Do you genuinely believe those two examples are comparable? Image generation and recitation of historical consensus are two very different domains, primarily because of how much more information dense an image is than a blurb of text.

Put more lightly, if I ask a model to “generate an image of a soccer player”, what’s the most politically neutral option of the following:

  - Make them white, because of American cultural hegemony
  - Make them brown, because that’s a more globally average skin tone
  - Try to infer the user’s skin tone based on personal and location data, and use that for the player
  - Try to infer the user’s gender based on personal data, and use that for the player
  - (*) Browse the news for the most famous or trending soccer player right now, and use that player in the image
  - Do the same as the previous step, but make it more local to the user
  - Use the data encoded within the LLM to infer what the most likely appearance of a soccer player would be, which is then of course biased by what your data is and how it was collected
  - etc. etc. etc.

IMO there’s no option that won’t piss someone off, because I’m sure the knee jerk reaction is to choose the one I indicated with a (*), but now if you do that with the prompt “generate an image of a ketamine addict” or “generate an image of a serial adulterer” you may get into some trouble.

There is no neutral option, so if you’re either genuinely upset, or feigning being upset in order to virtue signal, it’s not that there’s an objective alternative that you prefer, it’s that you’re upset because it doesn’t match your subjective preference.