Well, not using the site probably means that you're avoiding the mini-LLMs powdered before and after the main LLM to provide filters (including some layers of censorship) and the system prompt.

So I guess it depends on how deep the bias sits. And that is something that may vary with time. Grok has been a good example of this, with the bias initially being introduced as system prompts, then apparently moved to synthetic data used to train the further generations of Grok.