The researchers “found that, as language models grow, covert racism increases. Ethical guardrails, they learned, simply teach language models to be more discreet about their racial biases.”
Based on what they use to 'train' language models, I'm sure all our biases are magnified and as the models grow racism, sexism, homophobia, transphobia, ableism, saneism, classism, etc. all become more covert.
I mean, the most commonly avaliable huge dataset is the internet. And even with all the safeguards, human moderation, and weight system - it is mind boggling that the creators didn't forsee this outcome.
Although I wonder if LLMs would pick up on the subtler forms of -isms if trained on literature.
They did foresee it, they just ignored it, likely assuming they'd find a way to fix it later.
They also trained LLMs on literature, directly infringing copyright of almost every published author with no compensation. Their value is wealth stolen from writers.
I'm not sure why @ValentinHoffman from the @alleninstitute.bsky.social would say this when AI and Algorithms are ALREADY being used to determine parol decisions in US Criminal Justice system
I'm guessing AI in general versus LLMs in particular? But even for LLMs we are inevitably close if not already there because many people use ChatGPT for just anything these days, so I assume judges, parole boards, and more are starting to rely on their racist pal who's fun to bias with.
How is anyone finding this surprising? they get their info by scraping the Internet, it's like saying, I'm surprised my oil soaked wood caught light when I threw it on the fire
