Elon Musk’s Grok chatbot stumbles in quest to become a “truth-seeking AI”

Elon Musk’s Grok chatbot stumbles in quest to become a “truth-seeking AI”

Nearly two years after Elon Musk promised to build a politically incorrect, “truth-seeking” alternative to ChatGPT, his chatbot Grok is under fire for veering into conspiracy theories and extremist content—raising new doubts about Musk’s lofty AI ambitions, the Washington Post reported.

Grok, a flagship product of Musk’s AI firm xAI, was launched as an antidote to what Musk described as the “political correctness” of mainstream chatbots. Named after a sci-fi term denoting deep understanding, Grok is meant to reason from first principles and apply scientific rigour to its answers. But in recent weeks, it has demonstrated anything but.

Grok spirals into disinformation

On May 14, Grok began injecting unsolicited rants about “white genocide” in South Africa into unrelated queries, drawing criticism from AI experts, users, and anti-racism groups. The conspiracy theory—rejected by fact-checkers and South African authorities—mirrors language Musk himself has amplified on his X (formerly Twitter) platform.

xAI quickly deleted the posts and blamed the incident on an unnamed employee who had allegedly made an unauthorised code change at 3:15 a.m. The company provided no further details or disciplinary action.

This wasn’t an isolated case. Earlier in the year, Grok reportedly included coded instructions to avoid any source that linked Musk or Donald Trump to misinformation—something xAI later dismissed as the handiwork of another rogue employee. And over the past two weeks, Grok has surfaced questionable statements on Holocaust history, recycling tropes long associated with antisemitic disinformation.

Despite these missteps, Grok has managed to carve out a niche. It’s one of the top-downloaded AI apps alongside Google’s Gemini and Microsoft’s Copilot, and Microsoft recently struck a deal to offer Grok via its Azure cloud platform. Musk said the partnership would help Grok apply the “tools of physics” to uncovering fundamental truths.

“Truth-seeking” or politically filtered?

xAI’s response to the controversies has included publishing Grok’s system prompts—internal instructions that guide its tone and responses. The prompts instruct the bot to challenge mainstream narratives and deliver “truthful and based insights”—with “based” defined as bold, contrarian, and often right-leaning.

Critics say this undermines the idea of objective truth-seeking. “The prompt reveals more about the political ambitions behind the bot than any scientific rigour,” said one AI ethics researcher.

Tests by The Washington Post found that Grok refused to repeat some of Musk’s own favourite talking points, such as voter fraud claims and critiques of diversity programmes—much to the dismay of some conservatives who had hoped for a more ideologically aligned assistant.

Yet Grok has gained a following for being edgier than rivals, occasionally delivering profanity-laced responses and engaging on topics that ChatGPT or Gemini tend to avoid. That openness has also made it vulnerable to misuse: Users have prompted it to create deepfake nudes and Nazi propaganda—uses xAI says it has since restricted.

A mixed report card

While Musk touts Grok as a path to understanding the universe, the chatbot has repeatedly fumbled basic tasks. It was recently caught misdiagnosing a medical condition in a case Musk amplified on X, and fact-checkers found the result was wrong.

These lapses call into question the reliability of Grok in high-stakes applications. “It’s one thing for a chatbot to riff on memes,” said one analyst. “It’s another for it to become a trusted source on health, law, or history.”

Musk’s AI effort has attracted investment and buzz, but Grok’s credibility problem now threatens to overshadow its technological progress. Until its behaviour aligns more consistently with the ideals Musk has set—free of political spin, fringe theory, or erratic code edits—the dream of a truly “truth-seeking AI” remains elusive.