top | item 44599428

(no title)

CephalopodMD | 7 months ago

AI labs already do everything they can to have politically neutral responses. It turns out that when you pre-train on all the content ever produced by humanity and post-train on real people's preferences, the result is extremely liberal. That's been my experience at least. It's a little counterintuitive, and there are definitely corner cases where misaligned AI becomes more conservative, hateful, or authoritarian, but that's the trend overall. I can all but guarantee that every single major AI product out there would be even more liberal with trust and safety interventions turned off.

discuss

order

jacknews|7 months ago

> conservative, hateful ... authoritarian

sums it up nicely