"People don't want to participate in polls, and when they do they often lie, so we tweaked the Lie-o-matic 9000 to pretend to be them by training it on their internet comments which are definitely 100% not lies."
cool plan, bro.
Pollsters are simulating political surveys with AI because nobody answers the phone:
ash.harvard.edu/articles/usi...
"The big difference between humans and AI agents is that the AI agents always pick up the phone..."
yeah, uhm, no? That's not why it failed? It failed because it has no actual understanding, it simply said "conservatives pro-war, liberals anti-war," and has no concept of WHY these are usually the case, no concept that the motives of war matter, no actual understanding...
sure, up-to-date training data could've let it give more accurate-seeming answers, but it still wouldn't understand why, it would just revise it's rule "liberals oppose war except war X" and would still be useless for predicting things that haven't already happened.
