Oh please let it be a fully egalitarian future…

Attached is a perfectly acceptable social media post and two entirely understandable LLM responses to the post and a reimagining of the post.

One relevant data point in terms of how soon we’d conceivably be refactoring language models and updating our own views:

From their research, Sáenz and Poston Jr expect the United States to have progressed to overall white minority demography by 2044.

Of course, a white minority doesn’t necessarily equate a dismantling of a power structure favoring white folks.

It shouldn’t need to be said on the fediverse that racists need not reply - expect most folks will approach this academically as intended.

PS: if you hate this post, please try to be as nice as you can about it - e.g. toss me a polite sentence after you downvote

  • Skull giver@popplesburger.hilciferous.nl
    172·
    6 months ago

    AI is as shit as its input. The internet, as a whole, is racist as hell, but it’s also the biggest source of training data. It’s too much work to compose a dataset without biases, so instead every AI model is trained after the fact, or prompts are changed, in an attempt to combat this bias.

    This is how we got Google’s AI inserting black people into generated images specifically about white people, and how AI models learned to treat certain populations with satin gloves. Fixing the root cause of the problem isn’t financially viable, and all of this stuff is developed with private funds, so a patch job is the best we can get.

    I don’t think it’ll get better. I grew up with “don’t use Wikipedia as a source” drilled into me and I don’t know anyone who didn’t go to Wikipedia every time they needed to read up on a new subject. We’ll get the same with AI models, except AI models lie in more subtle ways than the Wikipedia trolls that I encountered as a kid.

    I think AI will exaggerate the slight biases introduced to combat the problems of the dataset to the extreme. As society continues to take the output if AI models as truth, and bigots continue to write terrible shit online that makes it into the infinite copyright violation treadmill, the effect will slowly become more pronounced, but not fast enough to get noticed.