Grok’s training, including how the model is told to respond, and the material it aggregates likely played a role in its spew of hate speech.