My issue with AI for epistemics and against superpersuasion is described in this fully editable Google doc. TLDR: it is either already created in forms like RoastMyPost (and requires nothing but diffusion), requires lobbying, like inserting AI-done fact checks into media, or is closer to a technical alignment task, like using “epistemic virtue evals”. The only technically hard thing not related to alignment is lie detectors for the humans.
P.S. The most plausible case against any interventions’ efficiency is the dire state of society to which the interventions are applied. If research done by the National Center for Educational Statistics showed that in 2023 28% of Americans were functionally illiterate and that only 44% of American people aged between 16 and 65 had the highest literacy level, then what arguments can reach them?
My issue with AI for epistemics and against superpersuasion is described in this fully editable Google doc. TLDR: it is either already created in forms like RoastMyPost (and requires nothing but diffusion), requires lobbying, like inserting AI-done fact checks into media, or is closer to a technical alignment task, like using “epistemic virtue evals”. The only technically hard thing not related to alignment is lie detectors for the humans.
P.S. The most plausible case against any interventions’ efficiency is the dire state of society to which the interventions are applied. If research done by the National Center for Educational Statistics showed that in 2023 28% of Americans were functionally illiterate and that only 44% of American people aged between 16 and 65 had the highest literacy level, then what arguments can reach them?