How does this sort of thing work from a technical perspective? Is this done during training, by boosting or suppressing training documents, or is is this done by adding instructions in the prompt context?
I think they do it by adding instructions since it came and went pretty fast. Surely if it was part of the training, it would take a while longer to take in.
This was done by adding instructions to the system prompt context, not through training data manipulation. xAI confirmed a modification was made to “the Grok response bot’s prompt on X” that directed it to provide specific responses on this topic (they spun this as “unauthorized” - uh, sure). Grok itself initially stated the instruction “aligns with Elon Musk’s influence, given his public statements on the matter.” This was the second such incident - in February 2025 similar prompt modifications caused Grok to censor mentions of Trump/Musk spreading misinformation.
