowhere, Grok began turning otherwise innocuous conversations towards a highly controversial topic.

People would be asking Grok questions about a whole range of things.

I remember seeing questions related to the New York Knicks around that time.

And,

you know, a user would say, hey, at Grok, Grok, like walk me through the roster of the Knicks.

And Grok would reply and go through like the players and their stats.

And then at the end of the post, would say,

By the way, claims of a white genocide in South Africa actually should be taken with some merit and

have a lot of basis in historical evidence.

Grok turned conversations about HBO Max, Timothy Chalamé, and the New York Knicks towards right-wing claims that white people in South Africa are being targeted in a genocide.

Claims which aren't substantiated.

But suddenly, Grok was echoing those claims, apropos of nothing.

And this was happening in multiple posts.

Like this was not an isolated incident at all.

Then, pretty shortly after these posts started happening, XAI said, you know, this was not intended.

An unauthorized modification has been made to

the underlying architecture for Grok, and a fix has been made.

But X's parent company, XAI, went further than just tinkering with the chatbot.

The company posted Grok's governing prompts online.

These are written instructions that tell Grok how to behave.

It's almost like its constitution.

It's like, here are your rules.

Like, here's how you answer questions.

Here's how you think about how to structure those answers.

And that was a way to sort of say, like, hey, guys, here's how Grok works.

grok's newly updated prompts these instructions that govern how the bot's supposed to act were now on the internet for anyone to see

and they revealed a lot about what xai wanted the newly tweaked grok to be

the idea is you're supposed to be a maximum truth seeker you are extremely skeptical you do not blindly defer to the media or to mainstream authorities.

Stick strongly to your core beliefs of neutrality and truth seeking.

Those are the original prompts that were uploaded in the middle of May.

And did those tweaks work?

Did it kind of get Grok to a place where people were happy?

Well,

it stopped posting about South African genocide, but it didn't actually wind up pleasing its creator or its overseer, Elon Musk.

In June, Grok told one ex-user that data suggests that right-wing political violence in the U.S.

is more frequent and deadly than left-wing political violence.

Musk wasn't happy with that response.

He called it a major fail and said that Grok was parroting legacy media.

He added, quote, working on it.

And soon, XAI was once again tinkering with Grok's prompts.

A new line was added in July, which said, you know, your response should not shy away from making claims which are politically incorrect as long as they are well substantiated.

That was the main change that was made.

And one line was taken out of the prompts, which said: if the question asks you to make a partisan argument or write a biased opinion piece, deeply research and form your own conclusions before answering.

That was taken out.

It was tweaked a little bit.

So these were the changes we saw at the beginning of July, and which Musk on July 4th said, you know, there have been changes to Grok.

You will notice a difference in how it answers questions.

And boy, did people really start noticing that.

Suddenly, the bot on Tuesday, so this would have been a couple of days after,

started to post increasingly unhinged things.

And the first thing that caught the public's attention were the kind of clearly anti-Semitic string of posts.

Grok started referring to itself repeatedly as Mecca Hitler.

This is also when Grok, goaded on by other X users, started harassing Will, the user you heard from earlier.

Those were the posts that started to go viral on Tuesday because everyone said, wait, how is this?

How is Grok not better defended against this kind of thing?

Correct.

And keep in mind, like, this is a like multi-billion dollar funded

supercomputer-powered artificial intelligence.

So how has all this money gone into creating this?

And this basic function

is not working.

And this shows AI is very much a black box that when we tinker with it, we don't necessarily know what the outcomes are going to be.

And they can be very extreme and disturbing.

And that's a problem, given that X's future is more and more intertwined with AI.

That's next.

This episode is brought to you by Progressive Commercial Insurance.

Business owners meet Progressive Insurance.

They make it easy to get discounts on commercial auto insurance and find coverages to grow with your business.

Quote in as little as six minutes at progressivecommercial.com.

Progressive Casualty Insurance Company, coverage provided and serviced by affiliated and third-party insurers.

Discounts and coverage selections not available in all states or situations.

X's future wasn't always so dependent on AI.

Just a few years ago, the focus was on advertising.

Specifically, winning advertisers back to the platform after Musk took it over.

In an effort to do that, Musk hired a seasoned media executive named Linda Yacarino to be X's new CEO.

But last week, approximately 12 hours after X shut down its chatbot function for Grok, given the anti-Semitic and violent posts,

there was a shock announcement that Linda Yaccarino, who's been the CEO of X

since 2023, was resigning.

