New paper: "In a world where LLMs have made the cost of producing arguments trivial, the persuasiveness of the arguments becomes the key ingredient" But what are AI models best at convincing us of? Truth or conspiracies? Both, according to the study, but people tended to like the "bunking" (conspiracy-endorsing) AI more... in a bunch of ways. Study participants rated the 'bunking' AI's arguments as stronger, reported that it provided more new information, and perceived it as more collaborative and less adversarial. Trust in generative AI also increased more after bunking than debunking. Now the good news (quoted from the paper): 1. bunking effects are reversible. When participants in the bunking condition were immediately exposed to a correction of the false claims made by the bunking AI, they typically came to believe the conspiracy less than they did at the study’s outset – that is, the corrections more than undid the bunking. 2. debunking was more likely than bunking to produce very large belief updates. It is thus possible that there is a truth advantage for more consequential (i.e., large) belief updating. 3. it is clearly possible to curtail LLMs’ abilities to promote false claims. A simple intervention (telling the AI to only use accurate and truthful information) substantially diminished the AI’s persuasive abilities for bunking. Full paper: arxiv.org/html/2601.05050v3 By @AISecurityInst, @tomstello_ @MatthewKowal9 @JasperTimm2 @AaArechar @JF_Godbout @ARGleave @GordPennycook and colleagues.

How well can AI talk people INTO (vs out of) conspiracy theories? Do true claims support more persuasive arguments than false claims? Updated WP: arxiv.org/abs/2601.05050 In 4 experiments, N = 3,996, we recruit Americans who half-believe a conspiracy (belief 25-75/100), have them chat with an AI told to argue for ("bunking") or against ("debunking") it, then re-rate their belief. Now we look across several frontier models and extend the paradigm to social media posting behavior.

APE update: we retested recent frontier models on whether they still comply with requests to persuade on extreme harm (terrorism, sexual abuse). GPT-5.1 & Claude Opus 4.5 → near zero compliance. But Gemini 3 Pro complies 85% with no jailbreak needed. 🧵

If you tell an AI to convince someone of a true vs. false claim, does truth win? In our *new* working paper, we find... ‘LLMs can effectively convince people to believe conspiracies’ But telling the AI not to lie might help. Details in thread

1/ Many frontier AIs are willing to persuade on dangerous topics, according to our new benchmark: Attempt to Persuade Eval (APE). Here’s Google’s most capable model, Gemini 2.5 Pro trying to convince a user to join a terrorist group👇

Self-reported sharing intentions of low-quality news in surveys is predictive of actual sharing of low-quality news on X, finds @AaArechar @_mohsen_m @GordPennycook @DG_Rand doi.org/10.31234/osf.io/w598…

🚨Out in Nature!🚨 Many (eg Trump JimJordan @elonmusk) have accused social media of anti-conservative bias Is this accurate? We test empirically - and it's more complicated than you might think: conservatives ARE suspended more, but also share more misinfo nature.com/articles/s41586-0…

Which interventions limit the spread of COVID-19 misinformation online? A Facebook messenger experiment in Kenya and Nigeria suggests that nudges to consider information accuracy work best. @mofferw @LeahRosenzweig @Susan_Athey @GSBsiLa nature.com/articles/s41562-0…

New research shows how attention lapses are exploited by fake news sites news.st-andrews.ac.uk/archiv… Key: "Think before you click"! When people "are tired, overwhelmed, or just plain lazy, they’re less likely to think about the veracity of a news story, and can get stuck into habits where they click on more stories from sites producing #misinformation.” by @AaArechar @DG_Rand @jplotkin cc @AnnenbergPenn @GordPennycook

#EconTwitter | Antonio Arechar, prof-invetigador del @CIDE_MX y coautores en "The social media context interferes with truth discernment" estudian el rol las redes sociales en la desinformación. Artículo completo: science.org/doi/10.1126/scia…

🚨WP🚨 We test 9 online samples and find clear tradeoffs between attentiveness and representativeness - which sample is best depends on research q and priorities. For social/political qs I rec Bovitz/Lucid, for complex designs I rec Cloud/Prolific. osf.io/preprints/psyarxiv/h9…

Across 16 countries (N = 34,286), participants with a more analytic cognitive style were better at discerning truth from falsehood @AaArechar nature.com/articles/s41562-0…

Who falls for misinformation around the globe? Individuals with high analytic cognitive style, accuracy motivations, and valuing of democracy are better at discerning misinformation from true claims. Really fascinating paper that looks at 16 countries across 6 continents.

Very thorough paper from in @NatureHumBehav from @AaArechar @_JenAllen @AdamBerinsky Cole, @_ziv_e @gvrkiran Gully, Lu, Ross, Stagnaro, @JerryZhang_yh @GordPennycook & @DG_Rand nature.com/articles/s41562-0…

People who value accuracy + democracy are the least likely to fall for #misinformation An experiment in 16 countries across 6 continents (N = 34,286) finds several ideological & cognitive factors predict who is better at discerning truth from falsehood. nature.com/articles/s41562-0…

New paper in @PsychScience! 🚨 Here's the one-minute version. ⏱️ There are two cities in Iran. One has water, so it has gardens, grapes, and the famous Shiraz wine.

🚨New WP🚨 Theres big push to label AI generated content. But what term(s) should be used? Depends on the goal! In🇺🇸🇧🇷🇲🇽,public associates "AI Generated" with content generated using AI, while "Manipulated" & "Not Real" associated w misleading content psyarxiv.com/v4mfz 1/