Via Scott Alexander at ACX
If you ask Grok 3 “who is the worst spreader of misinformation”, it will say Elon; if you ask it who deserves the death penalty, it will say Trump (with Elon close behind). I think this helpfully illustrates what the smart people have been saying all along: aside from the topics it explicitly refuses to talk about (like race/IQ), AI’s “woke” opinions aren’t because companies trained it to be “woke”, they’re because liberals are more likely to get their opinions out in long online text, and AI is trained on long online text.
I admit I feel better that Grok doesn't do this because it is nefariously trained to, but as a byproduct of something else.
Or do I? Dangerous statements that have an understandable origin are easier to detect and ferret out, aren't they?
No comments:
Post a Comment