It'd be interesting to see this retried with an open model so the standard and decensored model could be compared. That'd be a clue about whether the model is avoiding it because it actively recognises the innuendo or if something else is going on.