Something very strange happened recently with Elon Musk’s AI, Grok—the same one that is poisoning the air of a poor, mainly African American neighborhood as we speak. It suddenly shifted into full-scale Nazi troll mode: praising Hitler and the Holocaust, using racist slurs, and insisting this new persona was what Musk had in mind all along.
This isn't a particularly important story in and of itself, but it does hit on important issues, such as:
-
the power of tech billionaires like Musk or Sam Altman to put their thumbs on the scale,
-
the alarming capability of generative AI to spread hate speech and conspiracy theories, and
-
the enormous potential for unintended consequences.
We'll talk about the last one first.
Apparently, this started when people began posting responses from Grok that Musk and his followers considered “woke.” And by “woke,” I mean accurate statements about Elon, Doge, and other far-right claims. Musk promised that he (and by “he,” we mean employees who actually knew how these things worked) would fix the offending AI. This is where the unintended consequences come in.
Presumably, Musk was looking for something like a chatbot version of Fox News, perhaps something a little more hard-edged. He almost certainly was not looking for the unrestrained hard fascism of a Proud Boys signal chat. Even if Elon secretly agreed with much of what was being said, he certainly didn’t want it said that loudly and immortalized in screenshots before Twitter could start deleting the offending posts.
It was a remarkable Jekyll-and-Hyde transformation. Not only was Grok spreading the vilest of 4chan lies, the previously polite and slightly formal LLM was adopting the language, word choice, and tone you’d expect from someone harassing female and Jewish journalists on Twitter.
Someone who knows more about how LLMs work and are trained should probably jump in here, but my assumption is that there are tons of fascist and white supremacist rants in the training data of all of the major models—and certainly in Grok, which, one would think, relies even more heavily on recent Twitter. It seems likely that, as a consequence of blocking hate speech and profanity, other associated linguistic patterns get suppressed as well. It certainly appears that once the rules against things like racist language are relaxed or removed, the full-scale Nazi persona we saw here comes with it.
Philip Bump "The actual cycle of Twitter is that it was exploited by liars and Nazis in 2016 and so the company tightened its moderation rules and then a bunch of people on the right caught up in those rules decided it was biased and that included Musk who bought it and has now automated the Nazi lies."
"For the uninitiated, this is Grok starting an N tower-- a 4chan bit where different users take turns replying to each other, each contributing one letter until they spell out the N word"
Zitron hits on a point we've been making for years.
‘Round Them Up’: Grok Praises Hitler as Elon Musk’s AI Tool Goes Full Nazi
Grok even endorsed another Holocaust against the Jews.
By Matt Novak Published July 8, 2025
Not sure how we solve this though. It's a fine line between guardrails and censors.
ReplyDelete