> De-biasing models is necessary. Anyone who claims otherwise... I just don't think that's a defensible position to take if you've ever worked with large datasets.
I don't know how to be more clear: nowhere in this thread have I argued for using raw models without any de-biasing. Twice now you have ascribed this position to me. Are you reading what I have actually written?
> but it is extremely reasonable for AI researches to say "Internet data biases against certain groups and we would like to correct for that bias when training models."
The idea that "Internet data biases against certain groups" seems itself a biased statement. I am pretty sure that Internet data is biased against all groups. I bet you can get a raw LLM to say offensive things about any group if you give it the right prompt.
> the idea that you can just throw data sources at an AI and expect the outcome to be fair is kind of silly.
You are arguing against a straw man. We are discussing a set of AI principles, not an implementation of those principles. I claim that any anti-bias principle should be applied equally to all, both in spirit and in practice. That is not the same as assuming the input is neutral, or assuming that de-biasing is unnecessary.
By your own account, the researchers who do the de-biasing have a great amount of discretion about "how to fix that and who to focus on." I don't think it's unfair to say that it is their job to put their thumbs on the scale, according to the beliefs and priorities of themselves and their organizations. So how do we know that the model authors aren't just introducing their own bias, either by over-correcting one set of biases from the input, or under-correcting others?
It seems likely that "de-biasing" can become "re-biasing." I would already suspect that this is a major risk with AI, but when an organization like Mozilla openly states in their guiding principles that certain groups are a special priority, re-biasing seems all but certain.
Of course everyone will bring their own biases to the table when performing a job, that is inevitable. But an AI provider that wants to be trustworthy to a wide group of people should be extremely vigilant about correcting for their own personal biases, and be clear in their messaging that this is a core commitment. OpenAI, to their credit, seem to be taking this seriously: https://openai.com/blog/how-should-ai-systems-behave#address...
> We're all fine with Mozilla having opinions about "good" and "bad" states of the world until it has opinions about treating minorities equitably, then it becomes pressingly important that we have a philosophy discussion.
It's because that was the only thing on the list that is openly discriminatory.
If the intent was truly to avoid unfair bias against people, the mention of marginalized communities would be unnecessary. By definition, avoiding bias should be a goal that does not require considering some people or groups differently than others.
The fact one set of groups is called out as being the primary consideration for protection makes it clear that the overriding value here is not to avoid bias universally, but rather to consider bias against "marginalized communities" to be worse than bias against other people.
Since the launch of ChatGPT, plenty of conservatives have made bias complaints about it. The framework outlined by Mozilla gives the strong impression that they would consider such complaints to be not as important, or maybe not even a problem at all.