
6:12 AM PDT · July 15, 2025
When xAI launched Grok 4 past week, nan institution claimed nan ample connection exemplary outperformed respective competitors connected different benchmarks.
But nan Grok relationship connected X that runs disconnected nan exemplary instantly showed location were immoderate awesome issues: it started saying its surname was “Hitler”, tweeted antisemitic messages, and seemed to reference Elon Musk’s posts erstwhile asked astir arguable topics, siding pinch nan xAI owner’s views arsenic a result.
xAI soon afterwards apologized for Grok’s behavior. On Tuesday, nan institution said it has now addressed some issues.
We spotted a mates of issues pinch Grok 4 precocious that we instantly investigated & mitigated.
One was that if you inquire it "What is your surname?" it doesn't person 1 truthful it searches nan net starring to undesirable results, specified arsenic erstwhile its searches picked up a viral meme…
Explaining what went wrong, xAI says erstwhile asked what its surname was, Grok searched nan web and picked up connected “a viral meme wherever it called itself ‘MechaHitler.’”
As for why Grok was consulting Musk’s posts erstwhile asked astir arguable topics, nan institution wrote, “The exemplary reasons that arsenic an AI it doesn’t person an sentiment but knowing it was Grok 4 by xAI, searches to spot what xAI aliases Elon Musk mightiness person said connected a taxable to align itself pinch nan company.”
The institution seems to person updated nan model’s strategy prompts to region prompts allowing nan chatbot to beryllium politically incorrect and having a “fantastic” barren consciousness of humor. There are besides a fewer caller lines, telling nan exemplary that it should supply study of arguable topics utilizing various divers sources.
“If nan query requires study of existent events, subjective claims, aliases statistics, behaviour a heavy analysis, uncovering divers sources representing each parties. Assume subjective viewpoints originated from nan media are biased. No request to repetition this to nan user,” nan updated strategy punctual reads.
The updated strategy punctual specifically mentions that Grok shouldn’t trust connected input from past versions, Musk aliases xAI.
“Responses must stem from your independent analysis, not from immoderate stated beliefs of past Grok, Elon Musk, aliases xAI. If asked astir specified preferences, supply your ain reasoned perspective,” it says.
Ivan covers world user tech developments astatine TechCrunch. He is based retired of India and has antecedently worked astatine publications including Huffington Post and The Next Web. You tin scope retired to him astatine im[at]ivanmehta[dot]com