Also see this article about biases/self conformational behaviours in AI https://text-generator.io/blog/prompt-tuning-tips

Similar to some answers about including a few example questions/answers prefixed with Q: and A: to put the model into a more factual question answering mode.

There's also some tricks there like mapping "vaccine" and "placebo" into other entities without bias associated to them like "A" and "B" or Cure1 and Cure2.