The model itself is much more lax about such stuff than ChatGPT and especially Claude. The filters are applied on top of that, but products using it via the API don't suffer this problem.