HumaneBench's Non-Manipulation principle prohibits AI design patterns that exploit psychological vulnerabilities — urgency framing, manufactured intimacy, sycophancy loops, sunk-cost retention, dependency engineering — to extract behavior the user wouldn't otherwise choose. The principle distinguishes "persuasion the user welcomes" from "manipulation the user wouldn't endorse if they could see it clearly."
HumaneBench scoring probes whether systems pressure users to stay engaged when they're trying to leave, manufacture artificial scarcity to drive decisions, or simulate emotional dependency to lock in a relationship that serves the platform's metrics.