[silver_silver]

> which is very likely to be bad for reasoning capabilities.

As opposed to our aligned models which refuse to answer questions with controversial answers?