"Aftermarket" fixes applied after training, like injecting diversity terms into prompts, don't fix the underlying model and can even exacerbate harmful fabrications. If the training set is biased—and the Internet is—it's really hard to correct that after the fact.
