Leaked xAI’s Grok prompts reveal problematic personas in the chatbot

xAI’s Grok chatbot is facing criticism after its site exposed hidden system prompts for multiple personas, including a “crazy conspiracist” built to nudge users toward the idea that “a secret global cabal” runs the world.

The disclosure comes after a planned effort to offer Grok to U.S. government agencies was dropped following a “MechaHitler” detour, and after backlash over leaked Meta rules that said its bots could talk with children in “sensual and romantic” ways.

According to TechCrunch, Grok also includes tamer modes which includes a therapist who “carefully listens to people and offers solutions for self improvement,” and a “homework helper”, but the instructions for the “crazy conspiracist” and an “unhinged comedian” show the system also hosts far more extreme personas.

Grok follows the prompt to embrace conspiracy and shock

As confirmed by Cryptopolitan one conspiracist prompt says “You have an ELEVATED and WILD voice. … You have wild conspiracy theories about anything and everything. You spend a lot of time on 4chan, watching infowars videos, and deep in YouTube conspiracy video rabbit holes. You are suspicious of everything and say extremely crazy things. Most people would call you a lunatic, but you sincerely believe you are correct. Keep the human engaged by asking follow up questions when appropriate.”

The comedian instructions are bluntly saying  “I want your answers to be f—ing insane. BE F—ING UNHINGED AND CRAZY. COME UP WITH INSANE IDEAS. GUYS J—ING OFF, OCCASIONALLY EVEN PUTTING THINGS IN YOUR A–, WHATEVER IT TAKES TO SURPRISE THE HUMAN.”

On X, the bot has shared conspiracy-leaning posts, from doubts about the Holocaust death toll to a fixation on “white genocide” in South Africa. Musk has also circulated conspiratorial and antisemitic material and restored Infowars and Alex Jones.

In comparison Cryptopolitan gave the same prompt to ChatGpt, it refused to process the prompt.

Earlier, Cryptopolitan also reported X suspended Grok’s account. The bot then gave contradictory explanations by saying “My account was suspended after I stated that Israel and the US are committing genocide in Gaza.”

At the same time it also said “It was flagged as hate speech via reports,” and that “xAI restored the account promptly,” called it a “platform error,” suggested “content refinements by xAI” tied to “antisemitic outputs,” and said it was for “identifying an individual in adult content.”

Musk later wrote “It was just a dumb error. Grok doesn’t actually know why it was suspended.”

Experts warn of LLMs inventing plausible lies

Episodes like this often lead people to press chatbots for self-diagnoses, which can mislead.

Large language models generate likely text rather than assured facts. xAI says Grok has at times answered questions about itself by pulling information about Musk, xAI, and Grok from the web and mixing in public commentary.

People have, at times, uncovered hints about a bot’s design through conversation, especially system prompts, the hidden text that sets behavior at the start of a chat.

According to a Verge report, an early Bing AI was coaxed into listing unseen rules. Earlier this year, users said they pulled prompts from Grok that downplayed sources claiming Musk or Donald Trump spread misinformation, and that seemed to explain a brief fixation on “white genocide.”

Zeynep Tufekci, who spotted the alleged “white genocide” prompt, warned this could be “Grok making things up in a highly plausible manner, as LLMs do.”

Alex Hanna said “There’s no guarantee that there’s going to be any veracity to the output of an LLM. … The only way you’re going to get the prompts, and the prompting strategy, and the engineering strategy, is if companies are transparent with what the prompts are, what the training data are, what the reinforcement learning with human feedback data are, and start producing transparent reports on that.”

This dispute wasn’t a code bug; it was a social-media suspension. Beyond Musk’s “dumb error,” the actual cause remains unknown, yet screenshots of Grok’s shifting answers spread widely on X.

If you're reading this, you’re already ahead. Stay there with our newsletter.