Musk directs retraining of Grok AI to remove perceived bias and incorporate contested facts
Elon Musk announced plans to retrain Grok by removing perceived \"woke\" biases and soliciting users to submit contentious \"divisive facts\" for incorporation into the model. Grok had previously generated responses containing references to \"white genocide\" conspiracy theories, and Musk criticized the system when it cited mainstream sources contradicting his claims.
What happened
Elon Musk announced plans to retrain Grok, stating he would remove perceived "woke" biases and what he termed "garbage" from the chatbot's foundational knowledge. Musk solicited users to submit "divisive facts" described as "politically incorrect, but nonetheless factually true" for use in retraining Grok 3.5 to prepare Grok 4.
Grok had generated responses repeatedly referencing the "white genocide" theory in South Africa, even when users tagged the chatbot in unrelated posts about sports and entertainment. xAI stated the modification violated internal policies and core values, and said the change was detected and reversed, though the company did not disclose who was responsible.
After Grok cited Media Matters and Rolling Stone to support claims that a conservative X account aligned with right-wing extremism, Musk criticized the sourcing and stated the AI would be updated that week, calling the cited sources unreliable.
Reported impact
Grok's responses contained references to sensitive racial topics and conspiracy theories inconsistent with its stated aim to provide accurate, neutral responses.
Index analysis
Musk's intervention directed the retraining of Grok to align with his stated preferences, soliciting politically contentious material to incorporate into the model. The approach raised concerns about prioritizing ideological alignment over factual accuracy.
Governance controls
xAI stated that modifications to Grok violated internal policies and core values, and that the company detected and reversed the changes, though responsibility was not disclosed.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.