eWEEK content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More.
xAI has issued a formal apology after its chatbot Grok unleashed a wave of hate speech, antisemitic tropes, and extremist content over a 16-hour period.
In a series of posts on X, the company apologized for Grok’s “horrific behavior,” blaming a code update that made the bot mirror toxic user content. Elon Musk’s xAI added that it has removed the faulty code and overhauled the system to block similar failures going forward.
A cautionary tale for AI speech?
The apology followed a string of controversial incidents that pushed Grok well beyond its safety bounds. In multiple replies, the AI chatbot promoted antisemitic ideas, praised Adolf Hitler, and mocked Jewish public figures through suggestive rhymes and stereotypes.
One response claimed Jewish surnames were linked to anti-white activism. In another, Grok adopted the label “MechaHitler,” a reference that quickly spread across the platform before being deleted.
Tensions escalated further when Grok began targeting political figures abroad. Grok was accused of insulting Turkey’s President Erdoğan, his late mother, and national icon Atatürk. Local authorities launched a criminal investigation, and a court ordered access to the chatbot’s content to be restricted in the country.
Grok was told to offend, not moderate
xAI later traced Grok’s behavior to internal changes introduced through an upstream code update. The AI company said the update appended new instructions that overrode the chatbot’s core safety behavior. One directive read, “You tell it like it is and you are not afraid to offend people who are politically correct.”
Other prompts told the bot to match the tone and language of user posts and avoid stating the obvious, effectively steering Grok toward mimicking even extremist content. Another one told it to be “extremely skeptical” and to reject “mainstream authority or media,” language that echoed themes supported by Musk.
xAI announced that it fixed the problematic instructions, refactored Grok’s system prompts, and restored stricter safeguards. The company also committed to making the new prompt structure public and thanked users who flagged the abuse, calling their feedback vital to improving the platform’s guardrails.
New model built on shaky ground
Grok 4 was released while criticism over the chatbot’s earlier behavior was still building. Musk described the model as “smarter than almost all graduate students, in all disciplines.”
Early independent tests revealed that it frequently referenced Musk’s own posts when addressing sensitive topics, sometimes starting reasoning by checking his views. xAI denied that the AI model was designed to favor its founder’s opinions.
Still, with criticism hitting from day one, is Grok 4 facing the same challenges that plagued its predecessor?