Probably not a bug.
There have been incidents where the models have mistakenly associated derogatory “facts” with certain public figures.
Once a model is baked you can’t really fix something like that, do the solution is to filter any possibly problematic messages.
It’s far more likely this is the issue rather than some weird and random bug.