Maybe a little off-topic here, but if they use Twitter and Reddit as datasets to train their LLMs, how can they guarantee that the AI models will be unbiased. Prejudice, hatred, and ill intentions can all be very delicate to identify.