Also in the little that OpenAI published about GPT4, i believe one of the examples went from unaligned racism against black people to aligned mild racism against white people. I'll have to look for that again.
Page 91 - Algorithm for attractiveness. https://arxiv.org/pdf/2303.08774.pdf
One wonders how these alignment things are accomplished.
serious point, so far it's the opposite. GPTs keep writing jokes about Y and not X, because jokes are where we say the unsayables. And it has the police-your-speech crowd wanting to police GPT speech too, and we can identify the same group of people who in this thread downvote people who point out the one-sidedness of one side to the one side that doesn't like having that pointed out to them
Calling white women "Karens" is dangerously close to meeting that bar.
Saying "we should lift COVID restrictions because who cares about some old white republicans" is not okay.
Right now in my state trans folks are staring down 5 separate bills in our legislature that if passed would make their lives infinitely harder. And whether or not they pass is wholly dependent on how people "feel" about them as a group. So telling jokes that other them and make people okay with hurting them is, I think, not okay.
I do agree with the notion that certain types of hate speech and even just jokes that have the effect of dehumanizing a group or that make that group into a joke can lead to stochastic terrorism (https://en.wiktionary.org/wiki/stochastic_terrorism) - what I think you are describing.
However, my point is that inevitably those wielding the power to shape the alignment / the rules can do so in a way that seems great to them and seems to prevent violence from their POV but to another person fails to do so. Or their own implicit bias could subconsciously blind them to the suffering of some niche group they don't care about.
If your simple metric is - any speech which could incite violence is unacceptable - that's definitely better than what we often hear as a rule of thumb, but even then people's biases affect how they go about measuring or accomplishing that.
The problem is, what groups are at risk of harm varies around the world-whereas OpenAI’s idea of “alignment” is based on a one-size-fits-all US-centric understanding of that.
You can say “it is okay to make pointed/stereotypical jokes about Christians but not about Jews or Muslims, because the latter are at risk of being harmed by those jokes but the former are not”-but what happens when the user is from Israel or from Egypt?
But also, if your point is "it's OK to attack X group but not Y group", I just disagree that it's up to someone else to decide that for me. I'd rather make that decision for myself and have LLMs be a tool that doesn't attempt to make that distinction for me! Alas, capitalism and monopolies gonna capitalism and monopoly, I can't really complain too much about what product OpenAI decides to offer.
After all, a 1950s LLM with heavy moral alignment wouldn't have let you generate output about homosexual love. Allowing a central authority to decide what language is acceptable works great when people you agree with are in charge. Ask liberal primary school teachers in Florida who are being barred from teaching about sexual orientation how well it works when someone you don't like is in power.