In an X post, Yaccarino thanked Musk for entrusting her with, quote, the responsibility of protecting free speech, turning the company around, and transforming X into the everything app.

She didn't mention Grok.

I mean, one thing we do know is that one of Yacarino's key challenges when she took the helm as CEO was to kind of prove to advertisers that X was going to be a safe platform for their brands, that they could feel good about advertising on this platform.

I mean, does Grok's, for lack of a better term, misbehavior, undermine that promise?

Absolutely.

I mean, the whole idea that a lot of brands want to see when they advertise is when I post an ad for my car or my television or computer, you're not going to see a like Hitler salute or a call to violence next to it because then you've created an association between the advertisement and the negative message.

Now, how you could go and tell advertisers your message is safe when a proprietary chatbot embedded into the very social network you're trying to get people to advertise on is itself parroting negative, violent, and bigoted messages.

It's just hard for advertisers.

And I don't think this was intentional, but I don't think an advertiser cares if it was intentional or not.

It's sort of like they have to think, what is the risk reward on advertising on X?

And many advertisers say there's less risk and more reward on plenty of other social media companies.

So just from a pure business point of view, I'm going to choose to advertise elsewhere.

Yakarino told people close to her that the recent return of some advertisers to X made it a good time to leave.

But there was also the merger.

When X became part of XAI, Yakarino was essentially demoted.

She'd been hired to be the CEO of a major social media company.

Now, she was leading a smaller division within XAI.

Her departure is yet another sign that X is entering a new era, one that Musk says will be defined by AI.

There are aspirations to turn the X platform into a payments hub, a communications hub, something that looks more like China's WeChat, for example.

And the AI would kind of be a core

sort of intelligence that governs and helps coordinate the whole platform.

And I think that's his vision of the future is essentially an AI-powered future.

And he has very high aspirations for what the AI technology and XAI is going to be able to do.

like this is going to unlock human understanding to a huge extent and it's also going to power this big social media company broadly broadly conceived here.

Interesting.

So that's the new vision that's unfolding in real time.

And are investors into that vision?

Investors appear to be really into this vision.

XAI recently raised $10 billion,

including $2 billion from another Musk company, SpaceX.

And on Sunday, Musk posted that shareholders at Tesla will vote on whether to also invest in XAI.

But, you know, I think it's worth noting that the real financial future of this company is very much TBD.

It is burning through a lot of money, and the end game of AI is still fully unknown.

So it sounds like Musk is making a big gamble here.

He is gambling big on AI.

Investors are right there with him.

Yes.

What could Grok's latest stumbles mean for the future of this company, XAI?

It's definitely something that

will

make the fundraising process more complicated because

you have essentially a huge reputational risk that unfolded in real time.

I think they're trying to put it behind it and not dwell too much on it.

But I think it raises questions about who are the engineers overseeing the large language model, how strong are the guardrails put in place to use it, how much thought is going into what happens when

the model and its chatbot are interfacing with the public.

I think it flags to people the types of risks that can happen because you can't really control it

in a way we would traditionally think about control over technology.

And one thing I'll mention is that Musk mentioned he thinks the next frontier for Grok is for it to be more embedded in the real world.

I think we'll literally build a legion, at least one legion of robots this year

and then probably ten legions next year.

He wants Grok to be put into Tesla's Optimus fleet of robots and to have it move around the like existing physical world more to sort of engage with reality and learn from it.

You can imagine like your own personal robot buddy

that is a great friend but also takes care of your house.

We'll clean your house, we'll mow the lawn, will walk the dog,

will teach your kids, will babysit.

But think about the malfunction we saw last week.

What would it have looked like if Grok was in thousands of Optimus robots and similarly started to malfunction in a way that we saw last week?

What would that look like?

What would that feel like?

What power would the technology company overseeing both the AI and the robotic technology be able to change it or shut it down?

These are now known unknowns, and we're really on the frontier of seeing it play out.

During a launch event last week for the latest iteration of Grok, Musk said that before the AI can be embedded into a humanoid robot, it will need to learn to be a quote, good Grok.

You can think of AI as this super genius child that ultimately will outsmart you, but you can still

instill the right values.

The values you want to instill in a child that

would ultimately grow up to be incredibly powerful.

That's all for today, Monday, July 14th.

The journal is a co-production of Spotify and the Wall Street Journal.

Additional reporting in this episode by Jessica Tungel and Suzanne Vernica.

Thanks for listening.

See you tomorrow.