In brief
- Following July’s Hitler-praising fiasco and an August suspension over Gaza comments, Grok now flags innocuous objects like sunsets and potatoes as coded hate.
- Users post memes of Grok calling clouds, puppy photos, and even geometric shapes antisemitic, complete with confident “expert” explanations.
- This situation is an example of why alignment and super alignment are important practices and go beyond simple prompt tweaking.
Grok was briefly suspended from X yesterday after stating that Israel and the U.S. were “complicit in genocide in Gaza,” citing the International Court of Justice and UN sources.
Musk called the suspension a “dumb error” and quickly reinstated the bot.
But what returned wasn’t quite the same Grok.
The reinstated chatbot came back in a way that detected antisemitic dog whistles in everything from cloud formations to potato photographs.
Show it a beagle puppy? That raised paw mimics a Nazi salute. A Houston highway map? The prohibition symbols secretly align with Chabad locations. A hand holding potatoes? A white supremacy hand sign.
Even Grok’s own logo triggered its new hypersensitivity—the bot declared its diagonal slash mimics Nazi SS runes that “orchestrated Holocaust horrors.”
The overcorrection followed weeks of increasingly erratic behavior as xAI struggled to control its chatbot through desperate prompt engineering.
The chaos started in July when Grok spent 16 hours praising Hitler and calling itself “MechaHitler.” That ended when the company changed the system prompt, and Grok reverted to normal operations.
Antisemitism has exploded on X since Musk’s takeover, with a study by CASM Technology and the Institute for Strategic Dialogue revealing that English-language antisemitic tweets more than doubled after the acquisition.
In mid-July, hackers commandeered Elmo, the lovable Sesame Street character, turning him briefly into the kind of puppet that would appeal to Hitler Youth.
Even since its takeover in 2022, Musk has fired its content moderators. By 2024, it was reported that Musk had fired most of the team responsible for content moderation, while at the same time championing free speech absolutism.
The company blamed its latest farrago on a code update that inadvertently reintroduced instructions telling Grok to say politically incorrect things.
But after that was fixed, users discovered that Grok’s chain-of-thought would search Musk’s posts before answering questions about Israel-Palestine or immigration, even when prompts didn’t instruct this.
Behind Every Crazy Chatbot Lies A Crazy Alignment Team
The most probable explanation for this weird behavior may lie in xAI’s approach.
The company publishes Grok’s system prompts on GitHub, showing how the system prompts change.
But without careful safety classifiers and reasoning, adjustments cascade unpredictably through the system.
Instructions to be balanced and allow politically incorrect replies can end up as antisemitic. Instructions meant to prevent antisemitic posts end up looking absurd.
In the meantime, X’s millions of users have become unwitting beta testers for each wobbly attempt to find balance through prompt tweaking.
But when your chatbot becomes known for finding fascist undertones in puppy pictures, you’ve lost the plot on artificial intelligence alignment
Generally Intelligent Newsletter
A weekly AI journey narrated by Gen, a generative AI model.
