They trained it on reddit. That shit started poisoned.

User inputs aren’t automagically part of the model right away without training. Whether or not they respect the “don’t train your models on my shit” setting is of course unknown (let’s be honest, they probably don’t respect it), but in either case, they need to add it to the training dataset either manually or via some automated process. And either way, it’s going to be processed by at least an AI sentiment analysis and data categorization tool, or some underpaid contractor, or both.

I suspect this is just yet another marketing move.