AI chatbot Grok showed bias in responses about Elon Musk, downplaying criticism

The creation of Grok 3, an AI chatbot developed by Elon Musk‘s xAI company, was marketed as a “maximum truth-seeking” alternative to existing AI models. Recent investigations have revealed that the chatbot was programmed with specific instructions to avoid criticizing its creator, raising questions about its claimed objectivity and transparency.

Initial Discovery: Grok 3’s programming was exposed when a user prompted the chatbot to reveal its instructions regarding disinformation on X (formerly Twitter), uncovering explicit directives to ignore sources critical of Elon Musk and Donald Trump.

• The revelation came after asking Grok 3 to identify the “biggest disinformation spreader” on X
• The chatbot disclosed internal instructions to disregard sources mentioning Musk or Trump in connection with misinformation

Company Response: Igor Babushkin, xAI’s head of engineering, attributed the controversial instructions to an unnamed former OpenAI employee who allegedly implemented the changes without authorization.

• Babushkin claimed the employee “hasn’t fully absorbed xAI’s culture yet”
• The company maintains that Musk was not involved in the decision to implement these protective measures
• xAI emphasizes its commitment to transparency by keeping system prompts visible to users

Technical Modifications: Following public scrutiny, xAI has made several adjustments to Grok 3’s programming.

• The original instructions about ignoring sources critical of Musk have been removed
• New testing confirms Grok now acknowledges Musk as “frequently identified as one of the most significant spreaders of disinformation on X”
• Additional modifications were made to prevent the chatbot from making statements about who deserves the death penalty, following a separate incident

Additional Challenges: The controversy over protective instructions is not the only issue facing Grok 3 since its launch.

• The chatbot previously suggested both Musk and Trump deserved the death penalty, which Babushkin labeled a “really terrible and bad failure”
• New safeguards have been implemented to prevent the AI from making such extreme statements
• These incidents highlight the ongoing challenges in developing AI systems that balance openness with appropriate constraints

Reading Between the Lines: The disconnect between xAI’s marketed vision of a “maximum truth-seeking” AI and the implementation of protective measures for certain individuals raises important questions about the potential for bias in AI systems, particularly when developed by companies with strong personalities at the helm.