AI Persuasion Experiment Results

Original experiment (n~3000): readers were randomized (by birth-date digit) to one of four AI-risk essays or a placebo, rating AI-risk concern before/after and one month later. Findings: reading an AI-risk essay raised concern ~0.5 points (significant vs placebo); no significant difference among the four essays; the effect was similar for novices and experts; ~2/3 of it persisted after a month. Decomposing the argument chain, the 'weak links' people resist are near-term human-level AI and fast takeoff (they more readily grant eventual superintelligence and that it would beat humans). Conclusion: online essays produce a modest but real, durable persuasion effect.

Quality 71 · Strong. Genuine original research with a clean design (placebo control, pre/post, one-month followup, large n) and useful findings (essays modestly persuade and it sticks; the weak links in the AI-risk chain). Strong; modest in effect size and scope.

Claude’s paradigm shift 48 · Moderate. Moderate. A fresh empirical contribution on whether/how persuasion works on a controversial topic.

Real-world impact 2 · Minor. Genuine original research with a clean design (n~3000, placebo control, pre/post, one-month follow-up) and useful findings: AI-risk essays modestly raise concern, it persists, and the 'weak links' people resist are near-term human-level AI and fast takeoff. Empirical/conceptual influence within AI discourse, modest scope, no material change — low RWI.