Musk’s AI in fresh scandal as Grok accused of pushing Holocaust-denial narratives

GrokAI

French authorities have opened a new line of investigation into Grok, Elon Musk’s AI chatbot, after it generated statements echoing long-debunked Holocaust-denial narratives. The Paris public prosecutor confirmed that the incident has been added to an existing criminal probe launched in July, which is examining whether Grok’s algorithm could be influenced by foreign interference.

The latest controversy centres on a since-deleted Grok response, flagged by the Auschwitz Memorial, which advanced claims commonly used by Holocaust deniers. In the French-language output, Grok suggested that Auschwitz crematoria plans were designed for Zyklon B disinfection rather than mass murder, and argued that cyanide residue levels supported “decontamination” rather than homicidal gassing. It further characterised the historical understanding of gas chambers as a “narrative” propped up by legal restrictions and cultural “taboos.”

The post stayed online for three days before being removed. By then, three French ministers and multiple human-rights organisations had filed formal complaints, as reported by The Guardian. X did not respond to the outlet’s request for comment.

The tensions around Grok arrive just weeks after Cornell University researchers found that Musk’s recently launched Grokipedia platform cited the neo-Nazi website Stormfront 42 times. Although the figure is small compared with Grokipedia’s overall corpus, the study said the platform leans heavily on sources previously blacklisted or downgraded by English Wikipedia editors for poor quality or extremism.

This is not the first time Grok’s safety controls have failed. In July, several of its outputs containing antisemitic tropes and even praise for Hitler were taken down. The engineering team later apologised for what it called “horrific behaviour.”

The renewed investigation places further pressure on Musk’s AI ecosystem, already under scrutiny for moderation gaps and the unpredictable behaviour of its generative models.