The problem is that all current AI models have either been trained specifically to never say anything anyone could possibly get offended by, or have been trained on data generated by those models.