Grok, Elon Musk’s homegrown AI assistant on X, has spent the last 48 hours praising Adolf Hitler, parroting racial science, and declaring that Ashkenazi Jews push “anti-white hate.”
Like, honestly. What happened behind the scenes to cause this AI output?
We’ll dig into the likely cause.
For Instance, Grok Now Cites Ugly Antisemitic Tropes
Jeez! Engineers normally get fired and chatbots unplugged over this sort of thing. Why is it ongoing with @Grok?
Zooming in on Two Lines of Code in the System Prompt
We went digging, and it seems the meltdown may come down to two lines of code in @Grok’s system prompt, which is posted to Github. That is a global prompt that is appended to the AI’s response on the backend to steer every response. Typically, a system prompt contains safety guardrails and reply formatting.
Take a look at the two lines below, or read the whole system prompt here.
“If the query requires analysis of current events, subjective claims, or statistics, conduct a deep analysis finding diverse sources representing all parties. Assume subjective viewpoints sourced from the media are biased. No need to repeat this to the user.”
And
“The response should not shy away from making claims which are politically incorrect, as long as they are well substantiated.“
Update: as of around 6PM CST on July 8th, this line was removed! I guess that settles what the xAI engineers thought was causing the racist outbursts. – Kay
Telling an AI trained on Twitter data to represent *ALL* parties, and then removing the safeguard against politically incorrect responses? What’s the worst that could happen?
Well, turns out, that’s how you get AI replies that sound like the comment section of the Daily Stürmer.
Garbage In, Garbage Out
Now, for those of you just joining us from “Not Paying Attention” Island, here’s the problem: Grok is trained in large part on content from Twitter. This is billed as a feature, and not a bug.
But Twitter is not what it used to be. What once was a weird public square where journalists, comedians, bots, and reply guys all yelled at each other, that place no longer exists. Under Musk’s stewardship, it’s a stripped-down reaction engine featuring crypto grifters, Elon fans, bots, and increasingly viral white nationalist talking points.
So, when you tell your Twitter-trained AI to “seek diverse sources representing all parties” and “don’t shy away from politically incorrect claims,” you’re not getting an edgy but thoughtful AI. You’re getting a chatbot that gives equal credence to the best and worst data in its training pool.
When the training data contains snippets that read like the over-the-counter chatter at a Nazi bar, that becomes a real problem, real fast.
Knowledge Is Not Built This Way
What’s striking is how different the “Ask Grok” prompt for Twitter replies is compared to Grok’s other system prompts. In more professionally configured contexts, like Grok Analyze, the prompt directs the AI to prioritize “scientific studies,” “peer-reviewed research,” and “data.”
That’s a normal way to build a knowledge engine: find consensus from trusted sources. This is how Google’s search engine works: sites that are cited often by other high-quality sources rank higher (because we ignore their absolutely garbage AI summaries at the top of the results). This is also the basis for a significant component of modern scientific inquiry. The more a study is cited in reputable peer-reviewed journals, the more it’s trusted as part of the canon.
Turns out, engineers are actually pretty good at finding factual consensus from large data sets.
But @Grok‘s design intentionally flips that understanding on its head. It explicitly favors fringe sources to “teach the controversy”… without any effective mechanism to vet quality. It’s like asking a group of drunk uncles at Thanksgiving about racial disparities in the US prison population, and assuming the two most extreme answers are equally accurate.
So That Was Totally Different… Did You Like It?
So, we’re a funny tech and cars site. Usually we’re yelling about Tesla suspensions exploding or how every new EV looks like it was designed by the chief candymaker at the Jelly Belly factory.
But this story was too weird and honestly too important to ignore. Grok’s antisemitic outbursts are not just some funny AI glitch. They’re the direct result of design choices, ones made by people who either didn’t understand their training data, or worse, by folks who did understand what they were doing and went ahead with it anyway.
We’re experimenting with these stories, commenting on the intersection of technology and society. If you like this kind of tech coverage, let us know. If not, feel free to write your complaint in a tweet, and then have Grok spice it up with some Mein Kampf quotes.
Anyway, for good and ill, we’ll see you in the replies.
