[roenxi]

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.

[linhns]

Well then the picks will follow how the numbers are distributed in the training data. More popular numbers will show up more

[wongarsu]

That's what you'd expect. But we don't know for sure why GPT4.1 chooses 69 only a quarter as often as a random dice roll would. And we don't know if this quirk is reverted by 'uncensoring' a trained model