I was live on BBC News talking about the psychology of conspiracies (and how to combat them with AI-based tools). Based on our recent Science paper with @dgrand.bsky.socialyoutu.be/xHN6YTzhKNo?...
YouTube video by BBC News
I was on NBC’s Nightly News tonight talking about our work (@dgrand.bsky.socialyoutu.be/uIC2DbZWr0Y?...
YouTube video by NBC News
And here is a Research Briefing summary of our work in Nature Human Behaviour from myself & @dgrand.bsky.socialrdcu.be/dT3dq
I'd expect it to work for incorrect /"epistemically suspect" beliefs of all kinds. I expect it to *increase* accurate beliefs
🚨Out in Science!🚨 Conspiracy beliefs famously resist correction, ya? WRONG: We show brief convos w GPT4 reduce conspiracy beliefs by ~20%! -Lasts over 2mo -Works on entrenched beliefs -Tailored AI response rebuts specific evidence offered by believers www.science.org/doi/10.1126/...
I think the problem w this that we've struggled with when trying to design such a thing is that there is so much heterogeneity in terms of which conspiracies are named, and which versions of a given consp are named. It makes it really hard to come up with canned debunks
Yep agreed, that is def a good idea. FWIW we do have a conditon in the follow up where we tell the AI to write less, and this actually increases the effect (prob bc participants are more inclined to actually read it)
Agreed!
Ya although we also have a recent NHB paper with a parallel finding for humans: fact checker warning labels decrease belief in, and sharing of, false claims even among those who distrust fact-checkers: www.nature.com/articles/s41...
Do fact-checker warning labels work for those who distrust fact-checkers? Martel and Rand analyse 21 experiments and find that warning labels reduce belief in, and sharing of, false posts both on aver...
Ya, we have a follow-up paper we are currently writing up with an active control where the AI just tries to persuade them without providing any reasons or arguments, and it doesn't work. Conversely telling the AI to give evidence/facts without being polite and friendly doesn't reduce effect at all